id string | question_type string | depth int64 | reasoning_type string | cognitive_skill string | venue string | consensus_tier string | question string | options list | correct_index int64 | seed_paper_id string | seed_arxiv_id string | seed_title string | gold_paper_ids list | gold_arxiv_ids list | bridge_paper_ids list | bridge_arxiv_ids list | distractor_types list | filter_agreement int64 | bridge_necessary bool | bridge_audit_verdict string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
mt_0001 | multi-target | 1 | RESULT | retrieve | ECCV | bronze | GScream contrasts two implicit-representation baselines as image-domain guidance strategies before introducing its 3D Gaussian feature propagation: one that updates the scene using multi-view inpainted images, and another that instead starts from a single inpainted reference and simulates view-dependent effects for oth... | [
"For the first strategy, the paper states that SPIn-NeRF and OR-NeRF βutilize patch-based LPIPS loss in their optimization objectiveβ and that prior approaches βrely on leveraging multi-view in-painted images to update NeRF/3DGS models.β For the second strategy, View-Sub βoffers an alternative by in-painting a sing... | 1 | 07a056ab69ac80d9ac35b7d4004a052592a30c06 | 2404.13679 | GScream: Learning 3D Geometry and Feature Consistent Gaussian Splatting for Object Removal | [
"04c36e581a4114096c07598eec310028206c4cc9",
"5b777875ce84db059600f3767c70ab55e8746222"
] | [
"2305.10503",
"2304.09677"
] | [] | [] | [
"seed_only",
"correct",
"wrong_paper",
"no_context"
] | 2 | null | null |
mt_0002 | multi-target | 1 | METHOD | reason | CVPR | bronze | The seed says earlier point-based editing methods share a two-step optimization scheme of motion supervision and point tracking, first in a GAN-based milestone and then in a diffusion-based adaptation with improved spatial control. Why did the GAN-based method choose intermediate generator features after a specific blo... | [
"In DragGAN, the handle-point loss is computed on intermediate StyleGAN features taken after the 6th synthesis block because these maps still preserve clear local semantics and spatial correspondence while being discriminative enough for reliable nearest-neighbor point tracking; to keep the image identity and textu... | 2 | 086133c46455f4fcc47aa115483707c20b09bda2 | 2404.01050 | Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation | [
"03a281a176413ed4d140293edc5bf04a3ad7a1f1",
"05b15934d837dc84afa96824742d3dcc7ec88e09"
] | [
"2306.14435",
"2305.10973"
] | [] | [] | [
"no_context",
"seed_only",
"correct",
"wrong_paper"
] | 2 | null | null |
mt_0004 | multi-target | 1 | RESULT | synthesize | ACL | bronze | Among the prior RecSys works the seed groups together as using instruction tuning and prompt-based LLM adaptation, one unifies multiple recommendation tasks with personalized prompts and tests transfer to unseen prompts and new domains, while the other systematically benchmarks off-the-shelf and fine-tuned language mod... | [
"In P5, a single T5 model trained with personalized natural-language prompts could unify many recommendation tasks and even generalize to unseen prompts/new domains, but its strongest gains came after task-specific prompt/instruction tuning rather than pure zero-shot prompting. Complementing that, the rating-predic... | 2 | 0def19d854d087e700dc9b4daad6d2f88b351bf7 | 2408.00802 | Leveraging LLM Reasoning Enhances Personalized Recommender Systems | [
"df602516e28a9ef0ef665ed0aef551984d8d770d",
"ed8832466a4295a6c110eb6beec9531d54b5aede"
] | [
"2203.13366",
"2305.06474"
] | [] | [] | [
"no_context",
"seed_only",
"correct",
"wrong_paper"
] | 2 | null | null |
mt_0005 | multi-target | 1 | MOTIVE | synthesize | NeurIPS | gold | The seed argues that earlier regularization-based sparse-view methods were an inadequate foundation for enhancement because one NeRF-based approach tends toward smooth, blurry renderings, while a depth-regularized 3D Gaussian approach reduces artifacts but remains noise-sensitive. Using the experiments and analyses in ... | [
"The frequency-regularized NeRF approach introduces slight blurriness and shows a trade-off where a longer frequency curriculum raises PSNR but lowers LPIPS. The depth-regularized Gaussian approach still improves quality at 6 views but does not work well at 9 views because depth-map errors negatively influence the ... | 0 | 0f0363d422a5c44ffd8d30661f6b6720d72f0b8f | 2410.16266 | 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors | [
"59015e5b1f0fbe6a41feb87b1fbf76f3c7339840",
"f710834a3e6de8f066b116a97711e9782532c15f"
] | [
"2303.07418",
"2403.06912"
] | [] | [] | [
"correct",
"wrong_paper",
"no_context",
"seed_only"
] | 3 | null | null |
mt_0006 | multi-target | 1 | RESULT | justify | ECCV | gold | The seed claims its single-epoch tuning is more training-efficient than earlier pseudo-word token-based zero-shot CIR approaches. Looking at the image-to-token method that trains a mapping network on CC3M with a fixed prompt, and the later projection-based method that explicitly compares itself against that family on t... | [
"The paper states that its TAT method reaches optimal performance with just a single epoch and is more training-efficient than pseudo-word token-based approaches, which require tens to hundreds of training epochs. It also emphasizes that this single-epoch LoRA tuning is highly efficient, with less than 0.5% trainab... | 2 | 11944c9e7b5668f0e50f2815006c6b076a968550 | 2405.00571 | Spherical Linear Interpolation and Text-Anchoring for Zero-shot Composed Image Retrieval | [
"9b897a8739f04fa4eff7431fd8a3c547b49f4878",
"15d6053cc60880524c6b01c5364e6e3f4217a522"
] | [
"2302.03084",
"2312.01998"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 3 | null | null |
mt_0007 | multi-target | 1 | GROUND | retrieve | ACL | bronze | The seed motivates its multi-level rephrasing benchmark by citing earlier evidence that retrieval systems are brittle to semantically equivalent query changes. What headline degradation figures are reported by the work that automatically generates several categories of syntax-changing query variations, and by the work ... | [
"The paper motivates its rephrasing benchmark by stating that retrieval models are highly vulnerable to query rephrasing, often leading to significant performance degradation, and it reports its own strong sensitivity findings through multi-level rephrasing evaluation. In this benchmark, ColQwen drops from 60.8 to ... | 2 | 1483ce1ff6cd42c0e60eebea49170182569e9d59 | 2502.12342 | REAL-MM-RAG: A Real-World Multi-Modal Retrieval Benchmark | [
"208e6ae33dacef760962e4d51a263d9a7652c5c4",
"d586af1899f39e2da2822be1af6e4c8317a6ebb5"
] | [
"2111.13057",
"2205.02303"
] | [] | [] | [
"seed_only",
"no_context",
"correct",
"wrong_paper"
] | 2 | null | null |
mt_0008 | multi-target | 1 | GROUND | contrast | CVPR | silver | The seed paper treats two earlier reconstruction-error-based detectors as the main foundations for its own method, then argues that both are inefficient because they depend on multi-step DDIM-style feature extraction. Under that framing, how do these two predecessors differ in what error signal they compute and in the ... | [
"The full-reconstruction approach computes the absolute difference between an input image and its DDIM-inverted-and-reconstructed counterpart, and it reports near-perfect cross-model generalization such as 99.9% ACC and 100% AP on average across multiple unseen diffusion generators. The stepwise approach instead co... | 0 | 339489f84d360666f6f8773416a35383c2f88978 | 2403.17465 | LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection | [
"d9d82eb6a8886226724ea230a7e5923d660a0bad",
"7e8e9b77de1e09829aaba30894b67f781b755711"
] | [
"2303.09295",
"2307.06272"
] | [] | [] | [
"correct",
"no_context",
"seed_only",
"wrong_paper"
] | 3 | null | null |
mt_0009 | multi-target | 1 | RESULT | retrieve | ECCV | gold | STAG4D contrasts the pioneering text-to-4D approach that optimizes from a text prompt using a video diffusion prior with the monocular video-to-4D approach that instead starts from a statically captured input video. What frozen diffusion supervision does each of those two earlier methods use during optimization, and wh... | [
"The earlier text-to-4D method is DreamGaussian4D, which optimizes a dynamic Gaussian representation under frozen video-diffusion guidance from ModelScope Text-to-Video and reports about 2 hours for its main 4D generation stage. The monocular video-to-4D baseline is 4D-fy, which starts from a single captured video ... | 2 | 349e53ad4deaae7c4ed164988057981655b12843 | 2403.14939 | STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians | [
"c7f4a93754a29426acc9bf9c03db27b6106eecaf",
"d684fbe07585be651cc93d3c00ae3fe6df3ac877"
] | [
"2311.02848",
"2301.11280"
] | [] | [] | [
"no_context",
"seed_only",
"correct",
"wrong_paper"
] | 3 | null | null |
mt_0010 | multi-target | 1 | MOTIVE | assess | ECCV | gold | RegionDrag motivates its region-based interface by grouping two earlier Stable-Diffusion point-drag methods as examples of editing through sparse handle/target points. Looking at the diffusion-based method that optimizes a single inverted latent with point tracking and the stochastic alternative that replaces latent op... | [
"For the latent-optimization method DragDiffusion, the paper explicitly says its sparse handle/target-point interface works when βthe object to move is deformable and shape-preservingβ and demonstrates this on real-image editing after DDIM inversion with supervised point tracking, but its own setup reveals limits: ... | 2 | 021f4d28ac16d074c484a0700d14efd93c6bed2d | 2407.18247 | RegionDrag: Fast Region-Based Image Editing with Diffusion Models | [
"03a281a176413ed4d140293edc5bf04a3ad7a1f1",
"1d8019f49b319aea98acd572f4ebb7fdeeddedb4"
] | [
"2306.14435",
"2311.01410"
] | [] | [] | [
"no_context",
"wrong_paper",
"correct",
"seed_only"
] | 3 | null | null |
mt_0011 | multi-target | 1 | MOTIVE | retrieve | CVPR | gold | DEADiff motivates its design by combining a non-reconstructive training idea from the earlier Q-Former-based subject method with a criticism of the image-prompt adapter that needs careful per-image balancing between text and image conditions. What exact training-pair construction does the former use to avoid trivial so... | [
"The earlier Q-Former-based subject method curates prompted-context-generation pairs by extracting the foreground subject, composing it onto a random background, and using the synthetic image as input while the original subject image is the output. The image-prompt adapter scales image conditioning with an explicit... | 0 | 08f20fadfb51974a6883b935593a93e2f39408b6 | 2403.06951 | DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations | [
"2854e5bab8e6f36e54c64456628a9559bf67019e",
"dc0c132b273456b288a785414db2fa72edf87b1a"
] | [
"2308.06721",
"2305.14720"
] | [] | [] | [
"correct",
"seed_only",
"wrong_paper",
"no_context"
] | 3 | null | null |
mt_0012 | multi-target | 1 | RESULT | reason | ECCV | bronze | BAMM contrasts two denoising-style baselines that it says both need motion length in advance, while differing in how they corrupt and recover motion: one adds structured Gaussian noise and adapts inpainting for editing, and the other replaces motion tokens with [MASK] tokens and iteratively refines them. Why did these ... | [
"The diffusion approach uses a gradual noising-and-denoising process because human motion is a many-to-many problem and predicting the clean sample enables geometric losses on positions, velocities, and foot contact, which supports coherent conditional editing by overwriting known parts during sampling. The masked-... | 0 | 0dde7e650ed368b2c37559b6fddc21f3617a40ee | 2403.19435 | BAMM: Bidirectional Autoregressive Motion Model | [
"f52af2c0cf521b48edd50670e9ba9cd02646d2b4",
"15736f7c205d961c00378a938daffaacb5a0718d"
] | [
"2312.00063",
"2209.14916"
] | [] | [] | [
"correct",
"wrong_paper",
"no_context",
"seed_only"
] | 2 | null | null |
mt_0014 | multi-target | 1 | GROUND | synthesize | SIGIR | gold | For the two foundational Slack-based collaborative search studies that the seed groups together as examples of mixed interactionβone unpacking why and when a human-operated system asks clarifying questions or offers search suggestions, and the other measuring the effects of those initiative levels on collaboratorsβwhat... | [
"Across the two Slack-based mixed-initiative collaborative search studies, the Wizard-of-Oz analysis showed that clarifying questions are most appropriate early or when the request is ambiguous/underspecified, while search suggestions are better when the task is well-scoped and the system can proactively advance th... | 1 | 1082b4734c07a1e0520c180ad2c1c6682924b3ff | 2402.06360 | CoSearchAgent: A Lightweight Collaborative Search Agent with Large Language Models | [
"f54fbe63b12219f46e39fef9733a5ed507cf9a78",
"4f329aa7610e0e0f71870a09f0ecf9fbcea93f45"
] | [
"2303.13484",
"2202.09728"
] | [] | [] | [
"no_context",
"correct",
"seed_only",
"wrong_paper"
] | 3 | null | null |
mt_0015 | multi-target | 1 | MOTIVE | synthesize | ECCV | gold | The seed points to two supervised, voxel-based methods for large-scale scene flow as motivation for its own self-supervised design, specifically citing them for the dynamic/static imbalance problem and the constraints of voxelized processing. Combining the evidence from the earlier scalable full-point-cloud baseline an... | [
"In FastFlow3Dβs scalable full-point-cloud baseline on Argoverse 2, the authors quantify the data-distribution problem directly: only about 8.6% of points are dynamic and 91.4% are static, and with this severe class imbalance a simple nearest-neighbor βcopy the previous pointβ static-flow baseline already achieves ... | 1 | 143a44051a8542585339f1f7ae08c2c1b457e745 | 2407.01702 | SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving | [
"434108bc94523af160c9c01e4f7310704aa4f722",
"64129711ce069151dea9d023e8b928095ea7a52a"
] | [
"2103.01306",
"2401.16122"
] | [] | [] | [
"no_context",
"correct",
"wrong_paper",
"seed_only"
] | 3 | null | null |
mt_0019 | multi-target | 1 | MOTIVE | retrieve | CVPR | silver | DiffMOT motivates its diffusion predictor by pointing to earlier tracking-by-detection systems that still depend on Kalman-filter-style linear motion handling, even when one improves association by keeping low-score detections and another tries to repair occlusion errors with observation-centric updates. What exact mot... | [
"The low-score-box association baseline sets the default detection score threshold to 0.6 and adopts Kalman filter to predict the new locations in the current frame of each track. The observation-centric baseline generates the virtual trajectory during occlusion with the constant-velocity assumption and calculates ... | 0 | 216ed181664b9b3579840c810be0cc56f82b58aa | 2403.02075 | DiffMOT: A Real-time Diffusion-based Multiple Object Tracker with Non-linear Prediction | [
"7d1ff4ac2390759cbe60dd46b2b9bcabd4a90db4",
"643914ae0ad9dcc5868348e4d867374aedbf9531"
] | [
"2110.06864",
"2203.14360"
] | [] | [] | [
"correct",
"wrong_paper",
"no_context",
"seed_only"
] | 3 | null | null |
mt_0020 | multi-target | 1 | MOTIVE | reason | ECCV | bronze | BAGS groups two earlier NeRF-based deblurring methods as approaches that sparsely estimate the blur kernel instead of predicting dense full-image kernels. Why did the earlier method with deformable sparse rays argue that sparsity was necessary in the first place, and why did the later method with physically shared rigi... | [
"BAGS states that NeRF-based deblurring methods suffered from slow training because only a limited number of rays can be rendered at once, and kernel modeling requires rendering a patch to model a single pixel, so prior work mitigated this by modeling the blur kernel sparsely. It also says those earlier approaches ... | 2 | 23d4b38f0ff080e943594882398796ffd3126985 | 2403.04926 | BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling | [
"7bc683fe1911f4987b845f7d2165d5888f0a50e9",
"dc32021df1c3692601fa2d4747e41f73005eda4f"
] | [
"2111.14292",
"2211.12046"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 2 | null | null |
mt_0023 | multi-target | 1 | MOTIVE | reason | ECCV | bronze | Click-Gaussian motivates its own global feature guidance by contrasting two earlier multi-object mask-lifting strategies for handling inconsistent SAM masks across views: one 3D Gaussian method that pre-assigns segment identities with a zero-shot tracker, and one NeRF-based method that uses a scale-conditioned feature ... | [
"In the 3D Gaussian segmentation pipeline built on SAM masks, object IDs had to be fixed before optimization because each per-view SAM proposal is independent, so the method first ran a zero-shot multi-object tracker to associate masks of the same object across images and assign a consistent segment label to the Ga... | 2 | 2d81e6af252b16238b10cf6009514a61ed60d77a | 2407.11793 | Click-Gaussian: Interactive Segmentation to Any 3D Gaussians | [
"5f7867681c6287cc4f4e7fb5260af8067b52b3d9",
"5232fd8b6e7065de424e61014b14e933257f4c24"
] | [
"2312.00732",
"2401.09419"
] | [] | [] | [
"no_context",
"wrong_paper",
"correct",
"seed_only"
] | 2 | null | null |
mt_0026 | multi-target | 1 | MOTIVE | reason | ACL | gold | MMInA argues that earlier benchmarks made web agents look promising on simpler settings, yet still motivated a new benchmark and a memory-based fix for multihop cross-website failures. Why did the cited standalone multicategory web environment conclude that realistic long-horizon web tasks remain hard for a single prom... | [
"The standalone realistic web environment found that long-horizon tasks stayed difficult because current agents lacked capabilities such as active exploration and failure recovery, and even the strongest model often stopped early by wrongly judging feasible tasks impossible. The multi-agent benchmark argued that as... | 0 | 34fc623e43f60b73827f4d0259013b96aeddfe6f | 2404.09992 | MMInA: Benchmarking Multihop Multimodal Internet Agents | [
"e41482f4ee984f17382f6cdd900df094d928be06",
"ce212cb873a54e5716da53a66b10298ac013008a"
] | [
"2307.13854",
"2308.05960"
] | [] | [] | [
"correct",
"no_context",
"seed_only",
"wrong_paper"
] | 3 | null | null |
mt_0027 | multi-target | 1 | MOTIVE | assess | ICLR | silver | ESC keeps the prompting setup fixed by inheriting the few-shot chain-of-thought exemplars from the earlier prompting paper and the evaluation protocol from the later voting-based decoding paper. Under that inherited setup, what conditions did the earlier prompting work place on when step-by-step exemplars improve reaso... | [
"Under the inherited setup, the paper states that large language models βhave exhibited strong reasoning capabilities β¦ especially with chain-of-thought (CoT) prompting,β and it uses harder arithmetic datasets while excluding MultiArith, SVAMP, AddSub and ASDiv because they are βrelatively simple.β It also reports ... | 2 | 36260206be303b3820df8fbba161591da0ca987c | 2401.10480 | Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning | [
"5f19ae1135a9500940978104ec15a5b8751bc7d2",
"1b6e810ce0afd0dd093f789d2b2742d047e316d5"
] | [
"2203.11171",
"2201.11903"
] | [] | [] | [
"seed_only",
"no_context",
"correct",
"wrong_paper"
] | 3 | null | null |
mt_0032 | multi-target | 1 | MOTIVE | contrast | SIGIR | bronze | The seed motivates WeightedGCL by criticizing unstable perturbations and uniform feature weighting in earlier graph contrastive recommenders. Among the cited methods that establish GCLβs effectiveness, how does the earlier approach that builds two graph views through node/edge/random-walk perturbations differ from the ... | [
"The earlier method is SGL (SIGIR 2021), which creates two stochastic views by applying node dropout, edge dropout, or random-walk sampling to the userβitem graph and then performs self-supervised contrastive learning over the corresponding user/item embeddings; it argues that random graph perturbations can preserv... | 1 | 45b93710eac997d24fa40bfce622fa12a6d9d8ce | 2504.04443 | Squeeze and Excitation: A Weighted Graph Contrastive Learning for Collaborative Filtering | [
"75739ed2ddebd7982042f516f407553f8d3110f8",
"1b06e507f7e2af3dc7cf28086ac34083709c3150"
] | [
"2010.10783",
"2302.08191"
] | [] | [] | [
"no_context",
"correct",
"wrong_paper",
"seed_only"
] | 2 | null | null |
mt_0035 | multi-target | 1 | MOTIVE | contrast | ECCV | gold | The seed highlights MS-COCO 30k as a benchmark for evaluating realism in compression and contrasts prior realism-oriented baselines with its own attempt to stay faithful to the input. Among the cited works, how does the diffusion-based method that adds details on top of an MSE autoencoder differ from the conditional ge... | [
"The paper describes HFD as a diffusion post-processing method that adds details to images compressed by an autoencoder-based codec, but notes that such reconstructions often contain high-frequency artifacts or entirely lose image content because the diffusion model cannot compensate for artifacts from the initial ... | 2 | 522091bb866f62af5f7b076a52aa70d942b1b13d | 2404.08580 | Lossy Image Compression with Foundation Diffusion Models | [
"40e21aeb07ae744fad3a47f658ff8dd8fbbc7f71",
"76a0cb56bed90ba3de90baea1a5f29fe91f6f1cb"
] | [
"2305.18231",
"2212.13824"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 3 | null | null |
mt_0037 | multi-target | 1 | MOTIVE | retrieve | ECCV | gold | The seed says it follows an earlier tuning-free high-resolution adaptation work when evaluating three pre-trained text-to-image models at four larger resolutions, including the larger model variant it also cites separately. What exact four higher-resolution settings does that earlier adaptation work use, and what train... | [
"Following the earlier high-resolution adaptation setup, the paper evaluates images at four larger-resolution settings corresponding to 4Γ, 6.25Γ, 8Γ, and 16Γ the pixel count of the training resolution. For the larger SDXL 1.0 model, it states that training is done at resolutions close to 1024 Γ 1024 pixels, and th... | 1 | 5c80f5947fed7c5c2b58a4aebfb7c3be4b3ce35d | 2403.12963 | FouriScale: A Frequency Perspective on Training-Free High-Resolution Image Synthesis | [
"b02d8250bf26701f0e300bb6d6c3cff7558b2754",
"d7890d1906d95c4ae4c430b350455156d6d8aed9"
] | [
"2310.07702",
"2307.01952"
] | [] | [] | [
"seed_only",
"correct",
"no_context",
"wrong_paper"
] | 3 | null | null |
mt_0038 | multi-target | 1 | RESULT | synthesize | ICLR | bronze | In the Table 1 comparison that motivates WildBench against earlier automatic evaluators, one cited benchmark is the manually created multi-turn benchmark used to argue that pairwise comparisons can reveal subtle differences, while the other is the benchmark line later revisited to debias automatic judgments for verbosi... | [
"The paper reports that pairwise and multi-turn evaluation can align closely with people, noting that GPT-4 reaches 85% agreement with humans under setup S2 without ties, which is even higher than human-human agreement of 81%. At the same time, its length-bias analysis finds that both GPT-4 and Prometheus show a si... | 3 | 5d12dfd7278cb8da26f9fd1956cad3c15cea9863 | 2406.04770 | WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild | [
"a0a79dad89857a96f8f71b14238e5237cbfc4787",
"eb375712bd37250c350ecd3f559e1879e87eb3e5"
] | [
"2306.05685",
"2404.04475"
] | [] | [] | [
"wrong_paper",
"seed_only",
"no_context",
"correct"
] | 2 | null | null |
mt_0040 | multi-target | 1 | MOTIVE | synthesize | NeurIPS | gold | The seed cites two works as motivation for reducing communication and computation overhead in federated foundation models: one is a broad assessment of parameter-efficient tuning methods that quantifies memory and trainable-parameter tradeoffs for large language models, and the other is a gradient-free federated prompt... | [
"The paperβs cited resource figure for a large-model PEFT setup shows that LLaMA-13B-Alpaca with QLoRA still requires 125.2M trainable parameters and 6922.3M memory on the user side, whereas black-box tuning is reported as suitable for LMaaS because its optimization cost is decoupled from PTM scale and it keeps use... | 3 | 60071051e176b6e2fadbededadbb08b05e125c13 | 2403.19211 | Dual-Personalizing Adapter for Federated Foundation Models | [
"83de13bd492b9e72c314e308f0d77014154a6a74",
"187f4521e6080f93fc2a26bf91b4e7d64f94e18a"
] | [
"2312.12148",
"2310.01467"
] | [] | [] | [
"wrong_paper",
"seed_only",
"no_context",
"correct"
] | 3 | null | null |
mt_0041 | multi-target | 1 | RESULT | synthesize | NeurIPS | bronze | The seed treats two game-based predecessors as evidence for the older diagnostic idea that world models can be assessed through legal next-token behavior and state-tracking probes. Looking at the chess testbed that evaluates legal move prediction from prompts and the Othello study that probes and intervenes on internal... | [
"In the chess testbed, a GPT-style language model trained only on PGN move text predicts the set of legal next moves with very high accuracy from a prompt positionβreporting roughly 99% legality on in-distribution gamesβso under a legal-next-token diagnostic it looks strongly aware of the underlying board state. In... | 3 | 61e5bafb567b36960a6e1584b3753d9de09b45aa | 2406.03689 | Evaluating the World Model Implicit in a Generative Model | [
"d5295f7ddcf281f3d30a7579d5ce482036a8e27c",
"453fc588d97958c6fefad96e79edd896873b3e09"
] | [
"2210.13382",
"2102.13249"
] | [] | [] | [
"no_context",
"seed_only",
"wrong_paper",
"correct"
] | 2 | null | null |
mt_0043 | multi-target | 1 | MOTIVE | justify | ECCV | gold | The seed contrasts two earlier feature-level synthesis lines: one that creates anomaly features by simply adding Gaussian noise to adapted normal features, and another that samples nearby codebook vectors in a quantized feature space to generate weak, near-in-distribution defects. What concrete evidence do these two wo... | [
"The paper states that prior feature-level synthesis methods include DSR, which βsamples in quantified feature space and synthesizes weak defects through the similarity comparison of codebook feature vector,β and SimpleNet, which βsynthesize anomalies by adding Gaussian noise to the normal features,β showing that f... | 1 | 69bcc8a20663b664df102fd0aad4a1bbbfc93589 | 2407.09359 | A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization | [
"a23c0a89bd21bd2481fedcdd6d1ac891c6c06bdc",
"a6a237a7625f2c3175e44070ee9bad6e917798b0"
] | [
"2303.15140",
"2208.01521"
] | [] | [] | [
"seed_only",
"correct",
"no_context",
"wrong_paper"
] | 3 | null | null |
mt_0045 | multi-target | 1 | RESULT | assess | CVPR | bronze | DAVE claims to beat all prior density-based counters on FSC147, especially the longstanding winner and the earlier method that introduced the benchmark and first argued for density estimation over detection-then-counting. Looking back at those two cited works, what conditions did each use to justify the strength of its... | [
"On FSC147 few-shot density-based counting, DAVE is evaluated under the standard protocol with disjoint object classes across train, validation, and test splits so that the target category is unseen during training, and under these conditions it is reported to outperform all few-shot density-based counters, includi... | 1 | 6c8ba7c878eb9f3ba96446e69508b8c4b8198029 | 2404.16622 | DAVE β A Detect-and-Verify Paradigm for Low-Shot Counting | [
"6b291dea9bb80fea41f4ffa17aaad52ab5d26ada",
"5713544d114fc9c5c1ffb1168fc64cd159101b44"
] | [
"2104.08391",
"2211.08217"
] | [] | [] | [
"seed_only",
"correct",
"no_context",
"wrong_paper"
] | 2 | null | null |
mt_0046 | multi-target | 1 | GROUND | synthesize | NAACL | gold | Among the step-by-step methods the seed cites as improving LLM reasoning, one adds verifier-guided search and self-thinking to a generator, while another breaks hard problems into simpler subproblems to improve easy-to-hard generalization. Under the seed paperβs framing of these as sequential reasoning advances, what c... | [
"The paper frames step-by-step methods as advances that improve reasoning capacity and reports that CoT helps break down complex reasoning into a sequential thought process, while Prolog generation surpasses this baseline especially on harder arithmetic settings: it gives a 10.9% average accuracy gain over CoT on G... | 2 | 6e4ba9f89a56ae9e16be5379dda7decd45c538ca | 2405.17893 | Arithmetic Reasoning with LLM: Prolog Generation & Permutation | [
"d33504263ff0b263a1562075711264228d83b181",
"5437e8adab596d7294124c0e798708e050e25321"
] | [
"2210.16257",
"2205.10625"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 3 | null | null |
mt_0047 | multi-target | 1 | MOTIVE | assess | CVPR | bronze | The seed contrasts a high-level next-best-pose system that reaches goals with a planner against a follow-up that adds a low-level module to rank sampled trajectories for tasks needing more specific motions. Across those two works, under what task conditions is plain goal-reaching planning presented as sufficient, and u... | [
"This paper presents plain goal-reaching planning as sufficient on simple tasks that only require goal information or no accurate trajectory control, where baselines already achieve competent performance. It presents planning as inadequate for tasks that require fine-grained, context-aware trajectoriesβparticularly... | 2 | 7359b10949dc8dd271388976b133aad214d59281 | 2403.03890 | Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation | [
"395a4db5fef867d5bd352585aa00b97004994972",
"b87c370eec32ffca8004e7fe239351caff370e23"
] | [
"2106.12534",
"2204.01571"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 2 | null | null |
mt_0050 | multi-target | 1 | MOTIVE | synthesize | ECCV | gold | The seed criticizes earlier imbalance methods for concentrating on selected worse-learnt modalities and sometimes suppressing the better one. For the prototype-based approach that accelerates a slow-learning modality using class prototypes and the on-the-fly gradient approach that monitors contribution discrepancy duri... | [
"The prototype-based approach states that its strategy is mainly for facilitation for the slow-learning modality, with the audio branch showing no noticeable change while the visual branch improves considerably, and it also adds an entropy regularization term to slow down the dominant modality in the early training... | 0 | 84bf95b68413cc262652ee7c056657fa8417a36a | 2407.09705 | Diagnosing and Re-learning for Balanced Multimodal Learning | [
"c259e512ced41063c3c032a6f9ebcabc63986cda",
"48967206cd2781e2a390754a5ca79230b8735d42"
] | [
"2211.07089",
"2203.15332"
] | [] | [] | [
"correct",
"wrong_paper",
"no_context",
"seed_only"
] | 3 | null | null |
mt_0052 | multi-target | 1 | MOTIVE | reason | NeurIPS | gold | The seed cites two diffusion-policy papers together to justify its claim that theoretically optimal policies can be strongly multimodal and therefore poorly served by deterministic or diagonal Gaussian parameterizations. Why did the offline method that couples diffusion with Q-learning and the online method that formul... | [
"The paper argues that offline datasets are often collected by a mixture of policies, so the true behavior policy can be strongly multimodal, skewed, or have dependencies across action dimensions, which makes simple policy classes inadequate. It specifically motivates a more expressive policy class by stating that ... | 3 | 88f0b92bdb29505055b73b751722da0d2cd3f0ab | 2405.15177 | Diffusion Actor-Critic with Entropy Regulator | [
"2cbea7615ebecea2c414d8fbad47d5d258a5c3b4",
"fe9fe9f15f24fbbb19b62bcd9a3418511a699b84"
] | [
"2208.06193",
"2305.13122"
] | [] | [] | [
"wrong_paper",
"seed_only",
"no_context",
"correct"
] | 3 | null | null |
mt_0054 | multi-target | 1 | RESULT | retrieve | ECCV | gold | For the two U-Net-based baselines singled out in IRSAMβs related workβone characterized by asymmetric context modulation for feature fusion and the other by integrating multiple nested U-Net networksβwhat SIRST/NUAA benchmark IoU and nIoU results do their own papers report? Give the reported numbers from each paperβs m... | [
"The asymmetric-context-modulation U-Net baseline is ACM (Asymmetric Contextual Modulation), whose paper reports on SIRST an IoU of 0.6347 and nIoU of 0.6130, and on NUAA-SIRST an IoU of 0.7232 and nIoU of 0.7457 in its main comparison table. The multiple-nested-U-Net baseline is DNANet, whose paper reports on SIRS... | 3 | 933cc893e31ef88b63b353404073663d29916609 | 2407.07520 | IRSAM: Advancing Segment Anything Model for Infrared Small Target Detection | [
"91c21dc7fd0e3b45c0bb1d83a946931e4eab1d51",
"77ccb7485bcec1d3fc732b19717b69b3c779339d"
] | [
"2009.14530",
"2212.00968"
] | [] | [] | [
"no_context",
"seed_only",
"wrong_paper",
"correct"
] | 3 | null | null |
mt_0055 | multi-target | 1 | METHOD | reason | CVPR | gold | Speedy-Splat says the iterations after densification can be used to βrefineβ a model after pruning, following both the principled sensitivity-based post-hoc pruning approach and the earlier global-significance compression pipeline. Why did each of those two methods include a prune-then-refine stage instead of stopping ... | [
"Speedy-Splat states that after densification, the remaining iterations can be used to βrefineβ models after pruning in the same way as PUP 3D-GS and LightGaussian, because these later iterations essentially fine-tune the model once pruning has been applied. In the paperβs description, this refine stage is needed b... | 2 | 941ada3e24e6f54bb49cfaeff998f4f5cbac5a38 | 2412.00578 | Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives | [
"3046a843ca616b00c6c11f1fec3a1585ed8fd7c9",
"951facdc297370af63da7f6c36f6b2114dd6b01f"
] | [
"2406.10219",
"2311.17245"
] | [] | [] | [
"seed_only",
"no_context",
"correct",
"wrong_paper"
] | 3 | null | null |
mt_0057 | multi-target | 1 | RESULT | reason | ECCV | gold | Glyph-ByT5 positions itself between earlier visual text rendering efforts by drawing inspiration from a character-aware encoder study and by benchmarking against a representative scene-text method that adds glyph conditions to an off-the-shelf diffusion model. Why did the character-aware line of work argue that changin... | [
"The character-aware motivation in ByT5 was that βcharacter-blindβ text encoders had limited spelling ability, and the paper further notes that typos, spelling variants, capitalization changes, and unknown characters can cause token representations to change completely and lead to mispredictions. It also reports th... | 3 | 9b13450581dc2a04366ad95fb60e30cd69453be2 | 2403.09622 | Glyph-ByT5: A Customized Text Encoder for Accurate Visual Text Rendering | [
"f8ef90a8cac1a8edf9477688aac24864c547d6cd",
"5fbe4c92791fbecb179c1ab79bba9a59b2e155ba"
] | [
"2212.10562",
"2305.18259"
] | [] | [] | [
"wrong_paper",
"no_context",
"seed_only",
"correct"
] | 3 | null | null |
mt_0058 | multi-target | 1 | MOTIVE | assess | ECCV | bronze | The seed contrasts very strong results on standard cross-view benchmarks with a remaining gap to real-world applications. Under that framing, how far do the claims of the transformer-based method trained on common benchmarks actually extend, and what concrete dataset conditions in the more realistic benchmark show why ... | [
"The benchmark-focused method evaluates on KITTI and Ford multi-AV with a restricted setup: it assumes city-scale retrieval has already narrowed the query to a 40 m Γ 40 m region around the satellite image center, uses paired query/satellite references for high-accuracy pose search, and even notes that orientation ... | 2 | a2493f457ba2a8200504ca2221481a1f2e966bea | 2408.05475 | Cross-view image geo-localization with Panorama-BEV Co-Retrieval Network | [
"360daf6da2d042e24f14a2aa53257591d851d305",
"7d2fe9c407a8e074964b23f42501fbd1991fa55b"
] | [
"2204.00097",
"2011.12172"
] | [] | [] | [
"wrong_paper",
"seed_only",
"correct",
"no_context"
] | 2 | null | null |
mt_0060 | multi-target | 1 | METHOD | contrast | NeurIPS | silver | The seed attributes slow convergence in vanilla low-rank adaptation to its default random initialization and cites both the original method paper and a later large-model study on code and math as context. Under that framing, how does the original low-rank adaptation approach initialize and position itself relative to f... | [
"In the original LoRA paper (Hu et al., 2021), the update is parameterized as \\(\\Delta W = BA\\) with one factor initialized to zero and the other with a small random Gaussian so the model starts exactly at the pretrained solution, and LoRA is positioned as matching or exceeding full fine-tuning while training on... | 2 | ac5d0faa5b07c884e21f423b8c4f22b3c1c3a01a | 2407.05000 | LoRA-GA: Low-Rank Adaptation with Gradient Approximation | [
"a8ca46b171467ceb2d7652fbfb67fe701ad86092",
"be12e6806e9190c3954cb34af7a6923b65cfedba"
] | [
"2106.09685",
"2405.09673"
] | [] | [] | [
"no_context",
"wrong_paper",
"correct",
"seed_only"
] | 3 | null | null |
mt_0061 | multi-target | 1 | RESULT | reason | NeurIPS | gold | The seed groups an earlier sentence-specific latent-vector method with a later baseline that extracts behavior-level vectors from many contrastive multiple-choice prompt pairs, treating both as frozen-model activation steering but criticizing direct prompt-based extraction for not always matching generation. Why did th... | [
"The behavior-level baseline argued that paired A/B prompts isolate the target behavior because only the answer option is varied while the rest of the prompt is kept constant, which isolates the internal representation most related to the target behavior and cancels out other confounding variables. The earlier sent... | 2 | af7acebf6c8b9adf38a9a2dd7cc6e6055b47ebcd | 2406.00045 | Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization | [
"edd1dc1e8d7989f36c0e54f69f4aeb5e597edc8b",
"42b6b7ae57b2b784be7fa78bf98b1c61d2b62751"
] | [
"2312.06681",
"2205.05124"
] | [] | [] | [
"wrong_paper",
"no_context",
"correct",
"seed_only"
] | 3 | null | null |
mt_0062 | multi-target | 1 | GROUND | reason | ECCV | silver | Texture-GS groups two earlier 3D Gaussian editing lines under the claim that appearance manipulation remains inconvenient because they keep the original entangled 3D-GS representation. Why did the SAM-assisted grouped-scene editing approach and the text-instruction-based delicate editing approach both choose to operate... | [
"The grouped-scene editing approach relies on the explicit, discrete structure of 3D Gaussians so that each Gaussian group can operate independently, enabling direct deletion, relocation, or color/style updates on selected parts while avoiding the difficulty of decomposing neural-field representations. The text-ins... | 0 | b16378d058d3c8ec639ce0cbf2eded59125f752e | 2403.10050 | Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splatting Editing | [
"5f7867681c6287cc4f4e7fb5260af8067b52b3d9",
"ab342050a0eb26fb4a0af5fff028e1a8ca54f575"
] | [
"2312.00732",
"2311.16037"
] | [] | [] | [
"correct",
"seed_only",
"wrong_paper",
"no_context"
] | 3 | null | null |
mt_0063 | multi-target | 1 | MOTIVE | justify | ECCV | bronze | The seed claims that its revised densification is motivated not just by the original Gaussian-splatting baseline, but also by the fact that the aliasing-focused extension still inherits the same adaptive density control. What concrete evidence from the original method and from the aliasing-focused extension justifies t... | [
"The original 3DGS adaptive density control grows primitives by tracking, for each Gaussian, the averaged positional gradient magnitude across rendered views and triggering either splitting for large primitives or cloning for smaller ones when that score exceeds a user-defined threshold. The aliasing-focused Mip-Sp... | 2 | b2ec0c7baaa5a6717f7eee14fa6b9dae348e8569 | 2404.06109 | Revising Densification in Gaussian Splatting | [
"2cc1d857e86d5152ba7fe6a8355c2a0150cc280a",
"7e305e37b49feb8899ae94f641c4f2d0d01ef054"
] | [
"2308.04079",
"2311.16493"
] | [] | [] | [
"seed_only",
"no_context",
"correct",
"wrong_paper"
] | 2 | null | null |
mt_0065 | multi-target | 1 | RESULT | assess | CVPR | bronze | The seed contrasts its dyadic, full face-body-hand, photorealistic setting with two monadic diffusion baselines: one that synthesizes co-speech body motion from audio and one that focuses only on facial motion while disentangling lip and non-lip dynamics. Looking at the empirical evidence those two baselines provide, w... | [
"The paperβs empirical results are for image synthesis: it evaluates unconditional diffusion models on LSUN bedroom, horse, and cat, and conditional models on ImageNet at 128Γ128, 256Γ256, and 512Γ512. Its strongest claims are supported by state-of-the-art FID and sFID results on these image-generation benchmarks, ... | 3 | b8bc88b75976f6a1909aeb7043e32f0ecd4567f9 | 2401.01885 | From Audio to Photoreal Embodiment: Synthesizing Humans in Conversations | [
"f352c11fad126a22deded3f904aa87eca5fc5df1",
"458e3d2be80c401fb47e562d9d57012bd63da1c3"
] | [
"2211.09707",
"2212.04248"
] | [] | [] | [
"wrong_paper",
"seed_only",
"no_context",
"correct"
] | 2 | null | null |
mt_0066 | multi-target | 1 | RESULT | reason | NeurIPS | bronze | The seed points to two earlier T2V benchmarks that already moved beyond single-number evaluationβone by expanding prompt scale and objective metrics, and the other by building a multi-dimensional, multi-category suiteβyet uses them to motivate a separate dynamics-centered protocol. Why did the prompt-and-metric expansi... | [
"The larger-prompt/objective-metric benchmark (e.g., EvalCrafter, with 700 prompts and metrics including FVD, CLIPScore, aesthetic quality, and motion-based measures) reports that videos with only small motion can still win because current T2V models trade off motion magnitude against text alignment and visual qual... | 2 | bd940782f2e9cd1c812ae8ada3d848df5056afe0 | 2407.01094 | Evaluation of Text-to-Video Generation Models: A Dynamics Perspective | [
"4e9a8141da2a8c603722b07d096109207f8e0b66",
"66d927fdb6c2774131960c75275546fd5ee3dd72"
] | [
"2311.17982",
"2310.11440"
] | [] | [] | [
"no_context",
"seed_only",
"correct",
"wrong_paper"
] | 2 | null | null |
mt_0067 | multi-target | 1 | MOTIVE | synthesize | ECCV | bronze | The seed motivates Face-Adapter by arguing that a diffusion-based reenactment refiner still cannot fix blur under large pose variation, even though it improves image quality, while a StyleGAN2-based reenactment line is used to improve generation quality. Using the diffusion refiner's own quantitative and qualitative ev... | [
"The paper states that FADM improves image quality as a diffusion-based refiner but still fails to resolve the blurring issue caused by large pose variation, while HyperReenact uses StyleGAN2 to improve generation quality yet remains weak on large-pose VoxCeleb-style reenactment. On the VoxCeleb1 benchmark, HyperRe... | 3 | be5f580d978b9c2b07d061217d54f2bb1edf6edf | 2405.12970 | Face Adapter for Pre-Trained Diffusion Models with Fine-Grained ID and Attribute Control | [
"f47a362d13c2e32ddd6204708398d6c5fdd9f338",
"4fa85f06ad62139aa10b16112ddffae257ae4ce0"
] | [
"2304.03199",
"2307.10797"
] | [] | [] | [
"seed_only",
"no_context",
"wrong_paper",
"correct"
] | 2 | null | null |
mt_0068 | multi-target | 1 | MOTIVE | reason | NAACL | gold | The seed groups two neuro-symbolic attempts to fix unfaithful reasoning: one that parses problems into declarative specifications and lets an automated theorem solver handle the rest, and another that translates questions into interleaved natural-language and symbolic reasoning chains executed by a deterministic solver... | [
"PaL argues for using the language model only to read the natural-language problem and generate interleaved natural-language and program statements, while offloading solving and calculation to a Python interpreter, because natural language understanding and decomposition require LLMs but βsolving and reasoning can ... | 1 | bf0479119679fb627883bffe6087722df5ba4e5b | 2409.17539 | Logic-of-Thought: Injecting Logic into Contexts for Full Reasoning in Large Language Models | [
"f27f6d1d521d189e78f5623098ced0deea613d33",
"b115c1e1e9e51f8ad7d47b745bc04e29a654b84d"
] | [
"2305.09656",
"2301.13379"
] | [] | [] | [
"wrong_paper",
"correct",
"no_context",
"seed_only"
] | 3 | null | null |
mt_0069 | multi-target | 1 | MOTIVE | contrast | ECCV | bronze | The seed groups two earlier CLIP-based DG methods as lightweight MLP approaches for conditional prompt learning, but distinguishes them by whether they adapt prompts per image or inject domain-specific features inferred from target-domain inputs. Under that framing, how do these two methods differ in what they conditio... | [
"The paper states that CoCoOp and DPL both learn an image-conditional vector with an MLP, but CoCoOp adds this residual vector to learn an instance-specific continuous prompt conditioned on the input image, while DPL concatenates the image-conditioned vector to add domain-specific features to the prompt template. I... | 2 | c36cb4a41369369d837ea170397f7818d02150dd | 2404.19286 | Soft Prompt Generation for Domain Generalization | [
"b879450f50a6113f44a5baf0bcd5b4331eeb7bbc",
"de0784f7c46a6f28bc3415c4c99a9e0ca5247ba3"
] | [
"2203.05557",
"2111.12853"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 2 | null | null |
mt_0075 | multi-target | 1 | RESULT | synthesize | ECCV | gold | For the two coupled 3D Gaussian SLAM baselines that the seed groups together as using dense photometric-error tracking, how do their shared-benchmark results on Replica and TUM-RGBD compare in terms of the tradeoff between tracking accuracy and runtime? Use the overlapping evaluation numbers to determine which one is f... | [
"The baseline with silhouette-masked RGB-and-depth tracking is more accurate on both shared benchmarks, reaching 0.36 cm ATE on Replica and 5.48 cm on TUM-RGBD, while the baseline with adaptive Gaussian expansion and coarse-to-fine tracking is faster at 8.34 FPS but less accurate at 0.50 cm on Replica and 3.7 cm on... | 0 | de975de4d6471983ce3e3969661979c2265b82d2 | 2403.12550 | Rgbd Gs-icp Slam | [
"a6c92dd8c537bc4a0c91cef8558e4a4f25416091",
"8b5f4902411123c5e68ded502f296cf325864831"
] | [
"2312.02126",
"2311.11700"
] | [] | [] | [
"correct",
"no_context",
"seed_only",
"wrong_paper"
] | 3 | null | null |
mt_0076 | multi-target | 1 | MOTIVE | reason | ECCV | bronze | The seed contrasts two strong sparse-query baselines that still do not explicitly model 3D scene structure: one improves detection by propagating object queries over time, and the other strengthens a fully sparse detector with adaptive BEV/image-space operations plus query denoising. Why did the first paper treat tempo... | [
"The paper states that recent methods such as temporal approaches and specialized losses improve multi-view 3D detection, but they βdo not explicitly consider the 3D structure of the scene,β while false positives arise because each pixel lacks accurate depth and its position embedding βcan only encode the ray direc... | 2 | e46831f597b8d151cfa74ac90fc1208757818609 | 2402.03634 | Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection | [
"73385e2aac9890073669759dfde8800b5704ab6e",
"1feb7e53dbb33cd6aa75bb20fcec62e153ccddf2"
] | [
"2303.11926",
"2308.09244"
] | [] | [] | [
"seed_only",
"wrong_paper",
"correct",
"no_context"
] | 2 | null | null |
mt_0077 | multi-target | 1 | RESULT | retrieve | ECCV | bronze | In DreamSceneβs explanation of FPS, the late-stage texture refinement combines a reconstruction method based on explicit 3D Gaussians with an earlier text-to-3D optimization strategy that uses monotonically non-increasing diffusion timesteps. What concrete speed figures do those two cited works report for their own sys... | [
"DreamScene cites the explicit 3D Gaussian reconstruction approach as βachieving superior real-time rendering,β and reports that its own reconstructive generation step based on that method βgenerates a representation with detailed and plausible textures using only 15 seconds.β For the timestep-scheduling line of wo... | 3 | e61a32cf3b253011cea5d4d88429ece876a1b6ae | 2404.03575 | DreamScene: 3D Gaussian-based Text-to-3D Scene Generation via Formation Pattern Sampling | [
"2cc1d857e86d5152ba7fe6a8355c2a0150cc280a",
"8dfe271d2186d5746d034b3cce12131f4d3f45f7"
] | [
"2308.04079",
"2306.12422"
] | [] | [] | [
"seed_only",
"no_context",
"wrong_paper",
"correct"
] | 2 | null | null |
mt_0078 | multi-target | 1 | GROUND | contrast | ICLR | gold | MotionClone frames two foundational text-to-video backbones together: one earlier latent-diffusion approach that adds both 3D-convolutional and temporal-attention layers to align frames over time, and one later approach that enhances a pre-trained text-to-image model with a plug-and-play motion module and optional ligh... | [
"MotionClone contrasts VideoLDM as a backbone that βintroduces a motion module that utilizes 3D convolutions and temporal attention to capture frame-to-frame correlationsβ with AnimateDiff, which βenhances a pre-trained T2I diffusion model with motion modeling capabilitiesβ by fine-tuning specialized temporal atten... | 3 | e872ad446a3c6986feea7bd5cbd5b5754c068ab6 | 2406.05338 | MotionClone: Training-Free Motion Cloning for Controllable Video Generation | [
"c1caa303549764d220ff17dc1785985dd1ba6047",
"f5a0c57f90c6abe31482e9f320ccac5ee789b135"
] | [
"2307.04725",
"2304.08818"
] | [] | [] | [
"seed_only",
"wrong_paper",
"no_context",
"correct"
] | 3 | null | null |
mt_0079 | multi-target | 1 | MOTIVE | reason | ECCV | gold | The seed contrasts two NeRF-based predecessors: one that keeps each training image fixed in time and adds small MLPs to learn blur effects inside rendering, and another that instead models each frame through a short camera trajectory during exposure. Why did the first approach justify learning a deformable sparse blur ... | [
"The paper contrasts prior methods by stating that Deblur-NeRF and a 3DGS predecessor βtreat the training images as fixed in time and focus on extracting sharper images from blurred inputs,β using additional learnable parameters in small MLPs, whereas BAD-NeRF instead βexplicitly model[s] the blur formation process... | 3 | eed0bb5e4738f1fc7620f181e68dae2552454a05 | 2403.13327 | Gaussian Splatting on the Move: Blur and Rolling Shutter Compensation for Natural Camera Motion | [
"7bc683fe1911f4987b845f7d2165d5888f0a50e9",
"314f1d10c228b32a3d64550f20841786c7f30c40"
] | [
"2111.14292",
"2211.12853"
] | [] | [] | [
"seed_only",
"no_context",
"wrong_paper",
"correct"
] | 3 | null | null |
mt_0083 | multi-target | 1 | RESULT | contrast | ICLR | gold | The seed contrasts one context-compression baseline that deletes unimportant tokens using the input question with another that adapts a pretrained LM to compress prior segments into reusable soft prompts. Under that framing, how do these two approaches differ in what they require to compress context and in the empirica... | [
"The question-aware token-deletion approach uses the question to guide coarse-to-fine prompt compression and reports that on NaturalQuestions it gains a performance boost of 17.1% with βΌ 4x fewer input tokens, while compressing βΌ10K-token prompts at 2xβ6x yields 1.4xβ2.6x end-to-end latency speedups. The soft-promp... | 0 | f69f494ab691481ec353da4be480b334fada6607 | 2401.03462 | Long Context Compression with Activation Beacon | [
"4c0428917aeee6aa7bd434f337d039f35996b736",
"2f7364d8e5cf94315bf8905f57de9c5543e9a4bf"
] | [
"2310.06839",
"2305.14788"
] | [] | [] | [
"correct",
"wrong_paper",
"seed_only",
"no_context"
] | 3 | null | null |
mt_0084 | multi-target | 1 | RESULT | synthesize | ECCV | gold | The seed groups two baselines under backpropagation-through-reward methods and suggests they are especially prone to reward hacking despite strong optimization performance. Using the human-preference reward model paper that also introduces a direct tuning baseline, together with the differentiable-reward fine-tuning pa... | [
"In the human-preference reward model paper, the 1.3B human feedback model was preferred over reference summaries 61% of the time, outperforming a supervised model 10Γ its size that reached 43%. The differentiable-reward fine-tuning paper reports an explicit collapse mode in which, with stronger optimization, true ... | 2 | f86b13cf11c4fb2d67e0cc1090ede5b38c6b85a1 | 2401.12244 | Large-scale Reinforcement Learning for Diffusion Models | [
"1b2355c3c674b26a977768a91a164384ad51bbb1",
"a395f1c0910a5622f6faaefb5b27e87bb1617759"
] | [
"2304.05977",
"2309.17400"
] | [] | [] | [
"wrong_paper",
"seed_only",
"correct",
"no_context"
] | 3 | null | null |
mt_0087 | multi-target | 2 | MOTIVE | contrast | CVPR | gold | The seed argues that earlier ways of reducing overfitting are limited because they do not directly optimize zero-shot robust accuracy, and it points to both lightweight input-space adaptation and self-supervised robustness repair as relevant precursors. Compared along that seed-defined axis, how does the approach that ... | [
"One approach learns a single image perturbation, or βsoft prompt,β by backpropagation while keeping the model parameters frozen. Separately, the paper reports that as models learn from more labeled ImageNet data, they become more accurate on ImageNet but these gains do not carry over to robustness benchmarks, with... | 3 | 15e318ceb6b972098590fc3b7214fe49fb3e3de4 | 2401.04350 | Pre-Trained Model Guided Fine-Tuning for Zero-Shot Adversarial Robustness | [
"aa8c61c9f6bb21c57e49611ccb995cfda1b53b10",
"191038455e19980163ac1ec3acd3c295fc2952b6"
] | [
"2203.17274",
"2204.03714"
] | [
"16596dd03fa40ba278f9533ea9986982dcc81fb6"
] | [
"2212.07016"
] | [
"wrong_paper",
"no_context",
"seed_only",
"correct"
] | 3 | true | pass |
mt_0092 | multi-target | 2 | RESULT | assess | NeurIPS | gold | DigiRL argues that earlier web-control resources either restrict exploration through cached states or evaluate narrower instruction-following setups than the autonomous, stochastic device-control setting it targets. For the benchmark built from snapshots of many real websites and the customer-support system built from ... | [
"For WebArena, the reported numbers come from a cached-snapshot benchmark over 134 tasks on four self-hosted websites, so agents are evaluated on a fixed, non-live web with constrained exploration rather than the open, stochastic Internet; the headline result is that strong prompting/GPT-4-style agents still solve ... | 3 | c6e3e5f794233da2b27ea5dd4af19e62b95c863f | 2406.11896 | DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning | [
"58f8925a8b87054ad0635a6398a7fe24935b1604",
"1905db8aef0d63d594b38a53939206fbef9fcd65"
] | [
"2306.06070",
"2103.16057"
] | [
"34fc623e43f60b73827f4d0259013b96aeddfe6f",
"61c723e7d91ca308b9eeacab35a2fce7548ef9fc",
"c844694387a89a477e7a8bbf918171cdc3b85672",
"e41482f4ee984f17382f6cdd900df094d928be06",
"e50583008fe4ec049e42fdc01727ce98f6d86a35"
] | [
"2307.13854",
"2312.08914",
"2312.13771",
"2401.01614",
"2404.09992"
] | [
"no_context",
"wrong_paper",
"seed_only",
"correct"
] | 3 | true | pass |
AgentHop
AgentHop is a diagnostic benchmark for agentic multi-hop question answering over scientific citation graphs: 1,011 four-option multiple-choice questions grounded in citation chains over 7,205 arXiv papers from nine major computer-science venues (2022β2025). Each question is answerable only by reading specific sections of one or two of these papers; the agent reaches the answer-bearing papers by following references from a given seed paper.
The dataset is paired with a four-axis decomposition framework (search, synthesis, tool-use pattern, resource management) that turns a single evaluation run into per-model failure-mode attribution rather than a single accuracy number.
Dataset Statistics
| Statistic | Value |
|---|---|
| Total items | 1,011 |
| Single-target (ST) | 568 |
| Multi-target (MT) | 443 |
| Source-paper pool | 7,205 papers |
| Source citation edges | >450,000 |
| Source venues | 9 (NeurIPS, ICML, ICLR, ACL, EMNLP, NAACL, CVPR, ECCV, SIGIR) |
| Publication years | 2022β2025 |
| Reasoning-type counts | GROUND 330, RESULT 292, MOTIVE 265, METHOD 124 |
| Consensus tiers | Gold / Silver / Bronze (filter-agreement signal at construction time) |
| Depths | 1-hop and 2-hop |
Files
AgentHop/
βββ qa/full.jsonl # 1,011 MCQ items (schema below)
βββ graphs/full.jsonl # per-sample citation neighbourhoods (nodes + edges)
βββ paper_pool/papers.jsonl # deduplicated section text for the 7,205-paper pool
βββ audit/recall_labels.jsonl # auditor-labelled answer-bearing sections per sample
βββ generator_control/ # 90-item split written by a different generator (see below)
β βββ qa/full.jsonl
β βββ graphs/full.jsonl
β βββ paper_pool/papers.jsonl
β βββ audit/recall_labels.jsonl
βββ loader.py # reference Python loader
βββ LICENSE # CC-BY 4.0
βββ README.md # this file
Schema (qa/full.jsonl)
Each line is one item.
| Field | Type | Description |
|---|---|---|
id |
string | Item identifier (e.g., st_0001, mt_0042) |
question |
string | The question stem |
options |
list of 4 strings | The four answer options |
correct_index |
int (0β3) | Index of the correct option in options |
question_type |
string | single-target or multi-target (a.k.a. target multiplicity) |
depth |
int | Maximum hop distance from seed to gold paper, 1 or 2 |
reasoning_type |
string | GROUND / METHOD / MOTIVE / RESULT |
cognitive_skill |
string | Construction-stage taxonomy label (e.g., retrieve) |
consensus_tier |
string | gold / silver / bronze β filter agreement at Stages 5β6 |
filter_agreement |
int | Numeric agreement score |
seed_arxiv_id |
string | arXiv ID of the seed paper the agent starts from |
seed_paper_id |
string | Semantic Scholar hash of the seed paper |
seed_title |
string | Seed paper title |
venue |
string | Source venue (e.g., NeurIPS, ICML) |
gold_arxiv_ids |
list of strings | arXiv IDs of the gold papers (1 for ST, 2 for MT) |
gold_paper_ids |
list of strings | Semantic Scholar hashes of the gold papers |
bridge_arxiv_ids |
list of strings | arXiv IDs of intermediate hop papers (depth-2 only) |
bridge_paper_ids |
list of strings | Semantic Scholar hashes of the bridge papers |
distractor_types |
list of strings | Engineered distractor labels: seed_only, wrong_paper, no_context, plus correct |
Schema (graphs/full.jsonl)
Each line is one sample's citation neighbourhood used as the agent's navigation pool.
| Field | Type | Description |
|---|---|---|
sample_id |
string | Matches qa/full.jsonl::id |
seed_paper_id |
string | Semantic Scholar hash of the seed paper |
nodes |
object | {paper_id: {arxivId, title, year, authors, abstract, ...}} |
edges |
object | {paper_id: [cited_paper_ids]} |
Schema (paper_pool/papers.jsonl)
Deduplicated section text for every paper appearing in any sample's citation neighbourhood.
| Field | Type | Description |
|---|---|---|
paper_id |
string | Semantic Scholar hash |
arxiv_id |
string | arXiv ID |
title |
string | |
year |
int | |
authors |
list | |
abstract |
string | |
sections |
list of {header, text} |
Section-structured body, parsed from arXiv via ar5iv |
Schema (audit/recall_labels.jsonl)
Auditor-labelled answer-bearing sections, used by the evaluation harness to compute the search-axis recall metric.
| Field | Type | Description |
|---|---|---|
sample_id |
string | Matches qa/full.jsonl::id |
analysis |
object | {section_recall_labels: [{arxiv_id, section}, ...], answer_validity, chain_coherence, query_quality, structural_integrity, synthesis_check} |
Quick start
from loader import load_agenthop, load_recall_labels
samples = load_agenthop("path/to/AgentHop")
labels = load_recall_labels("path/to/AgentHop")
print(len(samples)) # 1011
print(samples[0]["question"])
print(samples[0]["options"])
print(samples[0]["correct_index"])
print(samples[0]["paper_pool"]) # papers reachable from seed for this sample
print(labels[samples[0]["id"]]["analysis"]["section_recall_labels"])
To evaluate a model end-to-end (with the seven-tool sandbox, four-axis metrics, and resource caps), use the evaluation harness from the code repo.
Generator-control split (generator_control/)
The 1,011 main items were written by GPT-5.4, which is also one of the evaluated models. To measure whether that gives it an advantage, we rebuilt the pipeline on 98 fresh seeds (no overlap with the main set) with a different question writer, Inkling, keeping every other stage unchanged. Of the 94 items that passed Stage 6, 90 remain after the same human audit, triage, and section-length filter as the main set (56 single-target, 34 multi-target). GPT-5.4 scores 80.0% on these items (95% CI 70.6β87.0) against 79.0% on the main set; see the paper's Limitations section.
The split uses the same files and schema as the main set, so the loader works unchanged:
samples = load_agenthop("path/to/AgentHop/generator_control") # 90 items, ids gc_st_* / gc_mt_*
Two extra fields: generator ("Inkling") and source_id (the construction-time id). The split is not part of the main AgentHop score; report it separately. One item (gc_st_0056) was evaluated without its seed paper's text in the sandbox; the released pool includes it.
Intended Use
AgentHop is a measurement instrument for multi-step QA agents: locate relevant papers via citation following, read the right sections of those papers, combine the resulting evidence, and do so within a deployment-realistic resource budget. The headline accuracy is intended to be paired with the four-axis decomposition (search recall, conversion rate, tool-use pattern, resource management) that the harness produces. Aggregate accuracy alone collapses qualitatively different failure modes into a single number; the per-axis attribution surfaces which sub-ability is binding for each model.
Construction Pipeline
AgentHop is constructed via an eight-stage pipeline: four generation stages (seed selection β citation-chain expansion β corpus collection β Q/A generation β distractor generation) followed by four filtering stages (model-ensemble filtering β data sanity check β seven-auditor human review β post-audit triage). Full details are in the paper's appendix.
Limitations
- Single domain. Computer-science papers from nine major venues (2022β2025). Behaviour on other fields, on long-tail venues, and on open-ended response formats is not measured.
- Multiple-choice format. Each item is a four-option MCQ, not a free-form generation task.
- English-language source corpus. All papers and questions are in English.
- Generatorβevaluator overlap. GPT-5.4 generated the items and is also among the evaluated models. The effect is bounded but real and acknowledged in the paper's Limitations section.
- Parametric coverage. The strongest checkpoints already resolve a substantial fraction of items closed-book (Gemini-3 Pro reaches 0.734 without tools); the four-axis decomposition is most informative where parametric coverage is sparse.
Ethical Considerations
- Built from public arXiv papers under their respective deposit licenses. We use only paper identifiers, abstracts, and named-section text; we do not reproduce figures, tables, or any non-textual content.
- No personal information, no harmful or offensive content, and no generative model weights are released.
- Seven auditors (members of the research team, participating voluntarily) reviewed audit-prep items during the construction pipeline. No external paid annotation was conducted; no personal information of auditors is collected or released.
Changelog
- v1.1.0 (2026-11-01). Added the
generator_control/split (90 items). The main split is unchanged. - v1.0.1 (2026-11-01).
audit/recall_labels.jsonl: eightdirectlabels in seven samples (st_0001, st_0245, st_0302, st_0319, st_0330, st_0515, st_0521) named a table (e.g. "Table 3") instead of a readable section. Each now names the section that contains the table, and the table reference is kept at the start ofwhat_it_contains. No questions, options, or answers changed.croissant.json: author and provenance metadata added. - v1.0.0 (2026-05-04). Initial release.
Citation
@inproceedings{park2026agenthop,
title = {AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Question Answering over Scientific Citation Graphs},
author = {Park, Chanhee and Yoon, Jeongho and Han, Sungbin and Moon, Hyeonseok and Lim, Heuiseok},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
year = {2026}
}
- Downloads last month
- 320