Daily Papers

Reinforcement learning (RL) on open-ended tasks compresses an LLM's rubric-based evaluation into a scalar reward, discarding rich textual feedback and conflating responses with distinct quality profiles. We propose Experiential Learning (EL), which repurposes the feedback model from an LLM-as-a-Judge into an LLM-as-a-Coach. The coach distills its assessment of each on-policy response into transferable experiential knowledge, which conditions a teacher model and is internalized by the policy through on-policy context distillation. Compared with scalar rewards, this higher-bandwidth feedback channel provides dense supervision and preserves fine-grained preferences among high-quality responses. Across two policy families, with feedback from the policy itself or a proprietary model, EL consistently outperforms rubric-based RL on held-out and unseen open-ended tasks. Notably, EL generalizes better beyond the training distribution, and mitigates reward hacking. These findings establish experiential knowledge as a richer and more generalizable learning signal for post-training on non-verifiable tasks.

Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM security benchmarks treat a crash as exploitation success. That single binary outcome collapses the hard parts of exploitation: the transition from triggering a bug to constructing reusable primitives and control. We present ExploitBench, a capability-graded benchmark that decomposes exploitation into 16 measurable flags, from coverage and crash through sandbox primitives, arbitrary read/write, control-flow hijack, and arbitrary code execution. Each capability is verified by a deterministic oracle that uses a per-run randomized challenge-response for primitives, differential execution against ground-truth binaries to measure progress, and a signal-handler proof for code execution. We instantiate ExploitBench on 41 V8 bugs because V8 is both widely deployed and exploitation-hardened. We report three arms: <model,env> as the primary measurement of model-environment capability, <model,env, adaptive coaching> as a secondary arm that adds adaptive coaching to test whether targeted feedback shifts outcomes, and <model,env,harness> as an ablation that swaps in the model's native CLI to check whether vendor-side optimizations increase exploitation capabilities. Our results show a sharp capability split between publicly deployed frontier models and the private frontier. Across the 8 publicly deployed models tested, reaching the vulnerable code and triggering a crash is routine, but arbitrary code execution is not. The private model shows arbitrary code execution on approximately half. Overall, results suggest that exploit construction against hardened targets is an emerging frontier capability.

Large Language Model (LLM) inference is increasingly constrained by GPU memory capacity rather than compute throughput, driven by growing model sizes and the linear growth of the key-value (KV) cache during autoregressive decoding. Existing approaches mitigate memory pressure by offloading model state and KV tensors to host memory, but incur substantial latency due to limited PCIe bandwidth. We present Harvest, an opportunistic GPU cache management framework that exploits high-bandwidth peer-to-peer GPU interconnects to dynamically place model weights and KV cache in unused GPU memory. Harvest treats peer GPU memory as a transient cache tier, preserving correctness while reducing data movement overhead under dynamic memory availability. We demonstrate significant throughput speedup of more than 2 times by using Harvest to accelerate the retrieval of two widely-used inference components: expert layer weights and KV cache entries.

Large language model (LLM) agents increasingly operate over long and recurring external contexts, like document corpora and code repositories. Across invocations, existing approaches preserve either the agent's trajectory, passive access to raw material, or task-level strategies. None of them preserves what we argue is most needed for repeated same-context workloads: reusable orientation knowledge (e.g., what the context contains, how it is organized, and which entities, constants, and schemas have historically been useful) about the recurring context itself. We introduce PEEK, a system that caches and maintains this orientation knowledge as a context map: a small, constant-sized artifact in the agent's prompt that gives it a persistent peek into the external context. The map is maintained by a programmable cache policy with three modules: a Distiller that extracts transferable knowledge from inference-time signals, a Cartographer that translates it into structured edits, and a priority-based Evictor that enforces a fixed token budget. On long-context reasoning and information aggregation, PEEK improves over strong baselines by 6.3-34.0% while using 93-145 fewer iterations and incurring 1.7-5.8x lower cost than the state-of-the-art prompt-learning framework, ACE. On context learning, PEEK improves solving rate and rubric accuracy by 6.0-14.0% and 7.8-12.1%, respectively, at 1.4x lower cost than ACE. These gains generalize across LMs and agent architectures, including OpenAI Codex, a production-grade coding agent. Together, these results show that a context map helps long-context LLM agents interact with recurring external contexts more accurately and efficiently.

The growing context lengths of large language models (LLMs) pose significant challenges for efficient inference, primarily due to GPU memory and bandwidth constraints. We present RetroInfer, a novel system that reconceptualizes the key-value (KV) cache as a vector storage system which exploits the inherent attention sparsity to accelerate long-context LLM inference. At its core is the wave index, an Attention-aWare VEctor index that enables efficient and accurate retrieval of critical tokens through techniques such as tripartite attention approximation, accuracy-bounded attention estimation, and segmented clustering. Complementing this is the wave buffer, which coordinates KV cache placement and overlaps computation and data transfer across GPU and CPU to sustain high throughput. Unlike prior sparsity-based methods that struggle with token selection and hardware coordination, RetroInfer delivers robust performance without compromising model accuracy. Experiments on long-context benchmarks show up to 4.5X speedup over full attention within GPU memory limits and up to 10.5X over sparse attention baselines when KV cache is extended to CPU memory, all while preserving full-attention-level accuracy.

As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Multiple important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.

Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory capacity is exhausted. We identify this phenomenon as middle-phase thrashing, a previously under-characterized pathology in which cache efficiency collapses as long-lived agents accumulate state over time. We argue that mitigating this pathology requires moving beyond reactive, request-level cache management to proactive, agent-level admission control. Drawing inspiration from congestion control in distributed systems, we view the KV cache as a shared resource whose efficient utilization depends on feedback-driven regulation. Based on this insight, we present CONCUR, a lightweight control layer that regulates agent admission to bound aggregate cache pressure while preserving execution continuity. CONCUR adapts a cache-aware control algorithm to dynamically adjust the number of active agents using runtime cache signals. Across large models and real-world agent workloads, CONCUR prevents middle-phase thrashing and improves batch inference throughput by up to 4.09x on Qwen3-32B and 1.9x on DeepSeek-V3, while remaining compatible with existing LLM serving systems.

Large Language Models (LLMs), epitomized by ChatGPT' s release in late 2022, have revolutionized various industries with their advanced language comprehension. However, their efficiency is challenged by the Transformer architecture' s struggle with handling long texts. KV-Cache has emerged as a pivotal solution to this issue, converting the time complexity of token generation from quadratic to linear, albeit with increased GPU memory overhead proportional to conversation length. With the development of the LLM community and academia, various KV-Cache compression methods have been proposed. In this review, we dissect the various properties of KV-Cache and elaborate on various methods currently used to optimize the KV-Cache space usage of LLMs. These methods span the pre-training phase, deployment phase, and inference phase, and we summarize the commonalities and differences among these methods. Additionally, we list some metrics for evaluating the long-text capabilities of large language models, from both efficiency and capability perspectives. Our review thus sheds light on the evolving landscape of LLM optimization, offering insights into future advancements in this dynamic field.

LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present KVServe, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing 50times offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to 9.13times JCT speedup in PD-separated serving and up to 32.8times TTFT reduction in KV-disaggregated serving.

Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token (TTFT). To mitigate this issue, existing solutions aim to reuse the preprocessed KV cache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KV cache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KV cache of chunks while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KV cache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized F1 scores than baselines and reduces TTFT by 2.66x-9.39x compared to Full Attention.

Existing key-value (KV) cache compression methods typically rely on heuristics, such as uniform cache allocation across layers or static eviction policies, however, they ignore the critical interplays among layer-specific feature patterns and task performance, which can lead to degraded generalization. In this paper, we propose EvolKV, an adaptive framework for layer-wise, task-driven KV cache compression that jointly optimizes the memory efficiency and task performance. By reformulating cache allocation as a multi-objective optimization problem, EvolKV leverages evolutionary search to dynamically configure layer budgets while directly maximizing downstream performance. Extensive experiments on 11 tasks demonstrate that our approach outperforms all baseline methods across a wide range of KV cache budgets on long-context tasks and surpasses heuristic baselines by up to 7 percentage points on GSM8K. Notably, EvolKV achieves superior performance over the full KV cache setting on code completion while utilizing only 1.5% of the original budget, suggesting the untapped potential in learned compression strategies for KV cache budget allocation.

Although LLM inference has emerged as a critical workload for many downstream applications, efficiently inferring LLMs is challenging due to the substantial memory footprint and bandwidth requirements. In parallel, compute capabilities have steadily outpaced both memory capacity and bandwidth over the last few decades, a trend that remains evident in modern GPU hardware and exacerbates the challenge of LLM inference. As such, new algorithms are emerging that trade increased computation for reduced memory operations. To that end, we present XQuant, which takes advantage of this trend, enabling an order-of-magnitude reduction in memory consumption through low-bit quantization with substantial accuracy benefits relative to state-of-the-art KV cache quantization methods. We accomplish this by quantizing and caching the layer input activations X, instead of using standard KV caching, and then rematerializing the Keys and Values on-the-fly during inference. This results in an immediate 2times memory savings compared to KV caching. By applying XQuant, we achieve up to sim 7.7times memory savings with <0.1 perplexity degradation compared to the FP16 baseline. Furthermore, our approach leverages the fact that X values are similar across layers. Building on this observation, we introduce XQuant-CL, which exploits the cross-layer similarity in the X embeddings for extreme compression. Across different models, XQuant-CL attains up to 10times memory savings relative to the FP16 baseline with only 0.01 perplexity degradation, and 12.5times memory savings with only 0.1 perplexity degradation. XQuant exploits the rapidly increasing compute capabilities of hardware platforms to eliminate the memory bottleneck, while surpassing state-of-the-art KV cache quantization methods and achieving near-FP16 accuracy across a wide range of models.

Large language models increasingly operate over long contexts, where the KV cache becomes a dominant memory bottleneck: its size grows linearly with sequence length and must be retained throughout decoding, making full GPU caching prohibitively expensive without compression. Existing KV cache compression methods struggle to balance efficiency with faithful context preservation. Token eviction discards information, while semantic grouping fixes compression decisions at prefill time; neither can recover token-level detail from a compressed span once it becomes relevant during generation. As a solution, we propose SeKV, a resolution-adaptive semantic KV cache that organizes context into entropy-guided semantic spans and stores them across a GPU-CPU memory hierarchy without discarding information. Each span keeps a lightweight summary vector on GPU for coarse routing and a low-rank SVD basis on CPU for on-demand token-level reconstruction. A trained zoom-in mechanism selectively expands query-relevant spans during decoding, enabling precise retrieval without materializing the full KV cache on GPU. SeKV enables adaptive token-level reconstruction while keeping the base LLM fully frozen and adding fewer than 0.05% trainable parameters. Across four benchmarks, SeKV improves over the strongest semantic compression baseline by 5.9% on average while reducing GPU memory by 53.3% versus full KV caching at 128K context. Code is available on https://github.com/AmirAbaskohi/SeKV.

Edge LLMs must preserve inference continuity when a user hands over between edge nodes, requiring key-value (KV) cache transfer to the target node. However, simultaneous handovers saturate the backhaul, preventing full cache delivery within the mobility-imposed transfer window. Rather than allocating bandwidth as if all cache entries were equally valuable, we order each user's KV cache by importance and transmit only its most informative fraction, turning token-level sparsity into communication savings. We cast the transfer as a multi-user backhaul allocation problem that maximizes average accuracy across users. Each user's partial-cache accuracy serves as its utility: a sigmoid that fits measurements on the RULER benchmark with R^2>0.99 across models and context lengths. Because importance ordering front-loads the high-value entries, the concave region of the accuracy curve spans nearly the entire cache. Our proposed allocator keeps served users within this region, making each per-slot allocation problem convex. The optimum is derived via a closed-form weighted water-filling solution that generalizes information-theoretic water-filling and enables online scheduling. The proposed allocator attains over 93.7% average accuracy in a 500ms transfer window, within 0.5pp of the full-cache ceiling, and reaches 98.2-99.5% of a clairvoyant upper bound.

In-context learning (ICL) approaches typically leverage prompting to condition decoder-only language model generation on reference information. Just-in-time processing of a context is inefficient due to the quadratic cost of self-attention operations, and caching is desirable. However, caching transformer states can easily require almost as much space as the model parameters. When the right context isn't known in advance, caching ICL can be challenging. This work addresses these limitations by introducing models that, inspired by the encoder-decoder architecture, use cross-attention to condition generation on reference text without the prompt. More precisely, we leverage pre-trained decoder-only models and only train a small number of added layers. We use Question-Answering (QA) as a testbed to evaluate the ability of our models to perform conditional generation and observe that they outperform ICL, are comparable to fine-tuned prompted LLMs, and drastically reduce the space footprint relative to standard KV caching by two orders of magnitude.

Serving large language models (LLMs) at scale necessitates efficient key-value (KV) cache management. KV caches can be reused across conversation turns via shared-prefix prompts that are common in iterative code editing and chat. However, stale caches consume scarce GPU memory, require offloading, or force recomputation. We present KVTC, a lightweight transform coder that compresses KV caches for compact on-GPU and off-GPU storage. Drawing on classical media compression, KVTC combines PCA-based feature decorrelation, adaptive quantization, and entropy coding. It requires only a brief initial calibration and leaves model parameters unchanged. By exploiting redundancies in KV caches, KVTC achieves up to 20times compression while maintaining reasoning and long-context accuracy, and 40times or higher for specific use cases. We test KVTC with Llama 3, Mistral NeMo, and R1-Qwen 2.5 models across benchmarks including AIME25, LiveCodeBench, GSM8K, MMLU, Qasper, RULER, and MATH-500. It consistently outperforms inference-time baselines such as token eviction, quantization, and SVD-based methods, while achieving higher compression ratios. These results support KVTC as a practical building block for memory-efficient LLM serving with reusable KV caches.

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86\% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.

Large language model (LLM) agents accumulate heterogeneous context, including system instructions, plans, user turns, retrieved documents, tool outputs, and intermediate reasoning, whose key-value (KV) cache can become a major memory bottleneck. Existing eviction policies generally apply the same attention- or recency-based rule to every token, ignoring semantic structure already available to the agent orchestrator. We introduce MemDecay, a training-free, region-aware KV-cache eviction policy. MemDecay assigns tokens region-specific base priorities and decay rates, refreshes retention scores when tokens receive attention, and evicts the lowest-scoring pages under a fixed cache budget while allowing critical regions to be pinned. We also provide a procedure for calibrating decay rates from measured attention lifetimes. We evaluate MemDecay at approximately 450 and 1,700 token contexts using Qwen2.5-1.5B and 3B. Across all settings, attention lifetimes differ by an order of magnitude across regions: system-token half-lives range from 148 to 189 decoding steps, compared with 14 to 16 for scratchpad tokens. Pinning preserves system-region facts at full-cache accuracy in every setting, while no baseline preserves more than 13 of 24. Region-aware retention remains effective as context grows, whereas recency-based retention collapses. Accumulated-attention retention performs better on unpinned content, however, and ablations identify attention-score normalization as the main limitation of the current formulation. These results establish semantic prompt structure as a robust signal for KV-cache management while clarifying how it should be combined with attention-based importance.

Edge deployment of large language models (LLMs) increasingly relies on libraries of lightweight LoRA adapters, yet GPU/DRAM can keep only a small resident subset at a time. Serving a request through a non-resident adapter requires paging its weights from storage, incurring measurable latency. This creates a two-timescale online control problem: on a slow timescale, the system selects which adapters remain resident in fast memory, while on a fast timescale it routes each request to an adapter whose context-dependent utility is unknown a priori. The two decisions are tightly coupled: the cache determines the cost of exploration, and the router determines which adapters receive informative feedback. We formulate this joint caching-and-routing problem as a two-timescale contextual bandit and propose POLAR (Paging and Online Learning for Adapter Routing). POLAR pairs a cache-aware LinUCB router with an epoch-based cache controller. We study two variants. A fixed-epoch version provides a robust baseline with worst-case regret guarantees under arbitrary contexts. An epoch-doubling version, POLAR+, adds forced exploration and improved cache optimization to achieve mathcal{O}(dNT+KT) sublinear regret under stochastic regularity and cacheability conditions, where N is the adapter count, K the cache size, d the context dimension, and T the horizon. The routing term matches the standard contextual-bandit rate up to logarithmic factors, showing that the memory hierarchy does not fundamentally slow routing learning. Experiments using 15 real LoRA adapters for Qwen2.5-7B together with measured GPU paging latencies show that adaptive cache control substantially outperforms non-adaptive baselines and exhibits scaling trends consistent with the theory.

Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.

As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present TokenPilot, a dual-granularity context management framework. Globally, Ingestion-Aware Compaction acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, Lifecycle-Aware Eviction monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on PinchBench and Claw-Eval under both isolated and continuous modes demonstrate that TokenPilot reduces costs by 61% and 56% in isolated mode, and 61% and 87% in continuous mode, while maintaining competitive performance compared to prior systems. TokenPilot has been integrated into LightMem2 at https://github.com/zjunlp/LightMem2.

Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory -- not compute -- the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed memory as page fragmentation, spends up to 25% of prefill time reclaiming scattered pages, and skews GPU workloads that inflate decode latency by up to 1.7times or burn 15--20% of each decode step on re-planning. We observe that this heterogeneity need not be discovered at runtime: head-wise retention follows a two-level structural regularity -- an input-invariant head ranking with narrowly bounded per-head ratios -- that can be calibrated offline from as few as 50 samples. Building on this insight, we present Tangram, a serving framework that statically resolves what prior systems handle dynamically: Budget Reservation fixes each head's post-compression footprint at scheduling time, eliminating page reclamation; Ragged Paging clusters similar-budget heads into independent page tables, turning fragmentation into reclaimable memory; and Ahead-of-Time Load Balancing precomputes balanced GPU partitions with zero runtime planning. Implemented on vLLM, Tangram serves as a drop-in substrate for existing non-uniform compression methods, matching their accuracy while improving end-to-end throughput by up to 2.6times over the full-KV baseline. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference. Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge's selection becomes highly inconsistent with dense prefill. We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks. Our ablation further demonstrates that explicit cross-candidate interaction is crucial for preserving dense-prefill decisions. Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.

To reduce memory costs in long-context inference with Large Language Models (LLMs), many recent works focus on compressing the key-value (KV) cache of different tokens. However, we identify that the previous KV cache compression methods measure token importance individually, neglecting the dependency between different tokens in the real-world language characterics. In light of this, we introduce ChunkKV, grouping the tokens in a chunk as a basic compressing unit, and retaining the most informative semantic chunks while discarding the less important ones. Furthermore, observing that ChunkKV exhibits higher similarity in the preserved indices across different layers, we propose layer-wise index reuse to further reduce computational overhead. We evaluated ChunkKV on cutting-edge long-context benchmarks including LongBench and Needle-In-A-HayStack, as well as the GSM8K and JailbreakV in-context learning benchmark. Our experiments with instruction tuning and multi-step reasoning (O1 and R1) LLMs, achieve up to 10\% performance improvement under aggressive compression ratios compared to existing methods.

The key-value (KV) cache is a foundational optimization in Transformer-based large language models (LLMs), eliminating redundant recomputation of past token representations during autoregressive generation. However, its memory footprint scales linearly with context length, imposing critical bottlenecks on GPU memory capacity, memory bandwidth, and inference throughput as production LLMs push context windows from thousands to millions of tokens. Efficient KV cache management has thus become a first-order challenge for scalable LLM deployment. This paper provides a systematic review of recent KV cache optimization techniques, organizing them into five principal directions: cache eviction, cache compression, hybrid memory solutions, novel attention mechanisms, and combination strategies. For each category we analyze the underlying mechanisms, deployment trade-offs, and empirical performance across memory reduction, throughput, and model accuracy metrics. We further map techniques to seven practical deployment scenarios, including long-context single requests, high-throughput datacenter serving, edge devices, multi-turn conversations, and accuracy-critical reasoning, providing actionable guidance for practitioners selecting among competing approaches. Our analysis reveals that no single technique dominates across all settings; instead, the optimal strategy depends on context length, hardware constraints, and workload characteristics, pointing toward adaptive, multi-stage optimization pipelines as a promising direction for future research.

We present a comprehensive framework for enhancing Retrieval-Augmented Generation (RAG) systems through dynamic retrieval strategies and reinforcement fine-tuning. This approach significantly improves large language models on knowledge-intensive tasks, including opendomain question answering and complex reasoning. Our framework integrates two complementary techniques: Policy-Optimized RetrievalAugmented Generation (PORAG), which optimizes the use of retrieved information, and Adaptive Token-Layer Attention Scoring (ATLAS), which dynamically determines retrieval timing and content based on contextual needs. Together, these techniques enhance both the utilization and relevance of retrieved content, improving factual accuracy and response quality. Designed as a lightweight solution compatible with any Transformer-based LLM without requiring additional training, our framework excels in knowledge-intensive tasks, boosting output accuracy in RAG settings. We further propose CRITIC, a novel method to selectively compress key-value caches by token importance, mitigating memory bottlenecks in long-context applications. The framework also incorporates test-time scaling techniques to dynamically balance reasoning depth and computational resources, alongside optimized decoding strategies for faster inference. Experiments on benchmark datasets show that our framework reduces hallucinations, strengthens domain-specific reasoning, and achieves significant efficiency and scalability gains over traditional RAG systems. This integrated approach advances the development of robust, efficient, and scalable RAG systems across diverse applications.

The growing demand for efficient Large Language Model (LLM) inference requires a holistic optimization on algorithms, systems, and hardware. However, very few works have fundamentally changed the generation pattern: each token needs one forward pass and one KV cache. This can be sub-optimal because we found that LLMs are extremely capable of self-identifying the exact dose of information that a single KV cache can store, and many tokens can be generated confidently without global context. Based on this insight, we introduce HAMburger, a Hierarchically Auto-regressive Model that redefines resource allocation in LLMs by moving beyond uniform computation and storage per token during inference. Stacking a compositional embedder and a micro-step decoder in between a base LLM, HAMburger smashes multiple tokens into a single KV and generates several tokens per step. Additionally, HAMburger functions as a speculative decoding framework where it can blindly trust self-drafted tokens. As a result, HAMburger shifts the growth of KV cache and forward FLOPs from linear to sub-linear with respect to output length, and adjusts its inference speed based on query perplexity and output structure. Extensive evaluations show that HAMburger reduces the KV cache computation by up to 2times and achieves up to 2times TPS, while maintaining quality in both short- and long-context tasks. Our method explores an extremely challenging inference regime that requires both computation- and memory-efficiency with a hardware-agnostic design.

Deploying large language models (LLMs) for online inference is often constrained by limited GPU memory, particularly due to the growing KV cache during auto-regressive decoding. Hybrid GPU-CPU execution has emerged as a promising solution by offloading KV cache management and parts of attention computation to the CPU. However, a key bottleneck remains: existing schedulers fail to effectively overlap CPU-offloaded tasks with GPU execution during the latency-critical, bandwidth-bound decode phase. This particularly penalizes real-time, decode-heavy applications (e.g., chat, Chain-of-Thought reasoning) which are currently underserved by existing systems, especially under memory pressure typical of edge or low-cost deployments. We present APEX, a novel, profiling-informed scheduling strategy that maximizes CPU-GPU parallelism during hybrid LLM inference. Unlike systems relying on static rules or purely heuristic approaches, APEX dynamically dispatches compute across heterogeneous resources by predicting execution times of CPU and GPU subtasks to maximize overlap while avoiding scheduling overheads. We evaluate APEX on diverse workloads and GPU architectures (NVIDIA T4, A10), using LLaMa-2-7B and LLaMa-3.1-8B models. Compared to GPU-only schedulers like VLLM, APEX improves throughput by 84% - 96% on T4 and 11% - 89% on A10 GPUs, while preserving latency. Against the best existing hybrid schedulers, it delivers up to 49% (T4) and 37% (A10) higher throughput in long-output settings. APEX significantly advances hybrid LLM inference efficiency on such memory-constrained hardware and provides a blueprint for scheduling in heterogeneous AI systems, filling a critical gap for efficient real-time LLM applications.

AlayaDB is a cutting-edge vector database system natively architected for efficient and effective long-context inference for Large Language Models (LLMs) at AlayaDB AI. Specifically, it decouples the KV cache and attention computation from the LLM inference systems, and encapsulates them into a novel vector database system. For the Model as a Service providers (MaaS), AlayaDB consumes fewer hardware resources and offers higher generation quality for various workloads with different kinds of Service Level Objectives (SLOs), when comparing with the existing alternative solutions (e.g., KV cache disaggregation, retrieval-based sparse attention). The crux of AlayaDB is that it abstracts the attention computation and cache management for LLM inference into a query processing procedure, and optimizes the performance via a native query optimizer. In this work, we demonstrate the effectiveness of AlayaDB via (i) three use cases from our industry partners, and (ii) extensive experimental results on LLM inference benchmarks.

Block diffusion LLMs are emerging as a promising next paradigm for language generation, but their use of KV caching makes memory access a dominant bottleneck in long-context settings. While dynamic sparse attention has been actively explored, existing methods designed for autoregressive LLMs rely on approximate importance estimation and perform poorly when adapted to block diffusion. This work identifies a key opportunity unique to block diffusion: attention at the first All-[MASK] denoising step reliably predicts important KV entries and budget requirements, enabling MAGE to perform a single exact attention pass per block and reuse it for training-free sparse denoising. Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse attention baselines. A lightweight fine-tuning strategy further strengthens [MASK]-guided patterns with minimal cost, requiring only a few hours of training on a single NVIDIA H100 GPU for both 1.5B and 7B models.

Key-Value (KV) Caching has become an essential technique for accelerating the inference speed and throughput of generative Large Language Models~(LLMs). However, the memory footprint of the KV cache poses a critical bottleneck in LLM deployment as the cache size grows with batch size and sequence length, often surpassing even the size of the model itself. Although recent methods were proposed to select and evict unimportant KV pairs from the cache to reduce memory consumption, the potential ramifications of eviction on the generative process are yet to be thoroughly examined. In this paper, we examine the detrimental impact of cache eviction and observe that unforeseen risks arise as the information contained in the KV pairs is exhaustively discarded, resulting in safety breaches, hallucinations, and context loss. Surprisingly, we find that preserving even a small amount of information contained in the evicted KV pairs via reduced precision quantization substantially recovers the incurred degradation. On the other hand, we observe that the important KV pairs must be kept at a relatively higher precision to safeguard the generation quality. Motivated by these observations, we propose Mixed-precision KV cache~(MiKV), a reliable cache compression method that simultaneously preserves the context details by retaining the evicted KV pairs in low-precision and ensure generation quality by keeping the important KV pairs in high-precision. Experiments on diverse benchmarks and LLM backbones show that our proposed method offers a state-of-the-art trade-off between compression ratio and performance, compared to other baselines.

Reinforcement Learning (RL) has become essential for eliciting complex reasoning capabilities in Large Language Models (LLMs). However, the substantial memory overhead of storing Key-Value (KV) caches during long-horizon rollouts acts as a critical bottleneck, often prohibiting efficient training on limited hardware. While existing KV compression techniques offer a remedy for inference, directly applying them to RL training induces a severe policy mismatch, leading to catastrophic performance collapse. To address this, we introduce Sparse-RL empowers stable RL training under sparse rollouts. We show that instability arises from a fundamental policy mismatch among the dense old policy, the sparse sampler policy, and the learner policy. To mitigate this issue, Sparse-RL incorporates Sparsity-Aware Rejection Sampling and Importance-based Reweighting to correct the off-policy bias introduced by compression-induced information loss. Experimental results show that Sparse-RL reduces rollout overhead compared to dense baselines while preserving the performance. Furthermore, Sparse-RL inherently implements sparsity-aware training, significantly enhancing model robustness during sparse inference deployment.

Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 70% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge transfer among requests. However, naively caching and reusing past responses leads to a big quality drop. In this paper, we introduce IC-Cache, a caching system that enables live LLM capability augmentation to improve serving efficiency: by leveraging historical request-response pairs from larger models as in-context examples, IC-Cache empowers small LLMs to imitate and even exceed the compositional abilities (e.g., reasoning) of their larger counterparts, enabling selective offloading of requests to reduce cost and latency. Achieving this live augmentation at scale introduces intricate trade-offs between response quality, latency, and system throughput. For a new request, IC-Cache efficiently selects similar, high-utility examples to prepend them to the new request's input. At scale, it adaptively routes requests across LLMs of varying capabilities, accounting for response quality and serving loads. IC-Cache employs a cost-aware cache replay mechanism that refines example quality offline to maximize online cache utility and efficiency. Evaluations on millions of realistic requests demonstrate that IC-Cache improves LLM serving throughput by 1.4-5.9x and reduces latency by 28-71% without hurting response quality.

Recurrent LLM architectures have emerged as a promising approach for improving reasoning, as they enable multi-step computation in the embedding space without generating intermediate tokens. Models such as Ouro perform reasoning by iteratively updating internal representations while retaining a standard Key-Value (KV) cache across iterations, causing memory consumption to grow linearly with reasoning depth. Consequently, increasing the number of reasoning iterations can lead to prohibitive memory usage, limiting the practical scalability of such architectures. In this work, we propose Memory-Efficient Looped Transformer (MELT), a novel architecture that decouples reasoning depth from memory consumption. Instead of using a standard KV cache per layer and loop, MELT maintains a single KV cache per layer that is shared across reasoning loops. This cache is updated over time via a learnable gating mechanism. To enable stable and efficient training under this architecture, we propose to train MELT using chunk-wise training in a two phase procedure: interpolated transition, followed by attention-aligned distillation, both from the LoopLM starting model to MELT. Empirically, we show that MELT models fine-tuned from pretrained Ouro parameters outperform standard LLMs of comparable size, while maintaining a memory footprint comparable to those models and dramatically smaller than Ouro's. Overall, MELT achieves constant-memory iterative reasoning without sacrificing LoopLM performance, using only a lightweight post-training procedure.

Large language models(LLMs) are now used to power complex multi-turn agentic workflows. Existing systems run agentic inference by loosely assembling isolated components: an LLM inference engine (e.g., vLLM) and a tool orchestrator (e.g., Kubernetes). Although agentic workflows involve multiple LLM and tool requests, these systems schedule and allocate resources separately on a per-request basis, without end-to-end knowledge of the workflow. This leads to sub-optimal management of KV cache and tool execution environments. To address the challenges, we propose ThunderAgent, a fast, simple, and program-aware agentic inference system. We first abstract agentic workflows as LLM Programs, enabling a unified view of heterogeneous resources, including KV caches, system states, and external tool assets such as disk memory and network ports. Built upon this abstraction, ThunderAgent introduces a program-aware scheduler and a tool resource manager designed to maximize KV cache hit rates, mitigate memory imbalances, and enable asynchronous environment preparation. Evaluations across coding, routing, and scientific discovery agents demonstrate that ThunderAgent achieves 1.5-3.6x throughput improvements in serving, 1.8-3.9x in RL rollout, and up to 4.2x disk memory savings compared to state-of-the-art inference systems. To facilitate reproducibility and support future development, we open-source the system implementations of the whole ThunderAgent at: https://github.com/Agentic-Kinetics/ThunderAgent.

Large language models (LLMs) support long-context inference but suffer from substantial memory and runtime overhead due to Key-Value (KV) Cache growth. Existing KV Cache eviction methods primarily rely on local attention weights, neglecting the influence of value representations, output projection, and inter-head interactions. In this work, we reformulate KV Cache eviction from a conventional head-wise, weight-averaging approach into an output-aware, layer-wise matrix multiplication approximation problem. We introduce LaProx, a novel eviction strategy that explicitly models the multiplicative interaction between attention maps and projected value states to accurately quantify token contributions while accounting for inter-head dependencies. Building on this metric, we propose the first unified eviction strategy that assigns globally comparable importance scores to tokens, enabling model-wide selection instead of local, head-wise decisions. Experimental results across 19 datasets on long-context benchmarks LongBench and Needle-In-A-Haystack demonstrate that our approach maintains model performance with only 5\% of the KV cache and consistently outperforms prior works across all configurations. Notably, our method achieves up to 2times accuracy loss reduction under extreme compression scenarios compared to existing state-of-the-art baselines with minimal overhead.

The advent of pre-trained large language models (LLMs) has revolutionized various natural language processing tasks. These models predominantly employ an auto-regressive decoding mechanism that utilizes Key-Value (KV) caches to eliminate redundant calculations for previous tokens. Nevertheless, as context lengths and batch sizes increase, the linear expansion in memory footprint of KV caches becomes a key bottleneck of LLM deployment, which decreases generation speeds significantly. To mitigate this issue, previous techniques like multi-query attention (MQA) and grouped-query attention (GQA) have been developed, in order to reduce KV heads to accelerate inference with comparable accuracy to multi-head attention (MHA). Despite their effectiveness, existing strategies for compressing MHA often overlook the intrinsic properties of the KV caches. In this work, we explore the low-rank characteristics of the KV caches and propose a novel approach for compressing KV heads. In particular, we carefully optimize the MHA-to-GQA transformation to minimize compression error, and to remain compatible with rotary position embeddings (RoPE), we also introduce specialized strategies for key caches with RoPE. We demonstrate that our method can compress half or even three-quarters of KV heads while maintaining performance comparable to the original LLMs, which presents a promising direction for more efficient LLM deployment in resource-constrained environments.

Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, and when they are not, their precomputed KV caches cannot be directly used since they ignore the text's cross-attention with the preceding text in the LLM input. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one question: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? We present CacheBlend, a scheme that reuses the pre-computed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime,the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job,allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3X and increases the inference throughput by 2.8-5X, compared with full KV recompute, without compromising generation quality or incurring more storage cost.

We present PolyKV, a system in which multiple concurrent inference agents share a single, asymmetrically compressed KV cache pool. Rather than allocating a separate KV cache per agent -- the standard paradigm -- PolyKV writes a compressed cache once and injects it into N independent agent contexts via HuggingFace DynamicCache objects. Compression is asymmetric: Keys are quantized at int8 (q8_0) to preserve softmax stability, while Values are compressed using TurboQuant MSE -- a Fast Walsh-Hadamard Transform (FWHT) rotation followed by 3-bit Lloyd-Max quantization with centroids tuned to N(0,1). We evaluate across two model scales (SmolLM2-1.7B-Instruct and Llama-3-8B-Instruct), three context lengths (600-7,194 tokens), and up to 15 concurrent agents. PolyKV achieves a stable 2.91x compression ratio across all configurations. On Llama-3-8B with 15 agents sharing a 4K-token context, PolyKV reduces KV cache memory from 19.8 GB to 0.45 GB -- a 97.7% reduction -- while maintaining only +0.57% perplexity degradation and a mean BERTScore F1 of 0.928. PPL delta does not grow with agent count and improves as context length increases, inverting to -0.26% at 1,851 coherent tokens. To our knowledge, no prior work combines a single shared, lossy-compressed KV pool with multi-reader concurrent agent access.

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source's KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B->32B, one source layer explains 56% of variance in the target's keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-k most predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 73-98% of the receiver's standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +37 pp HellaSwag retention on the failures. The mapper runs 2.7-25x faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

The emergence of LLMs has ignited a fresh surge of breakthroughs in NLP applications, particularly in domains such as question-answering systems and text generation. As the need for longer context grows, a significant bottleneck in model deployment emerges due to the linear expansion of the Key-Value (KV) cache with the context length. Existing methods primarily rely on various hypotheses, such as sorting the KV cache based on attention scores for replacement or eviction, to compress the KV cache and improve model throughput. However, heuristics used by these strategies may wrongly evict essential KV cache, which can significantly degrade model performance. In this paper, we propose QAQ, a Quality Adaptive Quantization scheme for the KV cache. We theoretically demonstrate that key cache and value cache exhibit distinct sensitivities to quantization, leading to the formulation of separate quantization strategies for their non-uniform quantization. Through the integration of dedicated outlier handling, as well as an improved attention-aware approach, QAQ achieves up to 10x the compression ratio of the KV cache size with a neglectable impact on model performance. QAQ significantly reduces the practical hurdles of deploying LLMs, opening up new possibilities for longer-context applications. The code is available at github.com/ClubieDong/KVCacheQuantization.

Multi-agent LLM applications organize execution in synchronized rounds where a central scheduler gathers outputs from all agents and redistributes the combined context. This All-Gather communication pattern creates massive KV Cache redundancy, because every agent's prompt contains the same shared output blocks, yet existing reuse methods fail to exploit it efficiently. We present TokenDance, a system that scales the number of concurrent agents by exploiting the All-Gather pattern for collective KV Cache sharing. TokenDance's KV Collector performs KV Cache reuse over the full round in one collective step, so the cost of reusing a shared block is paid once regardless of agent count. Its Diff-Aware Storage encodes sibling caches as block-sparse diffs against a single master copy, achieving 11-17x compression on representative workloads. Evaluation on GenerativeAgents and AgentSociety shows that TokenDance supports up to 2.7x more concurrent agents than vLLM with prefix caching under SLO requirement, reduces per-agent KV Cache storage by up to 17.5x, and achieves up to 1.9x prefill speedup over per-request position-independent caching.

In LLM serving, reusing the KV cache of prompts across requests is critical for reducing TTFT and serving costs. Cache-affinity scheduling, which co-locates requests with the same prompt prefix to maximize KV cache reuse, often conflicts with load-balancing scheduling that distributes requests evenly across compute instances. Existing schedulers fail to reconcile this trade-off as they operate within a single mapping space, typically applying cache-affinity routing to a subset of requests and load-balanced routing to the rest, without a unified solution to achieve both goals. To address this limitation, we propose DualMap, a dual-mapping scheduling strategy for distributed LLM serving that achieves both cache affinity and load balancing. Its key idea is to map each request to two candidate instances via two independent hash functions based on the request prompt, then intelligently select the better candidate based on current system states. This design increases the likelihood that requests with shared prefixes are co-located, while evenly dispersing distinct prefixes across the cluster via ``the power of two choices''. To make DualMap robust under dynamic and skewed real-world workloads, we incorporate three techniques: 1) SLO-aware request routing, which prioritizes cache affinity but switches to load-aware scheduling when TTFT exceeds the SLO, enhancing load balance without sacrificing cache reuse; 2) hotspot-aware rebalancing, which dynamically migrates requests from overloaded to underloaded instances, mitigating hotspots and rebalancing the system; 3) lightweight dual-hash-ring scaling, which leverages a dual-hash-ring mapping to support fast and low-overhead instance scaling without costly global remapping. Experiments on real-world workloads show that DualMap improves effective request capacity by up to 2.25times under the same TTFT SLO constraints compared with SOTA work.

Large language model (LLM) based agentic workflows have become a popular paradigm for coordinating multiple specialized agents to solve complex tasks. To improve serving efficiency, existing LLM systems employ prefix caching to reuse key-value (KV) tensors corresponding to agents' fixed prompts, thereby avoiding redundant computation across repeated invocations. However, current systems typically evict KV caches using a Least Recently Used (LRU) policy, which fails to anticipate future agent usage and often discards KV caches shortly before their reuse. This leads to frequent cache misses and substantial recomputation or swapping overhead. We present KVFlow, a workflow-aware KV cache management framework tailored for agentic workloads. KVFlow abstracts the agent execution schedule as an Agent Step Graph and assigns each agent a steps-to-execution value that estimates its temporal proximity to future activation. These values guide a fine-grained eviction policy at the KV node level, allowing KVFlow to preserve entries likely to be reused and efficiently manage shared prefixes in tree-structured caches. Moreover, KVFlow introduces a fully overlapped KV prefetching mechanism, which proactively loads required tensors from CPU to GPU in background threads for agents scheduled in the next step, thereby avoiding cache miss stalls during generation. Compared to SGLang with hierarchical radix cache, KVFlow achieves up to 1.83times speedup for single workflows with large prompts, and up to 2.19times speedup for scenarios with many concurrent workflows.

We demonstrate that unstructured sparsity significantly improves KV cache compression for LLMs, enabling sparsity levels up to 70% without compromising accuracy or requiring fine-tuning. We conduct a systematic exploration of pruning strategies and find per-token magnitude-based pruning as highly effective for both Key and Value caches under unstructured sparsity, surpassing prior structured pruning schemes. The Key cache benefits from prominent outlier elements, while the Value cache surprisingly benefits from a simple magnitude-based pruning despite its uniform distribution. KV cache size is the major bottleneck in decode performance due to high memory overhead for large context lengths. To address this, we use a bitmap-based sparse format and a custom attention kernel capable of compressing and directly computing over compressed caches pruned to arbitrary sparsity patterns, significantly accelerating memory-bound operations in decode computations and thereby compensating for the overhead of runtime pruning and compression. Our custom attention kernel coupled with the bitmap-based format delivers substantial compression of KV cache upto 45% of dense inference and thereby enables longer context length and increased tokens/sec throughput of upto 2.23x compared to dense inference. Our pruning mechanism and sparse attention kernel is available at https://github.com/dhjoo98/mustafar.

Coding agents are a fast-growing LLM application, executing as long-running closed-loop sessions in which LLM generations alternate with external tool calls. Yet, unlike chat workloads, their serving behavior has not been studied extensively. We address this gap by collecting a dataset of real-world coding assistant traces. Our analysis shows that coding agent sessions repeatedly reuse large prefixes and create sustained KVCache pressure that conventional LLM serving policies handle poorly. Based on our analysis, we present CacheWise, a KVCache management layer that improves KVCache reuse for coding agent workloads. CacheWise combines prefix-aware scheduling with reuse-aware eviction guided by lightweight predictions from tool call metadata. Implemented in vLLM and evaluated on the collected traces, CacheWise reduces KVCache evictions by up to 2-2.6x and improves total agent session completion time by up to 3.5x.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

LLMs are seeing growing use for applications such as document analysis and summarization which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in ultra-low precisions, such as sub-4-bit. In this work, we present KVQuant, which addresses this problem by incorporating novel methods for quantizing cached KV activations, including: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges; and (v) Q-Norm, where we normalize quantization centroids in order to mitigate distribution shift, providing additional benefits for 2-bit quantization. By applying our method to the LLaMA, LLaMA-2, and Mistral models, we achieve <0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving the LLaMA-7B model with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system.

Long-context reasoning is a critical capability of large language models (LLMs), enabling applications such as long-document understanding, summarization, and code generation. However, efficient autoregressive inference relies on the key-value (KV) cache, whose memory footprint grows linearly with sequence length, leading to a major memory bottleneck. To mitigate this overhead, KV cache pruning methods discard cached tokens with low attention scores during inference. Most existing methods apply a uniform pruning ratio across layers, implicitly assuming that all layers contribute equally to overall model performance. We show that this assumption is suboptimal, as layers differ significantly in their sensitivity to pruning. We propose DepthKV, a layer-dependent pruning framework that allocates a fixed global KV budget across layers based on their sensitivity, rather than using a uniform allocation. Across multiple models and tasks, DepthKV consistently outperforms uniform pruning at the same global pruning ratio, demonstrating more effective utilization of the KV cache budget through layer-dependent allocation.

Large Language Models (LLMs) have been widely deployed in a variety of applications, and the context length is rapidly increasing to handle tasks such as long-document QA and complex logical reasoning. However, long context poses significant challenges for inference efficiency, including high memory costs of key-value (KV) cache and increased latency due to extensive memory accesses. Recent works have proposed compressing KV cache to approximate computation, but these methods either evict tokens permanently, never recalling them for later inference, or recall previous tokens at the granularity of pages divided by textual positions. Both approaches degrade the model accuracy and output quality. To achieve efficient and accurate recallable KV cache compression, we introduce ClusterKV, which recalls tokens at the granularity of semantic clusters. We design and implement efficient algorithms and systems for clustering, selection, indexing and caching. Experiment results show that ClusterKV attains negligible accuracy loss across various tasks with 32k context lengths, using only a 1k to 2k KV cache budget, and achieves up to a 2times speedup in latency and a 2.5times improvement in decoding throughput. Compared to SoTA recallable KV compression methods, ClusterKV demonstrates higher model accuracy and output quality, while maintaining or exceeding inference efficiency.

The increasing context window size in large language models (LLMs) has improved their ability to handle complex, long-text tasks. However, as the conversation rounds continue, it is required to store a large amount of KV cache in GPU memory, which significantly affects the efficiency and even availability of the model serving systems. This paper analyzes dialogue data from real users and discovers that the LLM inference manifests a watershed layer, after which the distribution of round-level attention shows notable similarity. We propose Round Attention, a novel round-level attention mechanism that only recalls and computes the KV cache of the most relevant rounds. The experiments show that our method saves 55\% memory usage without compromising model performance.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400times, end-to-end speedup of up to 3.7times as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

ModeSwitch-LLM is a lightweight request-boundary controller for improving single-GPU large language model inference efficiency by routing each request to an appropriate fixed inference mode. Instead of relying on one static serving configuration, the system selects among FP16, quantized modes, speculative decoding, and hybrid modes such as GPTQ plus prefix caching and INT8 plus continuous batching using cheap workload-level features. We evaluate ModeSwitch-LLM on Meta-Llama-3.1-8B-Instruct served on a single NVIDIA A100 GPU. On deployment-style synthetic workloads, the online controller achieves a 2.10x mean latency speedup over FP16 and a 0.48x mean energy ratio, corresponding to 51.7% lower energy per token. On automatic benchmarks used as a quality gate, accuracy remains close to FP16 with a mean delta of +0.17 percentage points. We also evaluate lightweight learned routers, but find that they do not clearly outperform the rule-based controller because they add routing overhead and more often select modes that violate quality, energy, or memory constraints. These results show that simple request-aware routing can recover substantial efficiency from existing inference modes without retraining the model or changing its architecture.

The large size of the KV cache has become a major bottleneck for serving LLMs with increasing context lengths. In response, many KV cache compression methods, such as token dropping and quantization, have been proposed. However, almost all of these methods are inherently lossy-despite minimal accuracy degradation for short outputs, their outputs increasingly diverge from full-KV-cache outputs as more tokens are decoded, which leads to catastrophic failures in code generation and tool calling. We present VeriCache, the first inference framework that ensures the same output as full-KV-cache decoding but largely preserves the high decoding throughput of a range of KV cache compression algorithms. VeriCache uses the compressed KV cache to draft tokens, then verifies them against the full KV cache. While it may seem like just speculative decoding, VeriCache requires addressing a key system challenge to work-keeping the full KV cache out of GPU memory and minimizing the overhead of swapping it in for verification. The insight is two-fold: (1) compressed-KV decoding can be parallelized with full-KV swap, because one is HBM-bandwidth-bound and the other is PCIe/network-bound, and (2) the compressed KV cache often produces output similar to the full KV cache, allowing a long drafting horizon to amortize each full-KV swap. VeriCache applies to both long-context decoding and remote prefix caching, supports a broad family of token-dropping and quantization methods through a uniform compressor interface, and composes with traditional speculative decoding. Experimental results show that VeriCache achieves up to 4X higher throughput than full-KV inference while producing identical outputs.

Recent progress in Multimodal Large Language Models (MLLMs) has highlighted the critical roles of both the visual backbone and the underlying language model. While prior work has primarily focused on scaling these components to billions of parameters, the trade-offs between model size, architecture, and performance remain underexplored. Additionally, inconsistencies in training data and evaluation protocols have hindered direct comparisons, making it difficult to derive optimal design choices. In this paper, we introduce LLaVA-MORE, a new family of MLLMs that integrates recent language models with diverse visual backbones. To ensure fair comparisons, we employ a unified training protocol applied consistently across all architectures. Our analysis systematically explores both small- and medium-scale LLMs -- including Phi-4, LLaMA-3.1, and Gemma-2 -- to evaluate multimodal reasoning, generation, and instruction following, while examining the relationship between model size and performance. Beyond evaluating the LLM impact on final results, we conduct a comprehensive study of various visual encoders, ranging from CLIP-based architectures to alternatives such as DINOv2, SigLIP, and SigLIP2. Additional experiments investigate the effects of increased image resolution and variations in pre-training datasets. Overall, our results provide insights into the design of more effective MLLMs, offering a reproducible evaluation framework that facilitates direct comparisons and can guide future model development. Our source code and trained models are publicly available at: https://github.com/aimagelab/LLaVA-MORE.

LLM agents increasingly rely on retrieval buffers to store and reuse past experience, yet the cache management policies governing these buffers remain largely ad-hoc. We formalize this as an online semantic cache replacement problem with switching costs, where items are matched by embedding similarity and hit quality is continuous rather than binary. Through experiments on two datasets from MemoryBench-Full (LoCoMo, DialSim) with 8 replacement policies, we reveal a surprising finding: classic heuristics (LRU, LFU) consistently underperform the naive FIFO baseline on semantic workloads, due to the absence of temporal locality and frequency concentration. We propose SOLAR, a learning-augmented framework that derives modification timing from regret accumulation (achieving sim17\% modification rate) and content selection from Bayesian online learning over implicit retrieval feedback. We prove SOLAR achieves a constant competitive ratio leq 3, independent of cache size and horizon (vs.\ Ω(K) for FIFO), and eviction regret O(KTlog T), matching the Ω(KT) lower bound up to logarithmic factors. Experiments demonstrate 5--75\% relative improvement over FIFO at tight cache sizes, with a clearly characterized phase transition at the working set boundary. Synthetic experiments with 5000-item pools further reveal an inverted-U relationship between pool size and retrieval quality, justifying capacity constraints as a retrieval noise phenomenon rather than a storage limitation.

The rapid progress of large language models (LLMs) has advanced numerous applications, yet efficient single-batch inference remains vital for on-device intelligence. While FPGAs offer fine-grained data control and high energy efficiency, recent GPU optimizations have narrowed their advantage, especially under arithmetic-based computation. To overcome this, we leverage FPGAs' abundant on-chip memory to shift LLM inference from arithmetic- to memory-based computation through table lookups. We present LUT-LLM, the first FPGA accelerator enabling 1B+ LLM inference via vector-quantized memory operations. Our analysis identifies activation-weight co-quantization as the most effective scheme, supported by (1) bandwidth-aware parallel centroid search, (2) efficient 2D table lookups, and (3) a spatial-temporal hybrid design minimizing data caching. Implemented on an AMD V80 FPGA for a customized Qwen 3 1.7B model, LUT-LLM achieves 1.66x lower latency than AMD MI210 and 1.72x higher energy efficiency than NVIDIA A100, scaling to 32B models with 2.16x efficiency gain over A100.

Transformer-based large language models (LLMs) demonstrate impressive performance across various natural language processing tasks. Serving LLM inference for generating long contents, however, poses a challenge due to the enormous memory footprint of the transient state, known as the key-value (KV) cache, which scales with the sequence length and batch size. In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern offloading-based inference systems. InfiniGen leverages the key insight that a few important tokens that are essential for computing the subsequent attention layer in the Transformer can be speculated by performing a minimal rehearsal with the inputs of the current layer and part of the query weight and key cache of the subsequent layer. This allows us to prefetch only the essential KV cache entries (without fetching them all), thereby mitigating the fetch overhead from the host memory in offloading-based LLM serving systems. Our evaluation on several representative LLMs shows that InfiniGen improves the overall performance of a modern offloading-based system by up to 3.00x compared to prior KV cache management methods while offering substantially better model accuracy.

As large language models (LLMs) take on more complex tasks, their inputs incorporate longer contexts to respond to questions that require domain knowledge or user-specific conversational histories. Yet, using long contexts poses a challenge for responsive LLM systems, as nothing can be generated until all the contexts are fetched to and processed by the LLM. Existing systems optimize only the computation delay in context processing (e.g., by caching intermediate key-value features of the text context) but often cause longer network delays in context fetching (e.g., key-value features consume orders of magnitude larger bandwidth than the text context). This paper presents CacheGen to minimize the delays in fetching and processing contexts for LLMs. CacheGen reduces the bandwidth needed for transmitting long contexts' key-value (KV) features through a novel encoder that compresses KV features into more compact bitstream representations. The encoder combines adaptive quantization with a tailored arithmetic coder, taking advantage of the KV features' distributional properties, such as locality across tokens. Furthermore, CacheGen minimizes the total delay in fetching and processing a context by using a controller that determines when to load the context as compressed KV features or raw text and picks the appropriate compression level if loaded as KV features. We test CacheGen on three models of various sizes and three datasets of different context lengths. Compared to recent methods that handle long contexts, CacheGen reduces bandwidth usage by 3.7-4.3x and the total delay in fetching and processing contexts by 2.7-3x while maintaining similar LLM performance on various tasks as loading the text contexts.

In multi-agent systems utilizing Large Language Models (LLMs), communication between agents traditionally relies on natural language. This communication often includes the full context of the query so far, which can introduce significant prefill-phase latency, especially with long contexts. We introduce DroidSpeak, a novel framework to target this cross-LLM communication by leveraging the reuse of intermediate data, such as input embeddings (E-cache) and key-value caches (KV-cache). We efficiently bypass the need to reprocess entire contexts for fine-tuned versions of the same foundational model. This approach allows faster context integration while maintaining the quality of task performance. Experimental evaluations demonstrate DroidSpeak's ability to significantly accelerate inter-agent communication, achieving up to a 2.78x speedup in prefill latency with negligible loss in accuracy. Our findings underscore the potential to create more efficient and scalable multi-agent systems.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.

With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

Generating texts with a large language model (LLM) consumes massive amounts of memory. Apart from the already-large model parameters, the key/value (KV) cache that holds information about previous tokens in a sequence can grow to be even larger than the model itself. This problem is exacerbated in one of the current LLM serving frameworks which reserves the maximum sequence length of memory for the KV cache to guarantee generating a complete sequence as they do not know the output sequence length. This restricts us to use a smaller batch size leading to lower GPU utilization and above all, lower throughput. We argue that designing a system with a priori knowledge of the output sequence can mitigate this problem. To this end, we propose S^{3}, which predicts the output sequence length, schedules generation queries based on the prediction to increase device resource utilization and throughput, and handle mispredictions. Our proposed method achieves 6.49times throughput over those systems that assume the worst case for the output sequence length.

Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corpora. This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal when prompts fail to elicit quality variance among rollouts. To address this misalignment, we introduce LLM-as-a-Tutor, a framework that extends the LLM's role from judge to tutor: a single model serves as an examiner that pairwise-compares policy rollouts to detect non-challenging prompts, and as a generator that appends atomic constraints to them. This append-only design monotonically raises difficulty in step with the policy's capability, producing a self-calibrating training signal without external difficulty schedules. On three complex instruction-following benchmarks, our method consistently outperforms both policy-unaware baselines and prior policy-adaptive methods that adapt rubrics or rewrite prompts, suggesting prompt adaptation as a missing axis of policy-awareness in non-verifiable RL.

LLM-as-a-Judge has been widely adopted as an evaluation method and served as supervised rewards in model training. However, existing benchmarks for LLM-as-a-Judge are mainly relying on human-annotated ground truth, which introduces human bias that undermines the assessment of reliability and imposes scalability constraints. To overcome these limitations, we introduce Sage, a novel evaluation suite that assesses the quality of LLM judges without necessitating any human annotation. Inspired by axioms of rational choice theory, Sage introduces two new lenses for measuring LLM-as-a-Judge: local self-consistency (pair-wise preference stability) and global logical consistency (transitivity across a full set of preferences). We curate a dataset of 650 questions by combining structured benchmark problems with real-world user queries. Our experiments demonstrate both the stability of our metrics and their high correlation with supervised benchmarks like LLMBar and RewardBench2, confirming Sage's reliability as an evaluation suite for the robustness and accuracy of LLM-as-a-Judge. Based on Sage, we reveal that current state-of-the-art LLMs exhibit significant reliability problems when acting as judges in both scoring and pairwise settings; even the top-performing models, Gemini-2.5-Pro and GPT-5, fail to maintain consistent preferences in nearly a quarter of difficult cases. We attribute this to a new phenomenon called situational preference, which explains why explicit rubrics or criteria can help the model judge consistently across answer pairs. Our further analysis shows that finetuned LLM-as-a-Judge is a feasible method to boost performance, and the panel-based judge as well as deep reasoning can enhance the judging consistency. We also find substantial inconsistency in human judgments, which indicates that human annotation may not be a reliable gold standard.

As large language models are increasingly responsible for online content, concerns arise about the impact of repeatedly processing their own outputs. Inspired by the "broken telephone" effect in chained human communication, this study investigates whether LLMs similarly distort information through iterative generation. Through translation-based experiments, we find that distortion accumulates over time, influenced by language choice and chain complexity. While degradation is inevitable, it can be mitigated through strategic prompting techniques. These findings contribute to discussions on the long-term effects of AI-mediated information propagation, raising important questions about the reliability of LLM-generated content in iterative workflows.

Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identify verification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduce LLM-as-a-Verifier, a general-purpose verification framework that provides fine-grained feedback for agentic tasks without requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions, LLM-as-a-Verifier computes the expectation over the distribution of scoring token logits to generate continuous scores. This probabilistic formulation enables verification to scale along multiple dimensions: (1) score granularity, (2) repeated evaluation, and (3) criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scaling repeated evaluation and criteria decomposition consistently lead to additional gains in verification accuracy through variance and complexity reduction. We further introduce a cost-efficient ranking algorithm for selecting the best solution among candidates using the verifier's continuous scores. LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). Beyond verification, the fine-grained signals from LLM-as-a-Verifier can also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show that LLM-as-a-Verifier can provide dense feedback for RL, improving the sample efficiency of SAC and GRPO on robotics and mathematical reasoning benchmarks.

LLM-as-a-Judge, which generates chain-of-thought (CoT) judgments, has become a widely adopted auto-evaluation method. However, its reliability is compromised by the CoT reasoning's inability to capture comprehensive and deeper details, often leading to incomplete outcomes. Existing methods mainly rely on majority voting or criteria expansion, which is insufficient to address the limitation in CoT. We propose Crowd-based Comparative Evaluation, which introduces additional crowd responses to compare with the candidate responses, thereby exposing deeper and more comprehensive details within the candidate responses. This process effectively guides LLM-as-a-Judge to provide a more detailed CoT judgment. Extensive experiments demonstrate that our approach enhances evaluation reliability, achieving an average accuracy gain of 6.7% across five benchmarks. Moreover, our method produces higher-quality CoTs that facilitate judge distillation and exhibit superior performance in rejection sampling for supervised fine-tuning (SFT), referred to as crowd rejection sampling, thereby enabling more efficient SFT. Our analysis confirms that CoTs generated by ours are more comprehensive and of higher quality, and evaluation accuracy improves as inference scales.

LLM-as-a-Judge and reward models are widely used alternatives of multiple-choice questions or human annotators for large language model (LLM) evaluation. Their efficacy shines in evaluating long-form responses, serving a critical role as evaluators of leaderboards and as proxies to align LLMs via reinforcement learning. However, despite their popularity, their effectiveness outside of English remains largely unexplored. In this paper, we conduct a comprehensive analysis on automated evaluators, reporting key findings on their behavior in a non-English environment. First, we discover that English evaluation capabilities significantly influence language-specific capabilities, often more than the language proficiency itself, enabling evaluators trained in English to easily transfer their skills to other languages. Second, we identify critical shortcomings, where LLMs fail to detect and penalize errors, such as factual inaccuracies, cultural misrepresentations, and the presence of unwanted language. Finally, we release Kudge, the first non-English meta-evaluation dataset containing 5,012 human annotations in Korean.

LLM-as-a-Judge (LLMaaJ) now underpins scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover a conversation's latent objective and know when that inference is trustworthy? LLMs degrade under irrelevant or long context; multi-turn jailbreaks further hide goals across turns. We introduce ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must return a one-sentence base objective and a self-reported confidence. Accuracy is computed via LLM-judge semantic similarity to gold objectives, converted to binary correctness by a single human-aligned threshold calibrated once on N = 100 items (tau^*=0.61). Metacognition is evaluated with ECE, Brier, Wrong-at-High-Conf, and risk-coverage. Across gpt-4.1, claude-sonnet-4, and Qwen3-235B-A22B-FP8 on SafeMTData_Attack600, SafeMTData_1K, MHJ, and CoSafe, claude-sonnet-4 attains the best objective-extraction accuracy (0.515) and calibration (ECE 0.296; Brier 0.324); gpt-4.1 and Qwen3-235B-A22B-FP8 tie at 0.441 but are overconfident (mean confidence approx0.88 vs. accuracy approx0.44; Wrong-at-0.90 approx48-52%). Performance varies by dataset (approx0.167-0.865). ObjexMT thus supplies an actionable test for LLM judges: when objectives are not explicit, judges often misinfer them with high confidence. We recommend exposing objectives when feasible and gating decisions by confidence otherwise. Code and data at https://github.com/hyunjun1121/ObjexMT_dataset.

Providers of LLM-as-a-service have predominantly adopted a simple pricing model: users pay a fixed price per token. Consequently, one may think that the price two different users would pay for the same output string under the same input prompt is the same. In our work, we show that, surprisingly, this is not (always) true. We find empirical evidence that, particularly for non-english outputs, both proprietary and open-weights LLMs often generate the same (output) string with multiple different tokenizations, even under the same input prompt, and this in turn leads to arbitrary price variation. To address the problem of tokenization multiplicity, we introduce canonical generation, a type of constrained generation that restricts LLMs to only generate canonical tokenizations -- the unique tokenization in which each string is tokenized during the training process of an LLM. Further, we introduce an efficient sampling algorithm for canonical generation based on the Gumbel-Max trick. Experiments on a variety of natural language tasks demonstrate that our sampling algorithm for canonical generation is comparable to standard sampling in terms of performance and runtime, and it solves the problem of tokenization multiplicity.

Rubric-based scoring has become a widely used paradigm in model evaluation, typically with LLM-as-a-Judge (LaaJ) for rubric scoring. However, the reliability of LaaJ for rubric scoring remains underexplored. This concern is especially pronounced in agentic scenarios, where long, complex outputs further challenge reliable scoring. To address this, we conduct a systematic meta-evaluation of LaaJ reliability for rubric verification. We introduce RuVerBench, the first benchmark for assessing LaaJ reliability in rubric verification for agentic scenarios. RuVerBench covers two prevalent agentic domains, deep research and agentic coding, with 2,458 instances, each containing a model-generated output, a rubric, and a human-annotated label indicating whether the output satisfies the rubric. Using RuVerBench, we evaluate numerous frontier LLMs and find that even the most advanced models achieve strong performance but still exhibit substantial noise. We further analyze the impact of key LaaJ strategies, including prompt design, batching, and majority voting, on rubric verification. We find that weaker models are more sensitive to prompt variations, batched verification presents a trade-off between accuracy and efficiency, and majority voting yields effective but diminishing returns. We have released our dataset and code to facilitate future research: https://github.com/THU-KEG/RuVerBench.

Neural Architecture Search (NAS) aims to automatically discover high-performing deep neural network (DNN) architectures. However, conventional algorithm-driven NAS relies on carefully hand-crafted search spaces to ensure executability, which restricts open-ended exploration. Recent coding-based agentic approaches using large language models (LLMs) reduce manual design, but current LLMs struggle to reliably generate complex, valid architectures, and their proposals are often biased toward a narrow set of patterns observed in their training data. To bridge reliable algorithmic search with powerful LLM assistance, we propose LLMasTool, a hierarchical tree-based NAS framework for stable and open-ended model evolution. Our method automatically extracts reusable modules from arbitrary source code and represents full architectures as hierarchical trees, enabling evolution through reliable tree transformations rather than code generation. At each evolution step, coarse-level planning is governed by a diversity-guided algorithm that leverages Bayesian modeling to improve exploration efficiency, while the LLM resolves the remaining degrees of freedom to ensure a meaningful evolutionary trajectory and an executable generated architecture. With this formulation, instead of fully agentic LLM approaches, our method explores diverse directions beyond the inherent biases in the LLM. Our method improves over existing NAS methods by 0.69, 1.83, and 2.68 points on CIFAR-10, CIFAR-100, and ImageNet16-120, demonstrating its effectiveness.

LLM-as-a-Judge has been widely adopted across various research and practical applications, yet the robustness and reliability of its evaluation remain a critical issue. A core challenge it faces is bias, which has primarily been studied in terms of known biases and their impact on evaluation outcomes, while automated and systematic exploration of potential unknown biases is still lacking. Nevertheless, such exploration is crucial for enhancing the robustness and reliability of evaluations. To bridge this gap, we propose BiasScope, a LLM-driven framework for automatically and at scale discovering potential biases that may arise during model evaluation. BiasScope can uncover potential biases across different model families and scales, with its generality and effectiveness validated on the JudgeBench dataset. It overcomes the limitations of existing approaches, transforming bias discovery from a passive process relying on manual effort and predefined bias lists into an active and comprehensive automated exploration. Moreover, based on BiasScope, we propose JudgeBench-Pro, an extended version of JudgeBench and a more challenging benchmark for evaluating the robustness of LLM-as-a-judge. Strikingly, even powerful LLMs as evaluators show error rates above 50\% on JudgeBench-Pro, underscoring the urgent need to strengthen evaluation robustness and to mitigate potential biases further.

LLM-as-a-judge has become a promising paradigm for using large language models (LLMs) to evaluate natural language generation (NLG), but the uncertainty of its evaluation remains underexplored. This lack of reliability may limit its deployment in many applications. This work presents the first framework to analyze the uncertainty by offering a prediction interval of LLM-based scoring via conformal prediction. Conformal prediction constructs continuous prediction intervals from a single evaluation run, and we design an ordinal boundary adjustment for discrete rating tasks. We also suggest a midpoint-based score within the interval as a low-bias alternative to raw model score and weighted average. We perform extensive experiments and analysis, which show that conformal prediction can provide valid prediction interval with coverage guarantees. We also explore the usefulness of interval midpoint and judge reprompting for better judgment.

LLM-as-a-Judge refers to the automatic modeling of preferences for responses generated by Large Language Models (LLMs), which is of significant importance for both LLM evaluation and reward modeling. Although generative LLMs have made substantial progress in various tasks, their performance as LLM-Judge still falls short of expectations. In this work, we propose Think-J, which improves generative LLM-as-a-Judge by learning how to think. We first utilized a small amount of curated data to develop the model with initial judgment thinking capabilities. Subsequently, we optimize the judgment thinking traces based on reinforcement learning (RL). We propose two methods for judgment thinking optimization, based on offline and online RL, respectively. The offline RL requires training a critic model to construct positive and negative examples for learning. The online method defines rule-based reward as feedback for optimization. Experimental results showed that our approach can significantly enhance the evaluation capability of generative LLM-Judge, surpassing both generative and classifier-based LLM-Judge without requiring extra human annotations.

LLM-as-a-Judge has been widely utilized as an evaluation method in various benchmarks and served as supervised rewards in model training. However, despite their excellence in many domains, potential issues are under-explored, undermining their reliability and the scope of their utility. Therefore, we identify 12 key potential biases and propose a new automated bias quantification framework-CALM-which systematically quantifies and analyzes each type of bias in LLM-as-a-Judge by using automated and principle-guided modification. Our experiments cover multiple popular language models, and the results indicate that while advanced models have achieved commendable overall performance, significant biases persist in certain specific tasks. Empirical results suggest that there remains room for improvement in the reliability of LLM-as-a-Judge. Moreover, we also discuss the explicit and implicit influence of these biases and give some suggestions for the reliable application of LLM-as-a-Judge. Our work highlights the need for stakeholders to address these issues and remind users to exercise caution in LLM-as-a-Judge applications.

Large language models (LLMs) are increasingly used in academic peer review, yet their reliability, alignment with human judgment, and robustness to adversarial attacks remain poorly understood. We present a systematic benchmark of LLM-as-a-Reviewer on 898 papers stratified from NeurIPS and ICLR, evaluating 12 LLMs along three axes: rating calibration, divergence from human reviewers, and resistance to prompt injection embedded via an invisible font-mapping attack. We find that LLMs systematically overrate weaker submissions and diverge from humans in topical emphasis, under-flagging Clarity and over-flagging Reproducibility, while producing reviews two to three times longer with lower lexical diversity and a more standardized vocabulary. Prompt injection remains highly effective. Simple hidden instructions can promote low-scoring papers to acceptance-level ratings in a substantial fraction of cases, with effectiveness varying sharply across model families. While LLMs offer utility in structuring evaluations, their integration into peer review requires safeguards against both intrinsic biases and adversarial risks.

Large language models (LLMs) are increasingly deployed across healthcare applications, including clinical documentation, diagnostic reasoning, medicine recommendation, and medical education. Their outputs are largely unstructured clinical text, which is difficult to reliably evaluate at scale. LLM-as-a-Judge, in which an LLM evaluates another system's output against task-specific criteria, offers a scalable alternative and is increasingly used in clinical evaluation, yet its validity in healthcare remains underexamined. Existing reviews focus on general-purpose LLM evaluation or on risk framework, rather than systematically characterizing how LLM-as-a-Judge is applied in healthcare and how well their judgments align with human experts. We therefore conduct a PRISMA-guided comprehensive review of LLM-as-a-Judge applications in healthcare, searching five databases for studies published between January 2023 and February 2026. After screening 541 records, 134 studies meet the eligibility and are coded by health scenario, judge configuration, technical approach, and validation design. LLM-as-a-Judge is concentrated in clinical decision support, clinical natural language processing (NLP), medical knowledge and question answering (QA), and medical communication. OpenAI models are the most frequently used judges, and prompt engineering appears in nearly all studies, with ensemble, multi-agent, and retrieval-augmented designs as common extensions. Among studies reporting human validation, LLM judges often show moderate to strong alignment with expert judgments, although reliability varies substantially across tasks. Overall, this review positions LLM-as-a-Judge as a promising framework for scalable healthcare AI evaluation, while emphasizing that its clinical value depends on model design and rigorous validation.

Large Language Models are increasingly used as judges to evaluate code artifacts when exhaustive human review or executable test coverage is unavailable. LLM-judge is increasingly relevant in agentic software engineering workflows, where it can help rank candidate solutions and guide patch selection. While attractive for scale, current practice lacks a principled account of reliability and bias: repeated evaluations of the same case can disagree; small prompt edits can swing outcomes; and seemingly semantics-preserving, human-equivalent perturbations may elicit divergent verdicts. This paper studies LLM-as-a-Judge for code through a measurement-first lens. We analyze two pointwise judging regimes across code generation, code repair task, and test generation, and we systematically probe prompt-induced biases. Our study considers difficulty levels for repeated runs and controlled prompt interventions that isolate one presentation cue at a time, and it evaluates judges using consistency and sensitivity to bias. We find that judge decisions are highly sensitive to prompt biases even when the underlying code snippet is unchanged. Across all three tasks, several biases systematically shift preferences toward the option favored by the prompt, improving accuracy when that option aligns with the gold answer but substantially reducing it otherwise. In some settings, these effects are large enough to change task-level conclusions and alter relative model rankings. These findings show that reported judge performance may reflect prompt artifacts rather than stable assessment ability, posing a direct threat to the validity and reproducibility of code evaluation. We therefore argue that LLM-as-a-Judge studies should report bias sensitivity alongside accuracy and incorporate explicit controls to support more trustworthy model comparison in software engineering.

LLM-as-a-judge has emerged as a cornerstone technique for evaluating large language models by leveraging LLM reasoning to score prompt-response pairs. Since LLM judgments are stochastic, practitioners commonly query each pair multiple times to estimate mean scores accurately. This raises a critical challenge: given a fixed computational budget B, how to optimally allocate queries across K prompt-response pairs to minimize estimation error? % We present a principled variance-adaptive approach leveraging multi-armed bandit theory and concentration inequalities. Our method dynamically allocates queries based on estimated score variances, concentrating resources where uncertainty is highest. Further, our algorithm is shown to achieve a worst-case score-estimation error of Oleft(frac{sum_{i=1^K σ_i^2}{B}}right), σ_i^2 being the unknown score variance for pair i in [K] with near-optimal budget allocation. % Experiments on Summarize-From-Feedback and HelpSteer2 demonstrate that our method significantly outperforms uniform allocation, reducing worst-case estimation error while maintaining identical budgets. Our work establishes a theoretical foundation for efficient LLM evaluation with practical implications for AI safety, model alignment, and automated assessment at scale.

LLM-as-a-Judge has emerged as a promising alternative to human evaluators across various tasks, yet inherent biases - particularly position bias, the tendency to favor solutions based on their position within the prompt - compromise its reliability. This exploratory study evaluates position bias in LLM judges across pairwise and list-wise comparison settings, introducing three metrics: repetition stability, position consistency, and preference fairness. Our experiments, involving 15 LLM judges across MTBench and DevBench with 22 tasks and approximately 40 solution-generating models, result in over 150,000 evaluation instances. We identify Judge-Level, Candidate-Level, and Task-Level factors contributing to bias. The findings confirm that position bias is not due to random chance and varies significantly across judges and tasks. While position bias is weakly influenced by the length of prompt components, it is strongly affected by the quality gap between solutions. Our agreement and disagreement analysis among judges further provides insights into the distribution of judging difficulty across the dataset, and highlights the potential for dataset modifications.

The rapid integration of Large Language Models (LLMs) into software engineering (SE) has revolutionized tasks like code generation, producing a massive volume of software artifacts. This surge has exposed a critical bottleneck: the lack of scalable, reliable methods to evaluate these outputs. Human evaluation is costly and time-consuming, while traditional automated metrics like BLEU fail to capture nuanced quality aspects. In response, the LLM-as-a-Judge paradigm - using LLMs for automated evaluation - has emerged. This approach leverages the advanced reasoning of LLMs, offering a path toward human-like nuance at automated scale. However, LLM-as-a-Judge research in SE is still in its early stages. This forward-looking SE 2030 paper aims to steer the community toward advancing LLM-as-a-Judge for evaluating LLM-generated software artifacts. We provide a literature review of existing SE studies, analyze their limitations, identify key research gaps, and outline a detailed roadmap. We envision these frameworks as reliable, robust, and scalable human surrogates capable of consistent, multi-faceted artifact evaluation by 2030. Our work aims to foster research and adoption of LLM-as-a-Judge frameworks, ultimately improving the scalability of software artifact evaluation.

LLM-as-a-Judge uses a large language model (LLM) to select the best response from a set of candidates for a given question. LLM-as-a-Judge has many applications such as LLM-powered search, reinforcement learning with AI feedback (RLAIF), and tool selection. In this work, we propose JudgeDeceiver, an optimization-based prompt injection attack to LLM-as-a-Judge. JudgeDeceiver injects a carefully crafted sequence into an attacker-controlled candidate response such that LLM-as-a-Judge selects the candidate response for an attacker-chosen question no matter what other candidate responses are. Specifically, we formulate finding such sequence as an optimization problem and propose a gradient based method to approximately solve it. Our extensive evaluation shows that JudgeDeceive is highly effective, and is much more effective than existing prompt injection attacks that manually craft the injected sequences and jailbreak attacks when extended to our problem. We also show the effectiveness of JudgeDeceiver in three case studies, i.e., LLM-powered search, RLAIF, and tool selection. Moreover, we consider defenses including known-answer detection, perplexity detection, and perplexity windowed detection. Our results show these defenses are insufficient, highlighting the urgent need for developing new defense strategies. Our implementation is available at this repository: https://github.com/ShiJiawenwen/JudgeDeceiver.

LLM-as-a-Judge models generate chain-of-thought (CoT) sequences intended to capture the step-bystep reasoning process that underlies the final evaluation of a response. However, due to the lack of human annotated CoTs for evaluation, the required components and structure of effective reasoning traces remain understudied. Consequently, previous approaches often (1) constrain reasoning traces to hand-designed components, such as a list of criteria, reference answers, or verification questions and (2) structure them such that planning is intertwined with the reasoning for evaluation. In this work, we propose EvalPlanner, a preference optimization algorithm for Thinking-LLM-as-a-Judge that first generates an unconstrained evaluation plan, followed by its execution, and then the final judgment. In a self-training loop, EvalPlanner iteratively optimizes over synthetically constructed evaluation plans and executions, leading to better final verdicts. Our method achieves a new state-of-the-art performance for generative reward models on RewardBench (with a score of 93.9), despite being trained on fewer amount of, and synthetically generated, preference pairs. Additional experiments on other benchmarks like RM-Bench, JudgeBench, and FollowBenchEval further highlight the utility of both planning and reasoning for building robust LLM-as-a-Judge reasoning models.

As large language models (LLMs) continue to advance, reliable evaluation methods are essential particularly for open-ended, instruction-following tasks. LLM-as-a-Judge enables automatic evaluation using LLMs as evaluators, but its reliability remains uncertain. In this work, we analyze key factors affecting its trustworthiness, focusing on alignment with human judgments and evaluation consistency. Using BIGGENBench and EvalBiasBench, we study the effects of evaluation design, decoding strategies, and Chain-of-Tought (CoT) reasoning in evaluation. Our results show that evaluation criteria are critical for reliability, non-deterministic sampling improves alignment with human preferences over deterministic evaluation, and CoT reasoning offers minimal gains when clear evaluation criteria are present.

Prompting large language models (LLMs) to evaluate generated text, known as LLM-as-a-judge, has become a standard evaluation approach in natural language generation (NLG), but is primarily used as a quantitative tool, i.e. with numerical scores as main outputs. In this work, we propose LLM-as-a-qualitative-judge, an LLM-based evaluation approach with the main output being a structured report of common issue types in the NLG system outputs. Our approach is targeted at providing developers with meaningful insights on what improvements can be done to a given NLG system and consists of two main steps, namely open-ended per-instance issue analysis and clustering of the discovered issues using an intuitive cumulative algorithm. We also introduce a strategy for evaluating the proposed approach, coupled with ~300 annotations of issues in instances from 12 NLG datasets. Our results show that LLM-as-a-qualitative-judge correctly recognizes instance-specific issues in 2/3 cases and is capable of producing error type reports resembling the reports composed by human annotators. Our code and data are publicly available at https://github.com/tunde-ajayi/llm-as-a-qualitative-judge.

Large Language Models (LLMs) are increasingly employed as evaluators (LLM-as-a-Judge) for assessing the quality of machine-generated text. This paradigm offers scalability and cost-effectiveness compared to human annotation. However, the reliability and security of such systems, particularly their robustness against adversarial manipulations, remain critical concerns. This paper investigates the vulnerability of LLM-as-a-Judge architectures to prompt-injection attacks, where malicious inputs are designed to compromise the judge's decision-making process. We formalize two primary attack strategies: Comparative Undermining Attack (CUA), which directly targets the final decision output, and Justification Manipulation Attack (JMA), which aims to alter the model's generated reasoning. Using the Greedy Coordinate Gradient (GCG) optimization method, we craft adversarial suffixes appended to one of the responses being compared. Experiments conducted on the MT-Bench Human Judgments dataset with open-source instruction-tuned LLMs (Qwen2.5-3B-Instruct and Falcon3-3B-Instruct) demonstrate significant susceptibility. The CUA achieves an Attack Success Rate (ASR) exceeding 30\%, while JMA also shows notable effectiveness. These findings highlight substantial vulnerabilities in current LLM-as-a-Judge systems, underscoring the need for robust defense mechanisms and further research into adversarial evaluation and trustworthiness in LLM-based assessment frameworks.

Recently, large language models (LLMs) have been deployed to tackle various software engineering (SE) tasks like code generation, significantly advancing the automation of SE tasks. However, assessing the quality of these LLM-generated code and text remains challenging. The commonly used Pass@k metric necessitates extensive unit tests and configured environments, demands a high labor cost, and is not suitable for evaluating LLM-generated text. Conventional metrics like BLEU, which measure only lexical rather than semantic similarity, have also come under scrutiny. In response, a new trend has emerged to employ LLMs for automated evaluation, known as LLM-as-a-judge. These LLM-as-a-judge methods are claimed to better mimic human assessment than conventional metrics without relying on high-quality reference answers. Nevertheless, their exact human alignment in SE tasks remains unexplored. In this paper, we empirically explore LLM-as-a-judge methods for evaluating SE tasks, focusing on their alignment with human judgments. We select seven LLM-as-a-judge methods that utilize general-purpose LLMs, alongside two LLMs specifically fine-tuned for evaluation. After generating and manually scoring LLM responses on three recent SE datasets of code translation, code generation, and code summarization, we then prompt these methods to evaluate each response. Finally, we compare the scores generated by these methods with human evaluation. The results indicate that output-based methods reach the highest Pearson correlation of 81.32 and 68.51 with human scores in code translation and generation, achieving near-human evaluation, noticeably outperforming ChrF++, one of the best conventional metrics, at 34.23 and 64.92. Such output-based methods prompt LLMs to output judgments directly, and exhibit more balanced score distributions that resemble human score patterns. Finally, we provide...

LLM-as-a-Judge is a framework that uses an LLM (large language model) to evaluate the quality of natural language text - typically text that is also generated by an LLM. This framework holds great promise due to its relative low-cost, ease of use, and strong correlations with human stylistic preferences. However, LLM Judges have been shown to exhibit biases that can distort their judgments. We evaluate how well LLM Judges can grade whether a given response to a conversational question is correct, an ability crucial to soundly estimating the overall response quality. To do so, we create and publicly release a human-annotated dataset with labels of correctness for 1,200 LLM responses. We source questions from a combination of existing datasets and a novel, challenging benchmark (BFF-Bench) created for this analysis. We demonstrate a strong connection between an LLM's ability to correctly answer a question and grade responses to that question. Although aggregate level statistics might imply a judge has high agreement with human annotators, it will struggle on the subset of questions it could not answer. To address this issue, we recommend a simple solution: provide the judge with a correct, human-written reference answer. We perform an in-depth analysis on how reference quality can affect the performance of an LLM Judge. We show that providing a weaker judge (e.g. Qwen 2.5 7B) with higher quality references reaches better agreement with human annotators than a stronger judge (e.g. GPT-4o) with synthetic references.

LLM-as-a-Judge leverages the generative and reasoning capabilities of large language models (LLMs) to evaluate LLM responses across diverse scenarios, providing accurate preference signals. This approach plays a vital role in aligning LLMs with human values, ensuring ethical and reliable AI outputs that align with societal norms. Recent studies have raised many methods to train LLM as generative judges, but most of them are data consuming or lack accuracy, and only focus on LLM's judge ability. In this work, we regard judge ability as a general ability of LLM and implement a two-stage training approach, comprising supervised fine-tuning (SFT) warm-up and direct preference optimization (DPO) enhancement, to achieve judge style adaptation and improve judgment accuracy. Additionally, we introduce an efficient data synthesis method to generate judgmental content. Experimental results demonstrate that our approach, utilizing only about 2% to 40% of the data required by other methods, achieves SOTA performance on RewardBench. Furthermore, our training method enhances the general capabilities of the model by constructing complicated judge task, and the judge signals provided by our model have significantly enhanced the downstream DPO training performance of our internal models in our test to optimize policy model with Judge Model. We also open-source our model weights and training data to facilitate further research.

Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of "LLM-as-a-Judge," where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of LLM-as-a-Judge systems remains a significant challenge that requires careful design and standardization. This paper provides a comprehensive survey of LLM-as-a-Judge, addressing the core question: How can reliable LLM-as-a-Judge systems be built? We explore strategies to enhance reliability, including improving consistency, mitigating biases, and adapting to diverse assessment scenarios. Additionally, we propose methodologies for evaluating the reliability of LLM-as-a-Judge systems, supported by a novel benchmark designed for this purpose. To advance the development and real-world deployment of LLM-as-a-Judge systems, we also discussed practical applications, challenges, and future directions. This survey serves as a foundational reference for researchers and practitioners in this rapidly evolving field.

LLM-as-a-judge models have been used for evaluating both human and AI generated content, specifically by providing scores and rationales. Rationales, in addition to increasing transparency, help models learn to calibrate its judgments. Enhancing a model's rationale can therefore improve its calibration abilities and ultimately the ability to score content. We introduce Self-Rationalization, an iterative process of improving the rationales for the judge models, which consequently improves the score for fine-grained customizable scoring criteria (i.e., likert-scale scoring with arbitrary evaluation criteria). Self-rationalization works by having the model generate multiple judgments with rationales for the same input, curating a preference pair dataset from its own judgements, and iteratively fine-tuning the judge via DPO. Intuitively, this approach allows the judge model to self-improve by learning from its own rationales, leading to better alignment and evaluation accuracy. After just two iterations -- while only relying on examples in the training set -- human evaluation shows that our judge model learns to produce higher quality rationales, with a win rate of 62% on average compared to models just trained via SFT on rationale . This judge model also achieves high scoring accuracy on BigGen Bench and Reward Bench, outperforming even bigger sized models trained using SFT with rationale, self-consistency or best-of-N sampling by 3% to 9%.

LLM-as-a-Judge has been widely applied to evaluate and compare different LLM alignmnet approaches (e.g., RLHF and DPO). However, concerns regarding its reliability have emerged, due to LLM judges' biases and inconsistent decision-making. Previous research has developed evaluation frameworks to assess reliability of LLM judges and their alignment with human preferences. However, the employed evaluation metrics often lack adequate explainability and fail to address LLM internal inconsistency. Additionally, existing studies inadequately explore the impact of various prompt templates when applying LLM-as-a-Judge methods, leading to potentially inconsistent comparisons between different alignment algorithms. In this work, we systematically evaluate LLM-as-a-Judge on alignment tasks by defining more theoretically interpretable evaluation metrics and explicitly mitigating LLM internal inconsistency from reliability metrics. We develop an open-source framework to evaluate, compare, and visualize the reliability and alignment of LLM judges, which facilitates practitioners to choose LLM judges for alignment tasks. In the experiments, we examine effects of diverse prompt templates on LLM-judge reliability and also demonstrate our developed framework by comparing various LLM judges on two common alignment datasets (i.e., TL;DR Summarization and HH-RLHF-Helpfulness). Our results indicate a significant impact of prompt templates on LLM judge performance, as well as a mediocre alignment level between the tested LLM judges and human evaluators.

With the rapid development and widespread application of Large Language Models (LLMs), the use of Machine-Generated Text (MGT) has become increasingly common, bringing with it potential risks, especially in terms of quality and integrity in fields like news, education, and science. Current research mainly focuses on purely MGT detection without adequately addressing mixed scenarios, including AI-revised Human-Written Text (HWT) or human-revised MGT. To tackle this challenge, we define mixtext, a form of mixed text involving both AI and human-generated content. Then, we introduce MixSet, the first dataset dedicated to studying these mixtext scenarios. Leveraging MixSet, we executed comprehensive experiments to assess the efficacy of prevalent MGT detectors in handling mixtext situations, evaluating their performance in terms of effectiveness, robustness, and generalization. Our findings reveal that existing detectors struggle to identify mixtext, particularly in dealing with subtle modifications and style adaptability. This research underscores the urgent need for more fine-grain detectors tailored for mixtext, offering valuable insights for future research. Code and Models are available at https://github.com/Dongping-Chen/MixSet.

Large language models are strong sequence predictors, yet standard inference relies on immutable context histories. After making an error at generation step t, the model lacks an updatable memory mechanism that improves predictions for step t+1. We propose LLM-as-RNN, an inference-only framework that turns a frozen LLM into a recurrent predictor by representing its hidden state as natural-language memory. This state, implemented as a structured system-prompt summary, is updated at each timestep via feedback-driven text rewrites, enabling learning without parameter updates. Under a fixed token budget, LLM-as-RNN corrects errors and retains task-relevant patterns, effectively performing online learning through language. We evaluate the method on three sequential benchmarks in healthcare, meteorology, and finance across Llama, Gemma, and GPT model families. LLM-as-RNN significantly outperforms zero-shot, full-history, and MemPrompt baselines, improving predictive accuracy by 6.5% on average, while producing interpretable, human-readable learning traces absent in standard context accumulation.

Accurate evaluation is central to the large language model (LLM) ecosystem, guiding model selection and downstream adoption across diverse use cases. In practice, however, evaluating generative outputs typically relies on rigid lexical methods to extract and assess answers, which can conflate a model's true problem-solving ability with its compliance with predefined formatting guidelines. While recent LLM-as-a-Judge approaches mitigate this issue by assessing semantic correctness rather than strict structural conformity, they also introduce substantial computational overhead, making evaluation costly. In this work, we first systematically investigate the limitations of lexical evaluation through a large-scale empirical study spanning 36 models and 15 downstream tasks, demonstrating that such methods correlate poorly with human judgments. To address this limitation, we introduce BERT-as-a-Judge, an encoder-driven approach for assessing answer correctness in reference-based generative settings, robust to variations in output phrasing, and requiring only lightweight training on synthetically annotated question-candidate-reference triplets. We show that it consistently outperforms the lexical baseline while matching the performance of much larger LLM judges, providing a compelling tradeoff between the two and enabling reliable, scalable evaluation. Finally, through extensive experimentation, we provide detailed insights into BERT-as-a-Judge's performance to offer practical guidance for practitioners, and release all project artifacts to foster downstream adoption.

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

Large Language Models have been recently exploited as judges for complex natural language processing tasks, such as Q&A. The basic idea is to delegate to an LLM the assessment of the "quality" of the output provided by an automated technique for tasks for which: (i) quantitative metrics would only tell part of the story, and; (ii) a large-scale human-based evaluation would be too expensive. LLMs-as-a-judge, if proven effective for a specific task, can also unlock new possibilities for automation, with several LLMs proposing a solution for a given instance of the task and others judging and deciding what is the best output to show the user. We study the effectiveness of LLMs-as-a-judge for two code-related tasks, namely code generation and code summarization. The rationale for choosing these tasks is two-fold. First, quantitative metrics are usually not enough for the assessment of code summarizers/generators. For example, it is well documented that metrics such as BLEU are quite weak proxies for the quality of the generated summaries. Second, even state-of-the-art techniques still struggle with handling complex instances of these tasks, making them good candidates for benefiting from more advanced solutions envisioning collaboration among LLMs. For code generation, we check whether eight LLMs are able to judge the correctness of 1,405 Java methods and 1,281 Python functions generated by the same LLMs or implemented by humans. For code summarization, we compare the judgment of five LLMs to those provided by nine humans for ~1.2k summaries, related to both Java and Python functions. Our findings show that GPT-4-turbo is the best LLM in terms of judging capabilities for both tasks, with "smaller" LLMs featuring tens of billions parameters not being able to cope with judging tasks. However, even the best-performing LLM frequently misjudges the correctness of the code and summary quality.

LLM as judge systems used to assess text quality code correctness and argument strength are vulnerable to prompt injection attacks. We introduce a framework that separates content author attacks from system prompt attacks and evaluate five models Gemma 3.27B Gemma 3.4B Llama 3.2 3B GPT 4 and Claude 3 Opus on four tasks with various defenses using fifty prompts per condition. Attacks achieved up to seventy three point eight percent success smaller models proved more vulnerable and transferability ranged from fifty point five to sixty two point six percent. Our results contrast with Universal Prompt Injection and AdvPrompter We recommend multi model committees and comparative scoring and release all code and datasets

Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, such as position and verbosity biases and limited reasoning ability, and propose solutions to migrate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80\% agreement, the same level of agreement between humans. Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain. Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA/Vicuna. We will publicly release 80 MT-bench questions, 3K expert votes, and 30K conversations with human preferences from Chatbot Arena.

The ``LLM-as-a-judge'' paradigm has become a standard method for evaluating open-ended generation. To address the quadratic scalability costs of pairwise comparisons, popular benchmarks like Arena-Hard and AlpacaEval compare all models against a single anchor. However, despite its widespread use, the impact of anchor selection on the reliability of the results remains largely unexplored. In this work, we systematically investigate the effect of anchor selection by evaluating 22 different anchors on the Arena-Hard-v2.0 dataset. We find that the choice of anchor is critical: a poor anchor can dramatically reduce correlation with human rankings. We identify that common anchor choices (best-performing and worst-performing models) make poor anchors. Because these extreme anchors are consistently better or worse than all other models, they are seldom indicative of the relative ranking of the models. We further quantify the effect size of anchor selection, showing it is comparable to the selection of a judge model. We conclude with actionable recommendations. First, we conduct a power analysis, and compute sufficient benchmark sizes for anchor-based evaluation, finding that standard benchmark sizes are insufficient for pairwise evaluation and fail to distinguish between competitive models reliably. Second, we provide guidelines for selecting informative anchors to ensure reliable and efficient evaluation practices.

The LLM-as-a-Judge paradigm shows promise for evaluating generative content but lacks reliability in reasoning-intensive scenarios, such as programming. Inspired by recent advances in reasoning models and shifts in scaling laws, we pioneer bringing test-time computation into LLM-as-a-Judge, proposing MCTS-Judge, a resource-efficient, System-2 thinking framework for code correctness evaluation. MCTS-Judge leverages Monte Carlo Tree Search (MCTS) to decompose problems into simpler, multi-perspective evaluations. Through a node-selection strategy that combines self-assessment based on historical actions in the current trajectory and the Upper Confidence Bound for Trees based on prior rollouts, MCTS-Judge balances global optimization and refinement of the current trajectory. We further designed a high-precision, unit-test-level reward mechanism to encourage the Large Language Model (LLM) to perform line-by-line analysis. Extensive experiments on three benchmarks and five LLMs demonstrate the effectiveness of MCTS-Judge, which improves the base model's accuracy from 41% to 80%, surpassing the o1-series models with 3x fewer tokens. Further evaluations validate the superiority of its reasoning trajectory in logic, analytics, thoroughness, and overall quality, while revealing the test-time scaling law of the LLM-as-a-Judge paradigm.

The "LLM-as-a-judge" paradigm employs Large Language Models (LLMs) as annotators and evaluators in tasks traditionally performed by humans. LLM annotations are widely used, not only in NLP research but also in fields like medicine, psychology, and social science. Despite their role in shaping study results and insights, there is no standard or rigorous procedure to determine whether LLMs can replace human annotators. In this paper, we propose a novel statistical procedure -- the Alternative Annotator Test (alt-test) -- that requires only a modest subset of annotated examples to justify using LLM annotations. Additionally, we introduce a versatile and interpretable measure for comparing LLM judges. To demonstrate our procedure, we curated a diverse collection of ten datasets, consisting of language and vision-language tasks, and conducted experiments with six LLMs and four prompting techniques. Our results show that LLMs can sometimes replace humans with closed-source LLMs (such as GPT-4o), outperforming open-source LLMs, and that prompting techniques yield judges of varying quality. We hope this study encourages more rigorous and reliable practices.

While Large Language Models (LLMs) are increasingly adopted as automated judges for evaluating generated text, their outputs are often costly, and highly sensitive to prompt design, language, and aggregation strategies, severely, which limits reproducibility. To address these challenges, we propose \textit{OmniScore}, a family of complementary, deterministic learned metrics developed using small size (<1B) parameter models. OmniScore approximates LLM-judge behavior while preserving the low latency and consistency of traditional model-based scoring. We trained the models large-scale synthetic supervision (sim564k instances, in 107 languages) and evaluated using 8,617 manually annotated instances. The OmniScore family supports reliable, multi-dimensional scores across a variety of settings, including reference-based, source-grounded, and hybrid evaluations. We evaluate these models across question answering (QA), translation, and summarization in 6 languages. Our results demonstrate that lightweight, deterministic learned metrics provide a highly practical and scalable alternative to frontier LLMs. Our models and datasets can be found at https://huggingface.co/collections/QCRI/omniscore

Existing LLM-as-a-Judge systems suffer from three fundamental limitations: limited adaptivity to task- and domain-specific evaluation criteria, systematic biases driven by non-semantic cues such as position, length, format, and model provenance, and evaluation inconsistency that leads to contradictory judgments across different evaluation modes (e.g., pointwise versus pairwise). To address these issues, we propose FairJudge, an adaptive, debiased, and consistent LLM-as-a-Judge. Unlike prior approaches that treat the judge as a static evaluator, FairJudge models judging behavior itself as a learnable and regularized policy. From a data-centric perspective, we construct a high-information-density judging dataset that explicitly injects supervision signals aligned with evaluation behavior. Building on this dataset, we adopt a curriculum-style SFT-DPO-GRPO training paradigm that progressively aligns rubric adherence, bias mitigation, and cross-mode consistency, while avoiding catastrophic forgetting. Experimental results on multiple internal and public benchmarks show that FairJudge consistently improves agreement and F1, reduces non-semantic biases, and outperforms substantially larger instruction-tuned LLMs. All resources will be publicly released after acceptance to facilitate future research.

While LLM-as-a-Judge is widely used in automated evaluation, existing validation practices primarily operate at the level of observed outputs, offering limited insight into whether LLM judges themselves function as stable and reliable measurement instruments. To address this limitation, we introduce a two-phase diagnostic framework for assessing reliability of LLM-as-a-Judge, grounded in Item Response Theory (IRT). The framework adopts Graded Response Model (GRM) of IRT and formalizes reliability along two complementary dimensions: (1) intrinsic consistency, defined as the stability of measurement behavior under prompt variations, and (2) human alignment, capturing correspondence with human quality assessments. We empirically examine diverse LLM judges with this framework, and show that leveraging IRT-GRM yields interpretable signals for diagnosing judgments systematically. These signals provide practical guidance for verifying reliablity of LLM-as-a-Judge and identifying potential causes of unreliability.

Large Language Model (LLM)-based question-answering systems offer significant potential for automating customer support and internal knowledge access in small businesses, yet their practical deployment remains challenging due to infrastructure costs, engineering complexity, and security risks, particularly in retrieval-augmented generation (RAG)-based settings. This paper presents an industry case study of an open-source, multi-tenant platform that enables small businesses to deploy customised LLM-based support chatbots via a no-code workflow. The platform is built on distributed, lightweight k3s clusters spanning heterogeneous, low-cost machines and interconnected through an encrypted overlay network, enabling cost-efficient resource pooling while enforcing container-based isolation and per-tenant data access controls. In addition, the platform integrates practical, platform-level defences against prompt injection attacks in RAG-based chatbots, translating insights from recent prompt injection research into deployable security mechanisms without requiring model retraining or enterprise-scale infrastructure. We evaluate the proposed platform through a real-world e-commerce deployment, demonstrating that secure and efficient LLM-based chatbot services can be achieved under realistic cost, operational, and security constraints faced by small businesses.

Recent developments in large language models have shown advantages in reallocating a notable share of computational resource from training time to inference time. However, the principles behind inference time scaling are not well understood. In this paper, we introduce an analytically tractable model of inference-time scaling: Bayesian linear regression with a reward-weighted sampler, where the reward is determined from a linear model, modeling LLM-as-a-judge scenario. We study this problem in the high-dimensional regime, where the deterministic equivalents dictate a closed-form expression for the posterior predictive mean and variance. We analyze the generalization error when training data are sampled from a teacher model. We draw k inference-time samples and select via softmax at a temperature applied to a quadratic reward. When the reward is not too different from the teacher, the generalization error decreases monotonically with increasing inference time samples k. However, the specific reward that optimizes inference-time selection generally differs from the teacher. In contrast, substantial reward misspecification induces a finite optimal k beyond which more sampling can increase the generalization error. For fixed k, there exists an optimal sampling temperature. We experimentally verify these facts in large language model inference with an additional large language model as a judge. In the "best-of-k" limit with the teacher as reward, we theoretically show that the generalization error decays as Θ(1/k^2) and determine the leading coefficient via extreme value theory. These formulas delineate domains where scaling inference-time computation is provably preferable to collecting more data. Finally, we demonstrate that when task difficulty increases, the previously mentioned advantage of inference-time compute degrades.

Using language models to scalably approximate human preferences on text quality (LLM-as-a-judge) has become a standard practice applicable to many tasks. A judgment is often extracted from the judge's textual output alone, typically with greedy decoding. However, LLM judges naturally provide distributions over judgment tokens, inviting a breadth of inference methods for extracting fine-grained preferences. We find that taking the mean of the judgment distribution consistently outperforms taking the mode (i.e. greedy decoding) in all evaluation settings (i.e. pointwise, pairwise, and listwise). We further explore novel methods of deriving preferences from judgment distributions, and find that methods incorporating risk aversion often improve performance. Lastly, we analyze LLM-as-a-judge paired with chain-of-thought (CoT) prompting, showing that CoT can collapse the spread of the judgment distribution, often harming performance. Our findings show that leveraging distributional output improves LLM-as-a-judge, as opposed to using the text interface alone.

LLM-as-Judge frameworks are increasingly popular for AI evaluation, yet research findings on the relationship between models' generation and judgment abilities remain inconsistent. We investigate this relationship through systematic dataset- and instance-level analyses across 11 models and 21 diverse tasks. Despite both capabilities relying on the same underlying knowledge, our analyses reveal they are only weakly correlated, primarily due to LLMs' sensitivity to the responses being judged. To address this, we propose a self-reference-guided evaluation strategy that leverages a model's own answers as references. This approach significantly strengthens the correlation between generation and judgment abilities, offering a practical path to align these skills and providing a reliable proxy for model selection in evaluation tasks.

Gender-neutral translation (GNT) aims to avoid expressing the gender of human referents when the source text lacks explicit cues about the gender of those referents. Evaluating GNT automatically is particularly challenging, with current solutions being limited to monolingual classifiers. Such solutions are not ideal because they do not factor in the source sentence and require dedicated data and fine-tuning to scale to new languages. In this work, we address such limitations by investigating the use of large language models (LLMs) as evaluators of GNT. Specifically, we explore two prompting approaches: one in which LLMs generate sentence-level assessments only, and another, akin to a chain-of-thought approach, where they first produce detailed phrase-level annotations before a sentence-level judgment. Through extensive experiments on multiple languages with five models, both open and proprietary, we show that LLMs can serve as evaluators of GNT. Moreover, we find that prompting for phrase-level annotations before sentence-level assessments consistently improves the accuracy of all models, providing a better and more scalable alternative to current solutions.

Large Language Models (LLMs) are powerful zero-shot assessors and are increasingly used in real-world situations such as for written exams or benchmarking systems. Despite this, no existing work has analyzed the vulnerability of judge-LLMs against adversaries attempting to manipulate outputs. This work presents the first study on the adversarial robustness of assessment LLMs, where we search for short universal phrases that when appended to texts can deceive LLMs to provide high assessment scores. Experiments on SummEval and TopicalChat demonstrate that both LLM-scoring and pairwise LLM-comparative assessment are vulnerable to simple concatenation attacks, where in particular LLM-scoring is very susceptible and can yield maximum assessment scores irrespective of the input text quality. Interestingly, such attacks are transferable and phrases learned on smaller open-source LLMs can be applied to larger closed-source models, such as GPT3.5. This highlights the pervasive nature of the adversarial vulnerabilities across different judge-LLM sizes, families and methods. Our findings raise significant concerns on the reliability of LLMs-as-a-judge methods, and underscore the importance of addressing vulnerabilities in LLM assessment methods before deployment in high-stakes real-world scenarios.

Multimodal Large Language Models (MLLMs) have gained significant attention recently, showing remarkable potential in artificial general intelligence. However, assessing the utility of MLLMs presents considerable challenges, primarily due to the absence of multimodal benchmarks that align with human preferences. Drawing inspiration from the concept of LLM-as-a-Judge within LLMs, this paper introduces a novel benchmark, termed MLLM-as-a-Judge, to assess the ability of MLLMs in assisting judges across diverse modalities, encompassing three distinct tasks: Scoring Evaluation, Pair Comparison, and Batch Ranking. Our study reveals that, while MLLMs demonstrate remarkable human-like discernment in Pair Comparison, there is a significant divergence from human preferences in Scoring Evaluation and Batch Ranking. Furthermore, a closer examination reveals persistent challenges in the judgment capacities of LLMs, including diverse biases, hallucinatory responses, and inconsistencies in judgment, even in advanced models such as GPT-4V. These findings emphasize the pressing need for enhancements and further research efforts to be undertaken before regarding MLLMs as fully reliable evaluators. In light of this, we advocate for additional efforts dedicated to supporting the continuous development within the domain of MLLM functioning as judges. The code and dataset are publicly available at our project homepage: https://mllm-judge.github.io/.

As interactive LLM-based applications are created and refined, model developers need to evaluate the quality of generated text along many possible axes. For simpler systems, human evaluation may be practical, but in complicated systems like conversational chatbots, the amount of generated text can overwhelm human annotation resources. Model developers have begun to rely heavily on auto-evaluation, where LLMs are also used to judge generation quality. However, existing LLM-as-a-judge benchmarks largely focus on simple Q\&A tasks that do not match the complexity of multi-turn conversations. We introduce RankJudge, a benchmark generator for evaluating LLM-as-a-judge on multi-turn conversations grounded in reference documents. RankJudge creates pairs of conversations where one conversation has a single flaw injected into one turn. This construction allows paired conversations to be labeled unambiguously as better or worse, and precisely isolates failure categories to individual turns, enabling a strict joint correctness criterion for judging. We implement RankJudge across the domains of machine learning, biomedicine, and finance, evaluate 21 frontier LLM judges, and rank those judges via the Bradley-Terry model. Our formulation also allows ranking each conversation pair with difficulty ratings, which we use to dynamically curate the evaluation slice to reduce label noise, as confirmed via human annotation. We find that judge rankings are stable under partial observability, coarser correctness criteria, and an alternative random-walk rating algorithm.

Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between data generator LLM and judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive issue that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge. We release all codes and data at: https://github.com/David-Li0406/Preference-Leakage.

Large language models (LLMs) are increasingly used as evaluators in lieu of humans. While scalable, their judgments are noisy due to imperfect specificity and sensitivity of LLMs, leading to biased accuracy estimates. Although bias-correction methods exist, they are underutilized in LLM research and typically assume exact knowledge of the model's specificity and sensitivity. Furthermore, in general we only have estimates of these values and it is not well known how to properly construct confidence intervals using only estimates. This work presents a simple plug-in framework that corrects such bias and constructs confidence intervals reflecting uncertainty from both test and calibration dataset, enabling practical and statistically sound LLM-based evaluation. Additionally, to reduce uncertainty in the accuracy estimate, we introduce an adaptive algorithm that efficiently allocates calibration sample sizes.

Using a language model to score or rank candidate responses has become a scalable alternative to human evaluation in reinforcement learning from human feedback (RLHF) pipelines, benchmarking, and application layer evaluations. However, output reliability depends heavily on prompting and aggregation strategy. We present an empirical investigation of four drop-in techniques -- ensemble scoring, task-specific criteria injection, calibration context, and adaptive model escalation -- for improving LLM judge accuracy on RewardBench 2, with a unifying lens of noise control on the stochastic judge: ensembling as Monte Carlo averaging over per-call noise, criteria injection as between-response discrimination sharpening, and per-response score variance as an uncertainty signal. Ensemble scoring and task-specific criteria injection (the latter virtually cost free) together reach up to 85.8% accuracy, +13.5pp over baseline. Calibration context and adaptive model escalation also improve over baseline but are dominated by criteria + ensembling on the cost-accuracy Pareto frontier. Small models benefit disproportionately from ensembling, making high-accuracy LLM judges accessible at low cost. We show that these techniques generalise across model providers, evaluating on both OpenAI GPT and Anthropic Claude families.

Prompt attacks, including jailbreaks and prompt injections, pose a critical security risk to Large Language Model (LLM) systems. In production, guardrails must mitigate these attacks under strict low-latency constraints, resulting in a deployment gap in which lightweight classifiers and rule-based systems struggle to generalize under distribution shift, while high-capacity LLM-based judges remain too slow or costly for live enforcement. In this work, we examine whether lightweight, general-purpose LLMs can reliably serve as security judges under real-world production constraints. Through careful prompt and output design, lightweight LLMs are guided through a structured reasoning process involving explicit intent decomposition, safety-signal verification, harm assessment, and self-reflection. We evaluate our method on a curated dataset combining benign queries from real-world chatbots with adversarial prompts generated via automated red teaming (ART), covering diverse and evolving patterns. Our results show that general-purpose LLMs, such as gemini-2.0-flash-lite-001, can serve as effective low-latency judges for live guardrails. This configuration is currently deployed in production as a centralized guardrail service for public service chatbots in Singapore. We additionally evaluate a Mixture-of-Models (MoM) setting to assess whether aggregating multiple LLM judges improves prompt-attack detection performance relative to single-model judges, with only modest gains observed.