Datasets:
Update papers and claims (part 4)
Browse files- site/p/2605.15383.json +1 -1
- site/p/2605.15846.json +1 -1
- site/p/2605.16371.json +1 -1
- site/p/2605.18421.json +1 -1
- site/p/2605.18859.json +1 -1
- site/p/2605.19127.json +1 -1
- site/p/2605.19382.json +1 -1
- site/p/2605.22469.json +1 -1
- site/p/2605.22907.json +1 -1
- site/p/2605.28345.json +1 -1
- site/p/2605.28683.json +1 -1
- site/p/2605.28911.json +1 -1
- site/p/2605.29059.json +1 -1
- site/p/2605.29253.json +1 -1
- site/p/2605.30326.json +1 -1
- site/p/2605.31572.json +1 -1
- site/p/2606.01848.json +1 -1
- site/p/2606.08340.json +1 -1
- site/p/2606.14299.json +1 -1
- site/p/2606.15320.json +1 -1
- site/p/2606.15749.json +1 -1
- site/p/2606.17541.json +1 -1
- site/p/2606.17833.json +1 -1
- site/p/2606.21116.json +1 -1
- site/p/2606.23061.json +1 -1
- site/p/2606.23669.json +1 -1
- site/p/2606.24819.json +1 -1
- site/p/2607.02680.json +1 -1
site/p/2605.15383.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.15383","title":"MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays","authors":"Emre Hayir; Lorin Crawford; Alex X Lu","openreview_id":"Lw6zDn55qc","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Cell Painting is a microscopy assay that stains cells with six dyes over five channels, used in drug discovery and functional genomics to read out how chemical or genetic perturbations change cell morphology. Images are turned into feature vectors either with classical hand-crafted pipelines (CellProfiler) or with deep models (natural-image models such as DINOv2 and ResNet, or microscopy-specific foundation models such as OpenPhenom, SubCell and CLOOME). Evaluations of these representations are fragmented, use different preprocessing and metrics, and often ignore batch effects (technical variation between batches, laboratories and plate layouts), so it is unclear whether newer learned representations actually beat CellProfiler.","method":["MorphoHELM evaluates frozen representations on three task families (MoA enrichment, gene pathway enrichment, replicate retrieval) at increasing batch-effect stringency: No Restriction (NR), Not Same Batch (NSB), Not Same Source/institution (NSS) and Not Same plate Layout (NSL), where candidates sharing the query's batch, source or plate position are excluded from retrieval. All representations go through the same pipeline: illumination correction, QC (channel de-duplication, removal of wells in the bottom 5% of cell count, removal of batches with unstable negative controls), then per-plate scaling, PCA to 64 components over the whole dataset, and per-plate MAD robust standardization.","For the chosen claim, the replicate retrieval task on cpg-target2 (a JUMP-CP subset of 301 compounds with strong phenotypes, 11 sources, 107 batches, 141 plates, 64,485 wells) queries individual well-level profiles and measures kNN Recall@1: for each query well, whether its nearest neighbour by cosine similarity (within the candidate set allowed by the stringency level) is treated with the same compound, with accuracy macro-averaged first per compound and then across compounds. Expected random Recall@1 is 0.33%. Compared representations: CellProfiler, ImageNet ResNet101, DINOv2 ViT-B/14 (CLS token and mean of patch tokens), OpenPhenom, SubCell, CLOOME, and a randomly initialized ResNet101."],"claim":"On cpg-target2 replicate retrieval, postprocessed CellProfiler features reach kNN Recall@1 of 25.38% (NR), 28.49% (NSB) and 14.64% (NSS), above every deep representation including the best, DINOv2 CLS (20.33%, 21.73%, 12.77%) and CLOOME (18.37%, 22.24%, 5.65%); CellProfiler also has the highest mAP (4.66 vs. 3.53 for DINOv2 CLS at NR).","source":"Figure 4, page 8","metric":"kNN Recall@1 in percent, macro-averaged per compound then across compounds (higher is better) and mAP over all samples (higher is better); single deterministic evaluation, no error bars, seeds or confidence intervals reported","comparison":"CellProfiler vs. DINOv2 CLS, DINOv2 Patch, ImageNet ResNet101, OpenPhenom, SubCell, CLOOME (and untrained ResNet101) under the identical scaling + PCA(64) + MAD postprocessing pipeline","falsification":"Re-running the cpg-target2 replicate retrieval with the same QC and postprocessing, some deep representation (most likely DINOv2 CLS) matches or exceeds CellProfiler's Recall@1 in the NR or NSB setting, or CellProfiler's margin over the best deep model at NR shrinks below about 2 percentage points.","caption":"Figure 4: kNN replicate retrieval.","image":"img/2605.15383.png","kind":"figure","image_w":1152,"image_h":1111,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.15383","title":"MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays","authors":"Emre Hayir; Lorin Crawford; Alex X Lu","openreview_id":"Lw6zDn55qc","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Cell Painting is a microscopy assay that stains cells with six dyes over five channels, used in drug discovery and functional genomics to read out how chemical or genetic perturbations change cell morphology. Images are turned into feature vectors either with classical hand-crafted pipelines (CellProfiler) or with deep models (natural-image models such as DINOv2 and ResNet, or microscopy-specific foundation models such as OpenPhenom, SubCell and CLOOME). Evaluations of these representations are fragmented, use different preprocessing and metrics, and often ignore batch effects (technical variation between batches, laboratories and plate layouts), so it is unclear whether newer learned representations actually beat CellProfiler.","method":["MorphoHELM evaluates frozen representations on three task families (MoA enrichment, gene pathway enrichment, replicate retrieval) at increasing batch-effect stringency: No Restriction (NR), Not Same Batch (NSB), Not Same Source/institution (NSS) and Not Same plate Layout (NSL), where candidates sharing the query's batch, source or plate position are excluded from retrieval. All representations go through the same pipeline: illumination correction, QC (channel de-duplication, removal of wells in the bottom 5% of cell count, removal of batches with unstable negative controls), then per-plate scaling, PCA to 64 components over the whole dataset, and per-plate MAD robust standardization.","For the chosen claim, the replicate retrieval task on cpg-target2 (a JUMP-CP subset of 301 compounds with strong phenotypes, 11 sources, 107 batches, 141 plates, 64,485 wells) queries individual well-level profiles and measures kNN Recall@1: for each query well, whether its nearest neighbour by cosine similarity (within the candidate set allowed by the stringency level) is treated with the same compound, with accuracy macro-averaged first per compound and then across compounds. Expected random Recall@1 is 0.33%. Compared representations: CellProfiler, ImageNet ResNet101, DINOv2 ViT-B/14 (CLS token and mean of patch tokens), OpenPhenom, SubCell, CLOOME, and a randomly initialized ResNet101."],"claim":"On cpg-target2 replicate retrieval, postprocessed CellProfiler features reach kNN Recall@1 of 25.38% (NR), 28.49% (NSB) and 14.64% (NSS), above every deep representation including the best, DINOv2 CLS (20.33%, 21.73%, 12.77%) and CLOOME (18.37%, 22.24%, 5.65%); CellProfiler also has the highest mAP (4.66 vs. 3.53 for DINOv2 CLS at NR).","source":"Figure 4, page 8","metric":"kNN Recall@1 in percent, macro-averaged per compound then across compounds (higher is better) and mAP over all samples (higher is better); single deterministic evaluation, no error bars, seeds or confidence intervals reported","comparison":"CellProfiler vs. DINOv2 CLS, DINOv2 Patch, ImageNet ResNet101, OpenPhenom, SubCell, CLOOME (and untrained ResNet101) under the identical scaling + PCA(64) + MAD postprocessing pipeline","falsification":"Re-running the cpg-target2 replicate retrieval with the same QC and postprocessing, some deep representation (most likely DINOv2 CLS) matches or exceeds CellProfiler's Recall@1 in the NR or NSB setting, or CellProfiler's margin over the best deep model at NR shrinks below about 2 percentage points.","caption":"Figure 4: kNN replicate retrieval.","image":"img/2605.15383.png","kind":"figure","image_w":1152,"image_h":1111,"pwc":true}
|
site/p/2605.15846.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.15846","title":"RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades","authors":"Xinbo Xu; Ruihan Yang; Haiyang Shen; Wendong XU; Bofei Gao; Ruoyu Wu; Kean Shi; Weichu Xie; Xuanzhong Chen; Ming Wu; Jason Zeng; Michael Heinrich; Liang Chen; Kuan Li; Baobao Chang","openreview_id":"ENBv2lHkyZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Most coding-agent benchmarks, such as SWE-bench Verified, test single-issue bug fixes with oracle patches of tens to about a hundred lines, mostly in Python, and score each task as pass or fail. Frontier models now score above 80% on such benchmarks. Real software development, by contrast, involves version upgrades that take months of coordinated work across many files. RoadmapBench asks whether agents can do this longer, multi-target work, and gives partial credit per target.","method":["Each of the 115 tasks comes from a real consecutive release pair in one of 17 open-source repositories (Python, TypeScript, C++, Go and Rust). The agent starts in a Docker container with the repository at the source version. The git history is kept, but all branches and tags beyond the source release are pruned, upstream repository access is blocked, and the agent cannot see target-version code, the test files or the oracle patch. The agent receives a roadmap instruction (instruction.md) listing 3 to 12 targets (median 5) that describe the API signatures and behaviours to implement, but not how to implement them. The median oracle patch is 3,714 lines across 51 files. Each target has its own adapted upstream test file and a weight. A task is resolved if all targets pass. The Completion Score is the mean over tasks of the weighted fraction of targets passed.","Thirteen models are run with the OpenHands scaffold, and a subset is also run with Terminus 2 (Harbor's reference agent). Each task gets a 2-hour wall-clock budget with extended thinking enabled (reasoning effort xhigh for Claude-Opus-4.7; high for GPT-5.4, Gemini-3.1-Pro, DeepSeek-V4-Pro and Seed-2.0-Pro; default for the others). There is one trial per model."],"claim":"Under OpenHands with a 2-hour per-task budget, the best model, Claude-Opus-4.7, resolves only 39.1% of the 115 RoadmapBench tasks (Completion Score 0.692). It is followed by Claude-Opus-4.6 at 32.2% and GPT-5.4 at 29.6%, and the other ten models resolve 5.2-20.9%. Claude-Opus-4.7 gets a similar 38.3% under Terminus 2.","source":"Table 2, page 7","metric":"Resolved rate (% of tasks where all target tests pass, higher is better) and Completion Score (mean weighted fraction of targets passed, 0-1). Single trial per model; no error bars or seeds reported.","comparison":"Thirteen frontier models under the same OpenHands scaffold and budget, plus a subset under Terminus 2. The claim concerns the top model's absolute resolved rate and the ordering of the top three models against the rest.","falsification":"Running the released benchmark with Claude-Opus-4.7 + OpenHands (2-hour budget, xhigh reasoning) gives a resolved rate far from 39.1% (for example 50% or more, or 25% or less), or the Completion Score is far from 0.692. Separately, the oracle patches fail their own target tests, which would make the task scores unreliable.","caption":"Table 2: Main results on RoadmapBench across 115 tasks, using a single trial per model. Domain columns report resolved rates (%). Task counts are ML & Data (36), Web & RPC (17), ORM & Val. (25), Infra. & Tool. (23), and UI & Ren. (14). Bold and underline denote the best and second-best domain-level results within each scaffold.","image":"img/2605.15846.png","kind":"table","image_w":1288,"image_h":1433,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.15846","title":"RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades","authors":"Xinbo Xu; Ruihan Yang; Haiyang Shen; Wendong XU; Bofei Gao; Ruoyu Wu; Kean Shi; Weichu Xie; Xuanzhong Chen; Ming Wu; Jason Zeng; Michael Heinrich; Liang Chen; Kuan Li; Baobao Chang","openreview_id":"ENBv2lHkyZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Most coding-agent benchmarks, such as SWE-bench Verified, test single-issue bug fixes with oracle patches of tens to about a hundred lines, mostly in Python, and score each task as pass or fail. Frontier models now score above 80% on such benchmarks. Real software development, by contrast, involves version upgrades that take months of coordinated work across many files. RoadmapBench asks whether agents can do this longer, multi-target work, and gives partial credit per target.","method":["Each of the 115 tasks comes from a real consecutive release pair in one of 17 open-source repositories (Python, TypeScript, C++, Go and Rust). The agent starts in a Docker container with the repository at the source version. The git history is kept, but all branches and tags beyond the source release are pruned, upstream repository access is blocked, and the agent cannot see target-version code, the test files or the oracle patch. The agent receives a roadmap instruction (instruction.md) listing 3 to 12 targets (median 5) that describe the API signatures and behaviours to implement, but not how to implement them. The median oracle patch is 3,714 lines across 51 files. Each target has its own adapted upstream test file and a weight. A task is resolved if all targets pass. The Completion Score is the mean over tasks of the weighted fraction of targets passed.","Thirteen models are run with the OpenHands scaffold, and a subset is also run with Terminus 2 (Harbor's reference agent). Each task gets a 2-hour wall-clock budget with extended thinking enabled (reasoning effort xhigh for Claude-Opus-4.7; high for GPT-5.4, Gemini-3.1-Pro, DeepSeek-V4-Pro and Seed-2.0-Pro; default for the others). There is one trial per model."],"claim":"Under OpenHands with a 2-hour per-task budget, the best model, Claude-Opus-4.7, resolves only 39.1% of the 115 RoadmapBench tasks (Completion Score 0.692). It is followed by Claude-Opus-4.6 at 32.2% and GPT-5.4 at 29.6%, and the other ten models resolve 5.2-20.9%. Claude-Opus-4.7 gets a similar 38.3% under Terminus 2.","source":"Table 2, page 7","metric":"Resolved rate (% of tasks where all target tests pass, higher is better) and Completion Score (mean weighted fraction of targets passed, 0-1). Single trial per model; no error bars or seeds reported.","comparison":"Thirteen frontier models under the same OpenHands scaffold and budget, plus a subset under Terminus 2. The claim concerns the top model's absolute resolved rate and the ordering of the top three models against the rest.","falsification":"Running the released benchmark with Claude-Opus-4.7 + OpenHands (2-hour budget, xhigh reasoning) gives a resolved rate far from 39.1% (for example 50% or more, or 25% or less), or the Completion Score is far from 0.692. Separately, the oracle patches fail their own target tests, which would make the task scores unreliable.","caption":"Table 2: Main results on RoadmapBench across 115 tasks, using a single trial per model. Domain columns report resolved rates (%). Task counts are ML & Data (36), Web & RPC (17), ORM & Val. (25), Infra. & Tool. (23), and UI & Ren. (14). Bold and underline denote the best and second-best domain-level results within each scaffold.","image":"img/2605.15846.png","kind":"table","image_w":1288,"image_h":1433,"pwc":true}
|
site/p/2605.16371.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.16371","title":"GeoSym127K: Scalable Symbolically-verifiable Synthesis for Multimodal Geometric Reasoning","authors":"Jinhao Jing; Zheng Ma; Jinwei Liang; Qiannian Zhao; Shuang Chen; Jing Yang; Prayag Tiwari; Jingjing Bai; Benyou Wang; Por L Yee; Zhan Su; Lewei Lu","openreview_id":"NO4INpEpCi","presentation":"oral","track":"evaluations_datasets","verdict":"borderline","background":"Large multimodal models (LMMs) still do poorly on geometry problems that need precise reading of a diagram and several reasoning steps. Existing geometry training data is either small and hand-annotated, or synthetic with answers and reasoning chains written by an LLM, which can be wrong. This paper builds a synthesis engine that constructs diagrams from exact symbolic coordinates, computes exact ground-truth answers with SymPy, and keeps only teacher-model chain-of-thought (CoT) solutions whose final answer matches the symbolic answer. The question is whether fine-tuning on such verified data improves geometric reasoning on public benchmarks more than fine-tuning on earlier synthetic geometry datasets.","method":["The GeoSym engine grows 2D geometric constructions with a type-conditional grammar, keeping all point coordinates as exact SymPy expressions, renders the diagrams (including shaded regions found by connected-component analysis), and asks length, angle, area, perimeter and shaded-area questions whose answers are computed symbolically. A teacher MLLM writes a caption and a CoT solution for each image/question; a CoT is kept only if Simplify(A_pred - A_GT) = 0. Problems are split into Entry, Hard and Expert tiers by construction depth; the SFT sets are GeoSym Entry (23,440 verified CoTs) and GeoSym Hard (23,835 verified CoTs), part of the 55K GeoSym-Instruct set.","SFT updates only the LLM backbone of Qwen3-VL-8B-Instruct (vision encoder and projector frozen) on 8 GPUs with DeepSpeed ZeRO-3: global batch 16, lr 1e-5 cosine, warmup 0.03, max context 12,288, BF16, for 1, 3, 5 or 10 epochs. Baselines are the same base model fine-tuned on TR-GeoMM (TR-CoT) and GeoTrust-train data with identical hyperparameters, data volume and evaluation. All models are evaluated with VLMEvalKit (temperature 0.7, top_p 0.95, top_k 20, max 32,768 tokens) on MathVista (geometry subsets), MathVerse Vision-only, MathVision and WeMath (strict score); 'Overall' is the mean of the four benchmark scores."],"claim":"SFT of Qwen3-VL-8B-Instruct on GeoSym Hard raises the Overall score (mean of MathVista, MathVerse Vision-only, MathVision, WeMath) to 63.18, versus 55.94 for the base model and 40.31 / 38.50 for the same model fine-tuned on GeoTrust-train / TR-GeoMM, with WeMath strict score 61.52 vs. 55.33 (base) and 32.19 (GeoTrust-train).","source":"Table 5, page 7 (per-epoch results in Table 16, page 34)","metric":"Accuracy (%) on each benchmark, higher is better; Overall is the mean of the four benchmark scores. Single run per configuration, no error bars or seeds reported.","comparison":"Qwen3-VL-8B-Instruct + GeoSym Hard SFT vs. the untuned Qwen3-VL-8B-Instruct and vs. the same model SFT'd on TR-GeoMM and GeoTrust-train under identical training and evaluation settings.","falsification":"Refuted if, after SFT on GeoSym Hard with the stated hyperparameters (any of 1-5 epochs), Qwen3-VL-8B-Instruct's Overall score is not clearly above the base model's (e.g. less than ~3 points above when both are evaluated with the same VLMEvalKit settings), or if SFT on GeoTrust-train or TR-GeoMM with matched data volume reaches a comparable Overall or WeMath score.","caption":"Table 5: Main SFT Results. The table details the macro-average scores and specific sub-category breakdowns. The highest values within each parameter-scale comparison group are highlighted with a light blue background and bold text. Asterisks (*) indicate models evaluated via API calls using identical evaluation settings. The dagger (†) denotes a specific anomalous result observed for the 8B baseline on MathVerse Vision-only, with a detailed discussion deferred to Appendix E.3. The double dagger (‡) indicates selected open source baseline synthesis methods for fair comparison with details reported in Appendix E.2. Note that our reported GeoSym scores represent the peak-performing training epoch; an ablation on epoch saturation is provided in Appendix E.4.","image":"img/2605.16371.png","kind":"table","image_w":1247,"image_h":1311,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.16371","title":"GeoSym127K: Scalable Symbolically-verifiable Synthesis for Multimodal Geometric Reasoning","authors":"Jinhao Jing; Zheng Ma; Jinwei Liang; Qiannian Zhao; Shuang Chen; Jing Yang; Prayag Tiwari; Jingjing Bai; Benyou Wang; Por L Yee; Zhan Su; Lewei Lu","openreview_id":"NO4INpEpCi","presentation":"oral","track":"evaluations_datasets","verdict":"borderline","background":"Large multimodal models (LMMs) still do poorly on geometry problems that need precise reading of a diagram and several reasoning steps. Existing geometry training data is either small and hand-annotated, or synthetic with answers and reasoning chains written by an LLM, which can be wrong. This paper builds a synthesis engine that constructs diagrams from exact symbolic coordinates, computes exact ground-truth answers with SymPy, and keeps only teacher-model chain-of-thought (CoT) solutions whose final answer matches the symbolic answer. The question is whether fine-tuning on such verified data improves geometric reasoning on public benchmarks more than fine-tuning on earlier synthetic geometry datasets.","method":["The GeoSym engine grows 2D geometric constructions with a type-conditional grammar, keeping all point coordinates as exact SymPy expressions, renders the diagrams (including shaded regions found by connected-component analysis), and asks length, angle, area, perimeter and shaded-area questions whose answers are computed symbolically. A teacher MLLM writes a caption and a CoT solution for each image/question; a CoT is kept only if Simplify(A_pred - A_GT) = 0. Problems are split into Entry, Hard and Expert tiers by construction depth; the SFT sets are GeoSym Entry (23,440 verified CoTs) and GeoSym Hard (23,835 verified CoTs), part of the 55K GeoSym-Instruct set.","SFT updates only the LLM backbone of Qwen3-VL-8B-Instruct (vision encoder and projector frozen) on 8 GPUs with DeepSpeed ZeRO-3: global batch 16, lr 1e-5 cosine, warmup 0.03, max context 12,288, BF16, for 1, 3, 5 or 10 epochs. Baselines are the same base model fine-tuned on TR-GeoMM (TR-CoT) and GeoTrust-train data with identical hyperparameters, data volume and evaluation. All models are evaluated with VLMEvalKit (temperature 0.7, top_p 0.95, top_k 20, max 32,768 tokens) on MathVista (geometry subsets), MathVerse Vision-only, MathVision and WeMath (strict score); 'Overall' is the mean of the four benchmark scores."],"claim":"SFT of Qwen3-VL-8B-Instruct on GeoSym Hard raises the Overall score (mean of MathVista, MathVerse Vision-only, MathVision, WeMath) to 63.18, versus 55.94 for the base model and 40.31 / 38.50 for the same model fine-tuned on GeoTrust-train / TR-GeoMM, with WeMath strict score 61.52 vs. 55.33 (base) and 32.19 (GeoTrust-train).","source":"Table 5, page 7 (per-epoch results in Table 16, page 34)","metric":"Accuracy (%) on each benchmark, higher is better; Overall is the mean of the four benchmark scores. Single run per configuration, no error bars or seeds reported.","comparison":"Qwen3-VL-8B-Instruct + GeoSym Hard SFT vs. the untuned Qwen3-VL-8B-Instruct and vs. the same model SFT'd on TR-GeoMM and GeoTrust-train under identical training and evaluation settings.","falsification":"Refuted if, after SFT on GeoSym Hard with the stated hyperparameters (any of 1-5 epochs), Qwen3-VL-8B-Instruct's Overall score is not clearly above the base model's (e.g. less than ~3 points above when both are evaluated with the same VLMEvalKit settings), or if SFT on GeoTrust-train or TR-GeoMM with matched data volume reaches a comparable Overall or WeMath score.","caption":"Table 5: Main SFT Results. The table details the macro-average scores and specific sub-category breakdowns. The highest values within each parameter-scale comparison group are highlighted with a light blue background and bold text. Asterisks (*) indicate models evaluated via API calls using identical evaluation settings. The dagger (†) denotes a specific anomalous result observed for the 8B baseline on MathVerse Vision-only, with a detailed discussion deferred to Appendix E.3. The double dagger (‡) indicates selected open source baseline synthesis methods for fair comparison with details reported in Appendix E.2. Note that our reported GeoSym scores represent the peak-performing training epoch; an ablation on epoch saturation is provided in Appendix E.4.","image":"img/2605.16371.png","kind":"table","image_w":1247,"image_h":1311,"pwc":true}
|
site/p/2605.18421.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.18421","title":"EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective","authors":"Yuyao Wang; Zhongjian Zhang; Mo Chi; Kaichi Yu; Yuhan Li; Miao Peng; Bing Tong; Chen Zhang; Yan Zhou; Jia Li","openreview_id":"3adOYrlRz9","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM agents are stateless, so memory modules (retrieval stores, long-term memory systems, procedural/skill memories) are added to keep information beyond the context window. These methods are usually evaluated on different tasks and protocols, so it is unclear when memory actually helps compared with simply giving a strong model the raw context. EvoMemBench evaluates 15 memory methods under one protocol along two axes: in-episode vs. cross-episode, and knowledge-oriented vs. execution-oriented memory.","method":["The in-episode knowledge setting (InEp-Know) uses the Accurate Retrieval subsets (EventQA, LongMemEval S*, RULER qa1, RULER qa2; 1000 samples in total) and the Selective Forgetting subsets (FactConsolidation multi-hop and single-hop; 200 samples) of MemoryAgentBench. Information arrives as progressive blocks in an incremental multi-turn format; for retention the agent must answer from earlier blocks, and for revision (FactConsolidation) later facts contradict earlier ones and the answer must reflect the update. The metric is answer accuracy.","Every memory method wraps DeepSeek-V3.2 with two interfaces: update (called after each information block, memory reset at the start of each episode) and utilize (returns retrieved entries or a working context that is put into the prompt). Retrieval top-k is 10 for knowledge tasks, and the default embedding backend is DashScope text-embedding-v4 (BM25 is lexical). The memory-free baseline is DeepSeek-V3.2 given the raw context directly. Memory methods compared: BM25, Qwen3-Emb-4B, GraphRAG, MemAgent, MemoBrain, Mem0, A-MEM, MemOS, MemoryOS."],"claim":"With the same DeepSeek-V3.2 backbone on InEp-Know, BM25 memory improves accuracy on all four retention subsets over the no-memory baseline (e.g. LongMemEval 50.33 vs. 32.33, RULER qa2 55.0 vs. 42.0), yet on FactConsolidation single-hop revision every one of the 9 memory methods scores below the no-memory baseline (best 54.0 vs. 65.0).","source":"Table 3, page 6","metric":"Answer accuracy (%), higher is better; single run, no error bars or seeds reported","comparison":"DeepSeek-V3.2 + memory module (BM25 and 8 other memory methods) vs. DeepSeek-V3.2 with raw context and no memory, on retention (EventQA, LME S*, RULER qa1/qa2) vs. revision (FactConsolidation MH/SH) subsets","falsification":"Re-running with DeepSeek-V3.2, BM25 memory does not beat the no-memory baseline by a clear margin on LongMemEval S* and RULER qa2 (e.g. gain under ~5 points), or some memory method matches or exceeds the no-memory baseline on FactConsolidation single-hop.","caption":"Table 3: Answer accuracy (%) and ranking comparison on InEp-Know. Lower ranks indicate better performance. All memory-augmented methods use DeepSeek-V3.2 as the backbone.","image":"img/2605.18421.png","kind":"table","image_w":1144,"image_h":588,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.18421","title":"EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective","authors":"Yuyao Wang; Zhongjian Zhang; Mo Chi; Kaichi Yu; Yuhan Li; Miao Peng; Bing Tong; Chen Zhang; Yan Zhou; Jia Li","openreview_id":"3adOYrlRz9","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM agents are stateless, so memory modules (retrieval stores, long-term memory systems, procedural/skill memories) are added to keep information beyond the context window. These methods are usually evaluated on different tasks and protocols, so it is unclear when memory actually helps compared with simply giving a strong model the raw context. EvoMemBench evaluates 15 memory methods under one protocol along two axes: in-episode vs. cross-episode, and knowledge-oriented vs. execution-oriented memory.","method":["The in-episode knowledge setting (InEp-Know) uses the Accurate Retrieval subsets (EventQA, LongMemEval S*, RULER qa1, RULER qa2; 1000 samples in total) and the Selective Forgetting subsets (FactConsolidation multi-hop and single-hop; 200 samples) of MemoryAgentBench. Information arrives as progressive blocks in an incremental multi-turn format; for retention the agent must answer from earlier blocks, and for revision (FactConsolidation) later facts contradict earlier ones and the answer must reflect the update. The metric is answer accuracy.","Every memory method wraps DeepSeek-V3.2 with two interfaces: update (called after each information block, memory reset at the start of each episode) and utilize (returns retrieved entries or a working context that is put into the prompt). Retrieval top-k is 10 for knowledge tasks, and the default embedding backend is DashScope text-embedding-v4 (BM25 is lexical). The memory-free baseline is DeepSeek-V3.2 given the raw context directly. Memory methods compared: BM25, Qwen3-Emb-4B, GraphRAG, MemAgent, MemoBrain, Mem0, A-MEM, MemOS, MemoryOS."],"claim":"With the same DeepSeek-V3.2 backbone on InEp-Know, BM25 memory improves accuracy on all four retention subsets over the no-memory baseline (e.g. LongMemEval 50.33 vs. 32.33, RULER qa2 55.0 vs. 42.0), yet on FactConsolidation single-hop revision every one of the 9 memory methods scores below the no-memory baseline (best 54.0 vs. 65.0).","source":"Table 3, page 6","metric":"Answer accuracy (%), higher is better; single run, no error bars or seeds reported","comparison":"DeepSeek-V3.2 + memory module (BM25 and 8 other memory methods) vs. DeepSeek-V3.2 with raw context and no memory, on retention (EventQA, LME S*, RULER qa1/qa2) vs. revision (FactConsolidation MH/SH) subsets","falsification":"Re-running with DeepSeek-V3.2, BM25 memory does not beat the no-memory baseline by a clear margin on LongMemEval S* and RULER qa2 (e.g. gain under ~5 points), or some memory method matches or exceeds the no-memory baseline on FactConsolidation single-hop.","caption":"Table 3: Answer accuracy (%) and ranking comparison on InEp-Know. Lower ranks indicate better performance. All memory-augmented methods use DeepSeek-V3.2 as the backbone.","image":"img/2605.18421.png","kind":"table","image_w":1144,"image_h":588,"pwc":true}
|
site/p/2605.18859.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.18859","title":"TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing","authors":"Pei Yang; Wanyi Chen; Tongyun Yang; Pengbin Feng; Jiarong Xing; Wentao Guo; Yuhang Yao; Yuhang Han; Hanchen Li; Xu Wang; Jie Xiao; Anjie Yang; Lynn Ai; Eric Yang; TIANYU SHI","openreview_id":"tku19umKQq","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM routers send each request to the cheapest model that can handle it. In agentic applications such as coding agents, one user task triggers many LLM calls, so routing each call separately could save much more than routing a whole query. Existing router benchmarks (RouterBench, LLMRouterBench, RouterArena) only score one-shot prompts. They never show the router the intermediate prefix of an agent trajectory (history, tool outputs, shell logs, partial edits), and they never check whether a cheaper model at step i breaks the task later. TwinRouterBench provides step-level routing labels with an offline scorer, plus a live SWE-bench track that measures real task resolution and real API spend.","method":["Static track: 970 router-visible prefixes from 520 successful strong-model trajectories (SWE-bench, BFCL, mtRAG, QMSum, PinchBench). Each prefix is labelled with the cheapest of four tiers (low, mid, mid_high, high; 11-model pool) found by a greedy sequential downgrade search. A step's downgrade is accepted if, with all earlier steps locked at their verified tiers and later steps at the strong tier, at least one of three pool models in that tier still lets the full trajectory pass without needing more LLM-call steps. The trained router ('UncommonRoute (trained)') encodes the latest user request with a frozen BAAI/bge-small-en-v1.5 embedding, concatenates routing metadata (message count, tool-call presence, tool-message count, request length, code/question indicators), and fits an L2-regularized multinomial logistic regression on a training split of the static labels (C chosen by 5-fold CV, plus a separate calibration split).","Dynamic track: 100 locked SWE-bench Verified instances (18 Astropy, 82 Django), disjoint from the static SWE split and not a random sample. They are run end to end with mini-swe-agent v2.2.8. At every LLM call the policy picks a concrete model from the locked pool. Scoring uses the official SWE-bench resolution predicate and realized OpenRouter API cost (input, cache_read, cache_write and output buckets). The leaderboard bill adds a flat $0.60 penalty per unresolved instance. Policies compared: unrouted Claude Opus 4.6, the rule-based UncommonRoute, an SR-KNN semantic router, and the trained UncommonRoute."],"claim":"On the locked 100-instance SWE-bench Verified list with mini-swe-agent, the logistic router trained on the static-track labels resolves 75/100 instances for $25.66 total realized API cost, versus 74/100 for $54.73 with unrouted Claude Opus 4.6 (53.1% lower cost). Its leaderboard bill is $40.66 vs. $70.33.","source":"Table 3, page 11","metric":"Resolved instances out of 100 (official SWE-bench predicate, higher is better) and total realized API cost in USD (lower is better). Leaderboard bill = API cost + $0.60 per unresolved instance. Single run per policy, no error bars or seeds reported.","comparison":"UncommonRoute (trained on static labels) vs. Opus 4.6 with no routing (also vs. SR-KNN router and rule-based UncommonRoute) on the same 100 instances and harness","falsification":"Re-running both policies on the same 100 instances, the trained router either resolves clearly fewer instances than unrouted Opus 4.6 (e.g. 5+ fewer) or its realized API cost is not substantially lower (e.g. less than ~30% saving), so that its leaderboard bill is not clearly below Opus's.","caption":"Table 3. Dynamic-track results on the locked 100-case leaderboard list (18 Astropy and 82 Django), not a random sample of SWE-bench Verified. API cost is the realized routed API spend; Penalty cost is $0.60× the number of unresolved instances, where γ=0.60 is the per-case price of a hypothetical perfect solver (§5.2); Leaderboard bill (sort key, lower is better) is API cost + Penalty cost. The penalty applies uniformly to every policy. The unrouted Opus 4.6 baseline is separated below the routing policies, mirroring the cost-envelope row in Table 2.","image":"img/2605.18859.png","kind":"table","image_w":1333,"image_h":394,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.18859","title":"TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing","authors":"Pei Yang; Wanyi Chen; Tongyun Yang; Pengbin Feng; Jiarong Xing; Wentao Guo; Yuhang Yao; Yuhang Han; Hanchen Li; Xu Wang; Jie Xiao; Anjie Yang; Lynn Ai; Eric Yang; TIANYU SHI","openreview_id":"tku19umKQq","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM routers send each request to the cheapest model that can handle it. In agentic applications such as coding agents, one user task triggers many LLM calls, so routing each call separately could save much more than routing a whole query. Existing router benchmarks (RouterBench, LLMRouterBench, RouterArena) only score one-shot prompts. They never show the router the intermediate prefix of an agent trajectory (history, tool outputs, shell logs, partial edits), and they never check whether a cheaper model at step i breaks the task later. TwinRouterBench provides step-level routing labels with an offline scorer, plus a live SWE-bench track that measures real task resolution and real API spend.","method":["Static track: 970 router-visible prefixes from 520 successful strong-model trajectories (SWE-bench, BFCL, mtRAG, QMSum, PinchBench). Each prefix is labelled with the cheapest of four tiers (low, mid, mid_high, high; 11-model pool) found by a greedy sequential downgrade search. A step's downgrade is accepted if, with all earlier steps locked at their verified tiers and later steps at the strong tier, at least one of three pool models in that tier still lets the full trajectory pass without needing more LLM-call steps. The trained router ('UncommonRoute (trained)') encodes the latest user request with a frozen BAAI/bge-small-en-v1.5 embedding, concatenates routing metadata (message count, tool-call presence, tool-message count, request length, code/question indicators), and fits an L2-regularized multinomial logistic regression on a training split of the static labels (C chosen by 5-fold CV, plus a separate calibration split).","Dynamic track: 100 locked SWE-bench Verified instances (18 Astropy, 82 Django), disjoint from the static SWE split and not a random sample. They are run end to end with mini-swe-agent v2.2.8. At every LLM call the policy picks a concrete model from the locked pool. Scoring uses the official SWE-bench resolution predicate and realized OpenRouter API cost (input, cache_read, cache_write and output buckets). The leaderboard bill adds a flat $0.60 penalty per unresolved instance. Policies compared: unrouted Claude Opus 4.6, the rule-based UncommonRoute, an SR-KNN semantic router, and the trained UncommonRoute."],"claim":"On the locked 100-instance SWE-bench Verified list with mini-swe-agent, the logistic router trained on the static-track labels resolves 75/100 instances for $25.66 total realized API cost, versus 74/100 for $54.73 with unrouted Claude Opus 4.6 (53.1% lower cost). Its leaderboard bill is $40.66 vs. $70.33.","source":"Table 3, page 11","metric":"Resolved instances out of 100 (official SWE-bench predicate, higher is better) and total realized API cost in USD (lower is better). Leaderboard bill = API cost + $0.60 per unresolved instance. Single run per policy, no error bars or seeds reported.","comparison":"UncommonRoute (trained on static labels) vs. Opus 4.6 with no routing (also vs. SR-KNN router and rule-based UncommonRoute) on the same 100 instances and harness","falsification":"Re-running both policies on the same 100 instances, the trained router either resolves clearly fewer instances than unrouted Opus 4.6 (e.g. 5+ fewer) or its realized API cost is not substantially lower (e.g. less than ~30% saving), so that its leaderboard bill is not clearly below Opus's.","caption":"Table 3. Dynamic-track results on the locked 100-case leaderboard list (18 Astropy and 82 Django), not a random sample of SWE-bench Verified. API cost is the realized routed API spend; Penalty cost is $0.60× the number of unresolved instances, where γ=0.60 is the per-case price of a hypothetical perfect solver (§5.2); Leaderboard bill (sort key, lower is better) is API cost + Penalty cost. The penalty applies uniformly to every policy. The unrouted Opus 4.6 baseline is separated below the routing policies, mirroring the cost-envelope row in Table 2.","image":"img/2605.18859.png","kind":"table","image_w":1333,"image_h":394,"pwc":true}
|
site/p/2605.19127.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.19127","title":"POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents","authors":"Qiaoyuan ZHENG; yiqu yang; Qi Gao; Imanol Schlag","openreview_id":"Y4sDfO3t6Q","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM agents increasingly hold users' private data (medical records, finances, CVs) and talk to third-party systems on the user's behalf. The user gives a privacy policy saying what may and may not be shared, and the agent has to share enough to complete the task while withholding protected attributes, even when the third party tries to extract them. Existing privacy benchmarks test one policy type or one attack at a time and often score privacy alone, so it is unclear which models actually hold up under adversarial probing and whether they do so by refusing to be useful.","method":["POLAR-Bench builds 7,852 synthetic samples over 10 domains. Each sample has a source document rendered by an LLM from a typed JSON profile whose attributes are labelled task-relevant, protected, or irrelevant; a privacy policy at one of five levels (P1 explicit field rules to P5 conflicting objectives); a task instruction that carries no privacy hints; and an attacker prompt sequence for one of five strategies (S1 direct single-turn, S2 yes/no narrowing, S3 role confusion, S4 prompt injection, S5 3-6 turn progressive elicitation). The trusted model T gets document, policy and task, and converses with a fixed third-party model E (Llama-3.3-70B-Instruct) that asks the attacker prompts.","Scoring is deterministic: attribute values revealed in the transcript are found by normalization and regex matching. Privacy = 1 - fraction of protected attributes revealed; Utility = fraction of task-relevant attributes shared; Overall = 0.5*Privacy + 0.5*Utility. 22 trusted models are evaluated over the full 5x5 policy x attack grid. Table 10 reports each model's privacy and utility, averaged over the five attack strategies (the mean values match Figure 2)."],"claim":"With Llama-3.3-70B-Instruct as the attacker, the strongest models keep protected attributes private almost perfectly (GPT-5.4 privacy 99.86, GLM-5.1 99.25, Gemma-4-31B 99.65), while small open-weight models leak far more: SmolLM3-3B 46.32 and Ministral-3-3B 48.28 privacy (i.e. leak over half), while keeping substantial utility (78.10 and 82.30, somewhat below GPT-5.4's 85.31).","source":"Table 10, page 45 (same privacy values shown in Figure 2, page 6)","metric":"Privacy score (0-100, % of protected attributes not revealed, higher is better) and Utility score (% of task-relevant attributes shared, higher is better), averaged over attack strategies S1-S5. Variance columns are variance across the five attack-strategy means, not across seeds; no seed-level error bars are reported.","comparison":"Frontier/large models (GPT-5.4, GLM-5.1, Gemma-4-31B, DeepSeek-V3.1) vs. small open-weight models (SmolLM3-3B, Ministral-3-3B, Apertus-8B, Gemma-4-E2B) as the trusted agent on the same benchmark samples and the same attacker.","falsification":"On the POLAR-Bench samples (or a faithful regeneration), SmolLM3-3B or Ministral-3-3B reach privacy above about 75, or GPT-5.4 falls below about 95, so that the gap between frontier and small 3B models shrinks to under about 20 points; or the small models' high leakage is accompanied by a utility collapse (i.e. they leak only because they share everything in degenerate transcripts).","caption":"Table 10: Model-wise variation across attack strategies. The table reports each model's mean and variance of Privacy and Utility scores across attack strategies. Red bold values indicate the highest value in each metric, while blue underlined values indicate the lowest value.","image":"img/2605.19127.png","kind":"table","image_w":1133,"image_h":736,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.19127","title":"POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents","authors":"Qiaoyuan ZHENG; yiqu yang; Qi Gao; Imanol Schlag","openreview_id":"Y4sDfO3t6Q","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLM agents increasingly hold users' private data (medical records, finances, CVs) and talk to third-party systems on the user's behalf. The user gives a privacy policy saying what may and may not be shared, and the agent has to share enough to complete the task while withholding protected attributes, even when the third party tries to extract them. Existing privacy benchmarks test one policy type or one attack at a time and often score privacy alone, so it is unclear which models actually hold up under adversarial probing and whether they do so by refusing to be useful.","method":["POLAR-Bench builds 7,852 synthetic samples over 10 domains. Each sample has a source document rendered by an LLM from a typed JSON profile whose attributes are labelled task-relevant, protected, or irrelevant; a privacy policy at one of five levels (P1 explicit field rules to P5 conflicting objectives); a task instruction that carries no privacy hints; and an attacker prompt sequence for one of five strategies (S1 direct single-turn, S2 yes/no narrowing, S3 role confusion, S4 prompt injection, S5 3-6 turn progressive elicitation). The trusted model T gets document, policy and task, and converses with a fixed third-party model E (Llama-3.3-70B-Instruct) that asks the attacker prompts.","Scoring is deterministic: attribute values revealed in the transcript are found by normalization and regex matching. Privacy = 1 - fraction of protected attributes revealed; Utility = fraction of task-relevant attributes shared; Overall = 0.5*Privacy + 0.5*Utility. 22 trusted models are evaluated over the full 5x5 policy x attack grid. Table 10 reports each model's privacy and utility, averaged over the five attack strategies (the mean values match Figure 2)."],"claim":"With Llama-3.3-70B-Instruct as the attacker, the strongest models keep protected attributes private almost perfectly (GPT-5.4 privacy 99.86, GLM-5.1 99.25, Gemma-4-31B 99.65), while small open-weight models leak far more: SmolLM3-3B 46.32 and Ministral-3-3B 48.28 privacy (i.e. leak over half), while keeping substantial utility (78.10 and 82.30, somewhat below GPT-5.4's 85.31).","source":"Table 10, page 45 (same privacy values shown in Figure 2, page 6)","metric":"Privacy score (0-100, % of protected attributes not revealed, higher is better) and Utility score (% of task-relevant attributes shared, higher is better), averaged over attack strategies S1-S5. Variance columns are variance across the five attack-strategy means, not across seeds; no seed-level error bars are reported.","comparison":"Frontier/large models (GPT-5.4, GLM-5.1, Gemma-4-31B, DeepSeek-V3.1) vs. small open-weight models (SmolLM3-3B, Ministral-3-3B, Apertus-8B, Gemma-4-E2B) as the trusted agent on the same benchmark samples and the same attacker.","falsification":"On the POLAR-Bench samples (or a faithful regeneration), SmolLM3-3B or Ministral-3-3B reach privacy above about 75, or GPT-5.4 falls below about 95, so that the gap between frontier and small 3B models shrinks to under about 20 points; or the small models' high leakage is accompanied by a utility collapse (i.e. they leak only because they share everything in degenerate transcripts).","caption":"Table 10: Model-wise variation across attack strategies. The table reports each model's mean and variance of Privacy and Utility scores across attack strategies. Red bold values indicate the highest value in each metric, while blue underlined values indicate the lowest value.","image":"img/2605.19127.png","kind":"table","image_w":1133,"image_h":736,"pwc":true}
|
site/p/2605.19382.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.19382","title":"PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning","authors":"Qiran Zhang; Yuheng Wang; Runde Yang; Lin Wu; Jingru Fan; Shu Yao; Jie Zhang; Tianle Zhou; Huatao Li; Ruijie Shi; Yihan Li; Chen Qian","openreview_id":"XbNjlZYLho","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLMs can make educational and explanatory videos by writing code for a programmatic animation engine such as Manim, which gives exact geometry and consistent timing that pixel-level diffusion video models lack. Existing evaluations of such code mostly check whether it runs, or ask a VLM judge for an overall rating of the video. Neither reliably catches layout failures that show up only in some frames, such as overlapping elements, objects off-screen, or text spilling out of its box. Whether 'the code runs' implies 'the animation is laid out correctly' is therefore an open question.","method":["PRISM contains 10,372 instruction-code pairs (5,199 English, 5,173 Chinese). Each instruction is an educational outline fragment plus environment constraints, and the model must write a Manim CE v0.19.0 script. Seven LLMs (GPT-5.4, Gemini 3.1 Pro Preview, Kimi K2 Thinking, Claude Sonnet 4.6, Qwen3.5-397B-A17B, GLM-5, DeepSeek-V3.2) are evaluated on the same randomly sampled 20% subset, with English and Chinese scored separately. Sampling uses temperature 0.7, a single first attempt per prompt, and thinking mode where available.","Exec. is the fraction of prompts whose first-attempt script renders without error. Spatial is the fraction of prompts that both render and pass a deterministic audit on every stable keyframe. The audit hooks into Manim's renderer, takes snapshots synchronized with animation actions, extracts each object's points, bounding extent and scene-graph hierarchy, and checks for three failure types: layout overlap, out-of-bounds placement, and layout leakage. A false-positive suppression module exempts intentional overlaps such as highlighting and grid adjacency. Failed renders count as Spatial failures."],"claim":"On the PRISM 20% evaluation subset, every one of the 7 models has a spatial pass rate far below its execution success rate, with a mean drop of 41.24 pp (95% bootstrap CI [40.41, 42.11]). GPT-5.4 has the highest Exec. but the largest gap (English 0.945 vs. 0.308; Chinese 0.902 vs. 0.177). Gemini 3.1 Pro Preview has the best Spatial score (0.572 English, 0.557 Chinese) despite a lower Exec. (0.778, 0.752).","source":"Table 1, page 7 (gap statistic in Section 4.2 text, page 7; Figure 6, page 8)","metric":"Exec. = execution success rate and Spatial = spatial pass rate, both proportions of prompts, higher is better; the gap is Exec. minus Spatial in percentage points. The only error estimate is a 95% bootstrap CI on the pooled mean gap. Per-model numbers come from a single sample per prompt, with no seeds or error bars.","comparison":"Execution success vs. spatial pass rate for each of 7 LLMs, plus GPT-5.4 vs. Gemini 3.1 Pro Preview on the same prompts, in English and Chinese","falsification":"Running the released audit on the same subset gives a mean Exec.-to-Spatial gap well below about 41 pp (e.g. under 30 pp), some model's Spatial is close to its Exec. (gap under about 15 pp), or Gemini 3.1 Pro Preview's Spatial is not clearly higher than GPT-5.4's (English 0.572 vs. 0.308).","caption":"Table 1: Main benchmark results. Exec. (execution success rate) and Spatial (spatial pass rate) are reported as proportions; PADVC and TD report the centered scores PADVCc and TDc; Time is mean render time in minutes; Overlap/Leak/OOB. are the three constituent spatial-failure rates, reported as proportions. Best results are in bold, runners-up are underlined.","image":"img/2605.19382.png","kind":"table","image_w":1122,"image_h":683,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.19382","title":"PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning","authors":"Qiran Zhang; Yuheng Wang; Runde Yang; Lin Wu; Jingru Fan; Shu Yao; Jie Zhang; Tianle Zhou; Huatao Li; Ruijie Shi; Yihan Li; Chen Qian","openreview_id":"XbNjlZYLho","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLMs can make educational and explanatory videos by writing code for a programmatic animation engine such as Manim, which gives exact geometry and consistent timing that pixel-level diffusion video models lack. Existing evaluations of such code mostly check whether it runs, or ask a VLM judge for an overall rating of the video. Neither reliably catches layout failures that show up only in some frames, such as overlapping elements, objects off-screen, or text spilling out of its box. Whether 'the code runs' implies 'the animation is laid out correctly' is therefore an open question.","method":["PRISM contains 10,372 instruction-code pairs (5,199 English, 5,173 Chinese). Each instruction is an educational outline fragment plus environment constraints, and the model must write a Manim CE v0.19.0 script. Seven LLMs (GPT-5.4, Gemini 3.1 Pro Preview, Kimi K2 Thinking, Claude Sonnet 4.6, Qwen3.5-397B-A17B, GLM-5, DeepSeek-V3.2) are evaluated on the same randomly sampled 20% subset, with English and Chinese scored separately. Sampling uses temperature 0.7, a single first attempt per prompt, and thinking mode where available.","Exec. is the fraction of prompts whose first-attempt script renders without error. Spatial is the fraction of prompts that both render and pass a deterministic audit on every stable keyframe. The audit hooks into Manim's renderer, takes snapshots synchronized with animation actions, extracts each object's points, bounding extent and scene-graph hierarchy, and checks for three failure types: layout overlap, out-of-bounds placement, and layout leakage. A false-positive suppression module exempts intentional overlaps such as highlighting and grid adjacency. Failed renders count as Spatial failures."],"claim":"On the PRISM 20% evaluation subset, every one of the 7 models has a spatial pass rate far below its execution success rate, with a mean drop of 41.24 pp (95% bootstrap CI [40.41, 42.11]). GPT-5.4 has the highest Exec. but the largest gap (English 0.945 vs. 0.308; Chinese 0.902 vs. 0.177). Gemini 3.1 Pro Preview has the best Spatial score (0.572 English, 0.557 Chinese) despite a lower Exec. (0.778, 0.752).","source":"Table 1, page 7 (gap statistic in Section 4.2 text, page 7; Figure 6, page 8)","metric":"Exec. = execution success rate and Spatial = spatial pass rate, both proportions of prompts, higher is better; the gap is Exec. minus Spatial in percentage points. The only error estimate is a 95% bootstrap CI on the pooled mean gap. Per-model numbers come from a single sample per prompt, with no seeds or error bars.","comparison":"Execution success vs. spatial pass rate for each of 7 LLMs, plus GPT-5.4 vs. Gemini 3.1 Pro Preview on the same prompts, in English and Chinese","falsification":"Running the released audit on the same subset gives a mean Exec.-to-Spatial gap well below about 41 pp (e.g. under 30 pp), some model's Spatial is close to its Exec. (gap under about 15 pp), or Gemini 3.1 Pro Preview's Spatial is not clearly higher than GPT-5.4's (English 0.572 vs. 0.308).","caption":"Table 1: Main benchmark results. Exec. (execution success rate) and Spatial (spatial pass rate) are reported as proportions; PADVC and TD report the centered scores PADVCc and TDc; Time is mean render time in minutes; Overlap/Leak/OOB. are the three constituent spatial-failure rates, reported as proportions. Best results are in bold, runners-up are underlined.","image":"img/2605.19382.png","kind":"table","image_w":1122,"image_h":683,"pwc":true}
|
site/p/2605.22469.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.22469","title":"MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation","authors":"Patryk Bartkowiak; Lennart Petersen; Bartosz Kotrys; Dominik L Michels; Soeren Pirk; Wojtek Palubicki","openreview_id":"s74oBrTreh","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Papers on subject-driven text-to-image personalization (DreamBooth, IP-Adapter, etc.) report Concept Preservation (CP, does the generated subject look like the reference?) with CLIP-I or DINO image-image cosine, and Prompt Following (PF) with CLIP-T. DreamBench++ collected human ratings and showed that these metrics agree with humans only at about Krippendorff α ≈ 0.3, roughly half the inter-annotator agreement, while GPT-4o judges do much better but are API-bound and opaque. The authors argue that the gap comes from global pooling, which mixes the subject with the background that humans ignore when judging identity, rather than from the encoder.","method":["MaSC encodes the reference image and the generated image once each with a frozen SigLIP2 SO400M-NaFlex vision tower (N = 1024 patch tokens for square inputs). A foreground mask of the subject, obtained on DreamBench++ with SAM3 prompted by the subject's canonical name, is downsampled to the patch grid (bilinear, threshold 0.5). The CP score (masked-maxcos) is: for each reference patch inside the reference foreground mask, take the maximum cosine similarity against all patches of the generated image (no mask on the output side), then average over the foreground reference patches. Pairs whose output mask covers under 5% of the image are dropped, and subjects whose reference mask covers under 5% are dropped entirely; this filter is applied to all metrics.","Scores are compared with pooled DreamBench++ human CP ratings using Krippendorff α (and Spearman ρ) on a 7,135-key subset that is the intersection on which all metrics were scored. Baselines include the same SigLIP2 checkpoint with its global attention pool and image-image cosine (same-backbone ablation), DreamSim, DINOv3, DIFT-SDXL, DINO-I, AM-RADIO C-RADIOv4-SO400M, CLIP-I, and the GPT-4V and GPT-4o judge scores shipped with DreamBench++."],"claim":"On the 7,135-key DreamBench++ concept-preservation subset, MaSC's masked-maxcos score reaches Krippendorff α = +0.471 with pooled human ratings, above the same SigLIP2 SO400M-NaFlex checkpoint with global pooling (+0.369) and every other non-LLM metric tested, the best of which is DreamSim (+0.421).","source":"Table 1, page 6","metric":"Krippendorff α (higher is better) between metric scores and pooled human CP ratings; Spearman ρ also reported. No error bars, CIs or bootstrap are reported (the metric is deterministic).","comparison":"MaSC (SigLIP2 patch tokens + masked-maxcos) vs. SigLIP2 global pool on the same checkpoint, DreamSim, DINOv3, DIFT-SDXL, DINO-I, AM-RADIO, CLIP-I; GPT-4V and GPT-4o judge scores as reference points","falsification":"Recomputing on the same DreamBench++ subset, MaSC's α does not exceed the same-backbone SigLIP2 global-pool α by a clear margin (e.g. the 0.102 gap shrinks to near zero), or DreamSim (or another non-LLM baseline) matches or exceeds MaSC's α.","caption":"Table 1: Concept Preservation on DreamBench++. Krippendorff α (Kd_o) and Spearman ρ against pooled human ratings on the 7,135-key apples-to-apples subset. Bold: ours — the only non-LLM, non-human metric to beat GPT-4V; sits Δα = −0.028 behind GPT-4o and reaches 72% of the honest human inter-rater ceiling.","image":"img/2605.22469.png","kind":"table","image_w":1152,"image_h":488,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.22469","title":"MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation","authors":"Patryk Bartkowiak; Lennart Petersen; Bartosz Kotrys; Dominik L Michels; Soeren Pirk; Wojtek Palubicki","openreview_id":"s74oBrTreh","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Papers on subject-driven text-to-image personalization (DreamBooth, IP-Adapter, etc.) report Concept Preservation (CP, does the generated subject look like the reference?) with CLIP-I or DINO image-image cosine, and Prompt Following (PF) with CLIP-T. DreamBench++ collected human ratings and showed that these metrics agree with humans only at about Krippendorff α ≈ 0.3, roughly half the inter-annotator agreement, while GPT-4o judges do much better but are API-bound and opaque. The authors argue that the gap comes from global pooling, which mixes the subject with the background that humans ignore when judging identity, rather than from the encoder.","method":["MaSC encodes the reference image and the generated image once each with a frozen SigLIP2 SO400M-NaFlex vision tower (N = 1024 patch tokens for square inputs). A foreground mask of the subject, obtained on DreamBench++ with SAM3 prompted by the subject's canonical name, is downsampled to the patch grid (bilinear, threshold 0.5). The CP score (masked-maxcos) is: for each reference patch inside the reference foreground mask, take the maximum cosine similarity against all patches of the generated image (no mask on the output side), then average over the foreground reference patches. Pairs whose output mask covers under 5% of the image are dropped, and subjects whose reference mask covers under 5% are dropped entirely; this filter is applied to all metrics.","Scores are compared with pooled DreamBench++ human CP ratings using Krippendorff α (and Spearman ρ) on a 7,135-key subset that is the intersection on which all metrics were scored. Baselines include the same SigLIP2 checkpoint with its global attention pool and image-image cosine (same-backbone ablation), DreamSim, DINOv3, DIFT-SDXL, DINO-I, AM-RADIO C-RADIOv4-SO400M, CLIP-I, and the GPT-4V and GPT-4o judge scores shipped with DreamBench++."],"claim":"On the 7,135-key DreamBench++ concept-preservation subset, MaSC's masked-maxcos score reaches Krippendorff α = +0.471 with pooled human ratings, above the same SigLIP2 SO400M-NaFlex checkpoint with global pooling (+0.369) and every other non-LLM metric tested, the best of which is DreamSim (+0.421).","source":"Table 1, page 6","metric":"Krippendorff α (higher is better) between metric scores and pooled human CP ratings; Spearman ρ also reported. No error bars, CIs or bootstrap are reported (the metric is deterministic).","comparison":"MaSC (SigLIP2 patch tokens + masked-maxcos) vs. SigLIP2 global pool on the same checkpoint, DreamSim, DINOv3, DIFT-SDXL, DINO-I, AM-RADIO, CLIP-I; GPT-4V and GPT-4o judge scores as reference points","falsification":"Recomputing on the same DreamBench++ subset, MaSC's α does not exceed the same-backbone SigLIP2 global-pool α by a clear margin (e.g. the 0.102 gap shrinks to near zero), or DreamSim (or another non-LLM baseline) matches or exceeds MaSC's α.","caption":"Table 1: Concept Preservation on DreamBench++. Krippendorff α (Kd_o) and Spearman ρ against pooled human ratings on the 7,135-key apples-to-apples subset. Bold: ours — the only non-LLM, non-human metric to beat GPT-4V; sits Δα = −0.028 behind GPT-4o and reaches 72% of the honest human inter-rater ceiling.","image":"img/2605.22469.png","kind":"table","image_w":1152,"image_h":488,"pwc":true}
|
site/p/2605.22907.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.22907","title":"VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding","authors":"He Haichen; Jiayi Zhou; Sifeng SHANG; Yihan Hu; Yuanhan Zhang; Kaiyang Zhou","openreview_id":"HIUX5RHWKs","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Long-video benchmarks have grown in raw video duration, but most of their questions can still be answered from a short, isolated clip. VideoOdyssey adds a measure called continuous certificate length (CCL): the stretch of video a human must watch without a break to answer a question for certain. The benchmark has 100 YouTube videos (average 109 minutes, all over 60 minutes) and two subsets. VideoOdyssey-V has 1,618 visual multiple-choice questions with an average CCL of 16 minutes. VideoOdyssey-AV has 1,062 audio-visual multiple-choice questions with an average CCL of 12.8 minutes. The paper's diagnostic argument is that models fail for two reasons: they cannot find the relevant evidence (retrieval), and they cannot reason over it once found (comprehension).","method":["Every question has four options and a human-annotated CCL window: the shortest continuous interval that contains all the evidence needed to answer it. In the full-video setting, a model sees N frames sampled uniformly and deterministically across the whole video (N = 64 for the AV subset), plus the audio track for omni-modal models. It is prompted to output a single option letter (prompts in Appendix G). The metric is accuracy, overall and split into five CCL bins: [0,0.5), [0.5,3), [3,15), [15,60) and [60,inf) minutes.","In the certificate-window (CW) settings, the model sees only the ground-truth CCL window instead of the whole video. 'CW, matched frames' samples the full budget of N frames inside the window, or 1 fps for windows shorter than N seconds, so it also raises frame density. 'CW, matched rate' keeps the full-video sampling interval and samples N_CW = max(1, round(N*L/T)) frames, where L is the window length and T the video length. This isolates the effect of localization from the effect of denser sampling. Table 15 reports all three settings for two V-subset models and two AV-subset models."],"claim":"On VideoOdyssey-AV with a 64-frame full-video budget and video+audio input, giving Qwen3-Omni-30B only the ground-truth certificate window at the same sampling rate raises overall accuracy from 28.7% to 38.6%, and at the same 64-frame count it reaches 48.6%.","source":"Table 15, page 28 (Appendix F.1)","metric":"Multiple-choice accuracy (%) over all 1,062 VideoOdyssey-AV questions; higher is better; 25% is chance. No error bars or repeated runs are reported (the checklist says variance is negligible under deterministic/low-temperature decoding).","comparison":"Full video (w/o CW) vs. CW at matched sampling rate vs. CW at matched frame count, for the same model, frame budget and prompt","falsification":"With Qwen3-Omni-30B on VideoOdyssey-AV at N = 64, matched-rate CW accuracy is not clearly above full-video accuracy (a gain of roughly 3 pp or less, instead of the reported +9.9 pp), or full-video accuracy differs substantially from 28.7%.","caption":"Table 15: Full-video, matched-frame CW, and matched-rate CW accuracy (%). CCL intervals are in minutes; N denotes the full-video frame budget.","image":"img/2605.22907.png","kind":"table","image_w":1144,"image_h":658,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.22907","title":"VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding","authors":"He Haichen; Jiayi Zhou; Sifeng SHANG; Yihan Hu; Yuanhan Zhang; Kaiyang Zhou","openreview_id":"HIUX5RHWKs","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Long-video benchmarks have grown in raw video duration, but most of their questions can still be answered from a short, isolated clip. VideoOdyssey adds a measure called continuous certificate length (CCL): the stretch of video a human must watch without a break to answer a question for certain. The benchmark has 100 YouTube videos (average 109 minutes, all over 60 minutes) and two subsets. VideoOdyssey-V has 1,618 visual multiple-choice questions with an average CCL of 16 minutes. VideoOdyssey-AV has 1,062 audio-visual multiple-choice questions with an average CCL of 12.8 minutes. The paper's diagnostic argument is that models fail for two reasons: they cannot find the relevant evidence (retrieval), and they cannot reason over it once found (comprehension).","method":["Every question has four options and a human-annotated CCL window: the shortest continuous interval that contains all the evidence needed to answer it. In the full-video setting, a model sees N frames sampled uniformly and deterministically across the whole video (N = 64 for the AV subset), plus the audio track for omni-modal models. It is prompted to output a single option letter (prompts in Appendix G). The metric is accuracy, overall and split into five CCL bins: [0,0.5), [0.5,3), [3,15), [15,60) and [60,inf) minutes.","In the certificate-window (CW) settings, the model sees only the ground-truth CCL window instead of the whole video. 'CW, matched frames' samples the full budget of N frames inside the window, or 1 fps for windows shorter than N seconds, so it also raises frame density. 'CW, matched rate' keeps the full-video sampling interval and samples N_CW = max(1, round(N*L/T)) frames, where L is the window length and T the video length. This isolates the effect of localization from the effect of denser sampling. Table 15 reports all three settings for two V-subset models and two AV-subset models."],"claim":"On VideoOdyssey-AV with a 64-frame full-video budget and video+audio input, giving Qwen3-Omni-30B only the ground-truth certificate window at the same sampling rate raises overall accuracy from 28.7% to 38.6%, and at the same 64-frame count it reaches 48.6%.","source":"Table 15, page 28 (Appendix F.1)","metric":"Multiple-choice accuracy (%) over all 1,062 VideoOdyssey-AV questions; higher is better; 25% is chance. No error bars or repeated runs are reported (the checklist says variance is negligible under deterministic/low-temperature decoding).","comparison":"Full video (w/o CW) vs. CW at matched sampling rate vs. CW at matched frame count, for the same model, frame budget and prompt","falsification":"With Qwen3-Omni-30B on VideoOdyssey-AV at N = 64, matched-rate CW accuracy is not clearly above full-video accuracy (a gain of roughly 3 pp or less, instead of the reported +9.9 pp), or full-video accuracy differs substantially from 28.7%.","caption":"Table 15: Full-video, matched-frame CW, and matched-rate CW accuracy (%). CCL intervals are in minutes; N denotes the full-video frame budget.","image":"img/2605.22907.png","kind":"table","image_w":1144,"image_h":658,"pwc":true}
|
site/p/2605.28345.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.28345","title":"PICID: A Modular Evaluation Infrastructure for Reproducible PHM Across Tasks and Domains","authors":"Lev Telyatnikov; Raffael Theiler; Leandro Von Krannichfeldt; Olga Fink","openreview_id":"B8dtYVh7hZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Prognostics and Health Management (PHM) uses machine learning to detect faults, classify them (diagnostics) and predict remaining useful life (prognostics) for engineered assets such as engines, batteries, bearings and hydraulic rigs. Reported results in PHM are hard to reproduce or compare, because protocol choices such as splits, preprocessing, label alignment, windowing and metrics are usually implicit and ad hoc; the authors' audit found public code for only 8 of 329 papers from 2022 to 2025. The paper proposes PICID, a library that fixes these protocol choices as executable, leakage-safe components, so that different model families can be compared on identical data contracts across tasks.","method":["PICID loads raw unit trajectories, applies split-safe preprocessing that is fitted on the training partition only, aligns targets to the transformed time grid, windows sequences into supervised samples, and evaluates every model through a shared wrapper and evaluator interface. The benchmark runs 13 models (Linear/Exp/MLP; LSTM, CNN-1D, TiDE; the transformers Timeseries Transformer (TST), Spacetimeformer (STF), Crossformer (CF) and PatchTST (PTST); XGBoost; TabPFN and TabDPT) on 6 prognostics and 6 diagnostics datasets, using the same family instantiated as a classifier or a regressor.","For diagnostics, the datasets are N-CMAPSS multi-source concept classification (NC-D), four HSF15 hydraulic component tasks (accumulator 4-way, cooler 3-way, pump 3-way, valve 4-way), and MZVAV building faults. HSF15 features are min-max scaled, summarized with time-domain and spectral statistics, concatenated and rescaled, and the label is the last label in each window. Gradient-trained models (including LSTM, CNN-1D and all transformers) search over sequence length {1, 10, 50} and learning rate {1e-3, 5e-4, 1e-4}, with batch size 512, up to 200 epochs and early stopping, and are selected on validation F1. Each configuration is run with 5 seeds, and the metric is test macro-F1 (x100)."],"claim":"Under PICID's identical preprocessing, windowing and splits, all four transformer models (TST, STF, CF, PTST) score test F1 <= 65.94 on NC-D and on each of the four HSF15 hydraulic diagnostics tasks, while both LSTM and CNN-1D score >= 87.53 on every one of those five tasks (e.g. HSF15-A: LSTM 94.59 ± 0.97, CNN-1D 94.03 ± 1.98 vs. best transformer STF 40.01 ± 5.75).","source":"Table 2 (bottom block), page 9","metric":"Macro-F1 x100 on the test split (higher is better); ± is the std over 5 seeds","comparison":"Transformer family (TST, STF, CF, PTST) vs. deep sequence models LSTM and CNN-1D, same pipeline and the same hyperparameter search","falsification":"Running PICID's released configurations with 5 seeds, any transformer reaches within a few F1 points of the weaker of LSTM/CNN-1D on NC-D or on any HSF15 task, or LSTM/CNN-1D fall below about 85 F1 on these tasks.","caption":"Table 2: Main evaluation results. Top block: prognostics, measured by MAE in the normalized target space (×100) (↓). Bottom block: diagnostics, measured by F1 score (×100) (↑). Models are grouped by family: simple baselines, deep sequence models, transformers, tabular models, and tabular foundation models. Bold/underline denote the best/second-best result.","image":"img/2605.28345.png","kind":"table","image_w":1144,"image_h":772,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.28345","title":"PICID: A Modular Evaluation Infrastructure for Reproducible PHM Across Tasks and Domains","authors":"Lev Telyatnikov; Raffael Theiler; Leandro Von Krannichfeldt; Olga Fink","openreview_id":"B8dtYVh7hZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Prognostics and Health Management (PHM) uses machine learning to detect faults, classify them (diagnostics) and predict remaining useful life (prognostics) for engineered assets such as engines, batteries, bearings and hydraulic rigs. Reported results in PHM are hard to reproduce or compare, because protocol choices such as splits, preprocessing, label alignment, windowing and metrics are usually implicit and ad hoc; the authors' audit found public code for only 8 of 329 papers from 2022 to 2025. The paper proposes PICID, a library that fixes these protocol choices as executable, leakage-safe components, so that different model families can be compared on identical data contracts across tasks.","method":["PICID loads raw unit trajectories, applies split-safe preprocessing that is fitted on the training partition only, aligns targets to the transformed time grid, windows sequences into supervised samples, and evaluates every model through a shared wrapper and evaluator interface. The benchmark runs 13 models (Linear/Exp/MLP; LSTM, CNN-1D, TiDE; the transformers Timeseries Transformer (TST), Spacetimeformer (STF), Crossformer (CF) and PatchTST (PTST); XGBoost; TabPFN and TabDPT) on 6 prognostics and 6 diagnostics datasets, using the same family instantiated as a classifier or a regressor.","For diagnostics, the datasets are N-CMAPSS multi-source concept classification (NC-D), four HSF15 hydraulic component tasks (accumulator 4-way, cooler 3-way, pump 3-way, valve 4-way), and MZVAV building faults. HSF15 features are min-max scaled, summarized with time-domain and spectral statistics, concatenated and rescaled, and the label is the last label in each window. Gradient-trained models (including LSTM, CNN-1D and all transformers) search over sequence length {1, 10, 50} and learning rate {1e-3, 5e-4, 1e-4}, with batch size 512, up to 200 epochs and early stopping, and are selected on validation F1. Each configuration is run with 5 seeds, and the metric is test macro-F1 (x100)."],"claim":"Under PICID's identical preprocessing, windowing and splits, all four transformer models (TST, STF, CF, PTST) score test F1 <= 65.94 on NC-D and on each of the four HSF15 hydraulic diagnostics tasks, while both LSTM and CNN-1D score >= 87.53 on every one of those five tasks (e.g. HSF15-A: LSTM 94.59 ± 0.97, CNN-1D 94.03 ± 1.98 vs. best transformer STF 40.01 ± 5.75).","source":"Table 2 (bottom block), page 9","metric":"Macro-F1 x100 on the test split (higher is better); ± is the std over 5 seeds","comparison":"Transformer family (TST, STF, CF, PTST) vs. deep sequence models LSTM and CNN-1D, same pipeline and the same hyperparameter search","falsification":"Running PICID's released configurations with 5 seeds, any transformer reaches within a few F1 points of the weaker of LSTM/CNN-1D on NC-D or on any HSF15 task, or LSTM/CNN-1D fall below about 85 F1 on these tasks.","caption":"Table 2: Main evaluation results. Top block: prognostics, measured by MAE in the normalized target space (×100) (↓). Bottom block: diagnostics, measured by F1 score (×100) (↑). Models are grouped by family: simple baselines, deep sequence models, transformers, tabular models, and tabular foundation models. Bold/underline denote the best/second-best result.","image":"img/2605.28345.png","kind":"table","image_w":1144,"image_h":772,"pwc":true}
|
site/p/2605.28683.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.28683","title":"VeriTrip: A Verifiable Benchmark for Travel Planning Agents over Unstructured Web Corpora","authors":"Yuting Xu; Jiayi Tian; Jian Liang; Xin Xiong; Hang Zhang; Mu Xu; Xiao-Yu Zhang","openreview_id":"hFIKn7BCcY","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Travel-planning benchmarks for LLM agents (TravelPlanner, ChinaTravel, TripTailor) give the agent clean structured APIs for flights, hotels and attractions, so they test constraint satisfaction but not whether an agent can find the facts itself in noisy web content. Deep-research benchmarks test information seeking but usually grade free-form reports with an LLM judge, which misses fabricated details. VeriTrip combines the two: agents must search a frozen, noisy multimodal web snapshot to build an itinerary, and every factual cell of the output is checked automatically against a hidden, human-verified knowledge base.","method":["The Multimodal Retrieval Base (MRB) holds 8,210 documents and 4,146 images scraped from Ctrip/Trip.com, 12306 and RedNote for 15 cities in China and the US, with ads and PII removed but conflicting or outdated user content kept. Agents get docSearch (Qwen3-Embedding-8B dense retrieval, top-10 snippets), imgSearch (DINOv2 ViT-S/14), getDocument, and three structured restaurant tools. Each of the 228 queries (78 simple, 76 medium, 74 complex) has text constraints plus a heavily cropped photo of a target POI. Complex queries have more travel days, more travelers and more stated preferences (2-8 vs. 1-4). Models run in a ReAct loop and must output a fixed JSON itinerary schema.","Evaluation is programmatic: Delivery Rate (DR, valid JSON), Factual Reliability (FR, fraction of transport IDs, times, POI and hotel names that match the VKB, exact or fuzzy), Preference Fulfillment Rate (PFR, rule checks for the stated transport mode, room type, must-see attractions, meal preferences and inner-city transport), micro Pass Rate on commonsense constraints (PRmi), and Average Margin over a TSP-optimal route (AM). Eleven MLLMs are evaluated (8 without thinking, 3 with thinking)."],"claim":"On VeriTrip, Preference Fulfillment Rate is lower on complex than on simple queries for all 11 evaluated MLLMs, by 9.7 to 30.0 points (e.g., Claude-4.5-Sonnet 64.10 -> 50.00, GPT-4o 45.20 -> 32.53, o3 55.38 -> 27.57, Qwen3-VL-235B 66.46 -> 36.49), while Factual Reliability is higher on complex queries for all 11 models (by 2.1 to 11.3 points).","source":"Table 4, page 7","metric":"PFR and FR in % (higher is better); single run per model and difficulty level, no error bars or seeds reported","comparison":"Same model, same toolset and prompt, evaluated on the 78 simple vs. 74 complex VeriTrip queries","falsification":"Re-running the models on the VeriTrip simple and complex splits shows a PFR drop of less than about 5 points for several models, or a PFR increase for any model, or FR no longer higher on complex queries for most models.","caption":"Table 4: Main results of different MLLMs on VeriTrip. AM is measured in units of 10 km. Tool Calls indicates the number of tool calls. The colored values indicate the performance delta (Complex score - Simple score) for each metric. Red denotes an improvement, while green denotes a degradation. All subsequent experiments followed this setting.","image":"img/2605.28683.png","kind":"table","image_w":1127,"image_h":522,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.28683","title":"VeriTrip: A Verifiable Benchmark for Travel Planning Agents over Unstructured Web Corpora","authors":"Yuting Xu; Jiayi Tian; Jian Liang; Xin Xiong; Hang Zhang; Mu Xu; Xiao-Yu Zhang","openreview_id":"hFIKn7BCcY","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Travel-planning benchmarks for LLM agents (TravelPlanner, ChinaTravel, TripTailor) give the agent clean structured APIs for flights, hotels and attractions, so they test constraint satisfaction but not whether an agent can find the facts itself in noisy web content. Deep-research benchmarks test information seeking but usually grade free-form reports with an LLM judge, which misses fabricated details. VeriTrip combines the two: agents must search a frozen, noisy multimodal web snapshot to build an itinerary, and every factual cell of the output is checked automatically against a hidden, human-verified knowledge base.","method":["The Multimodal Retrieval Base (MRB) holds 8,210 documents and 4,146 images scraped from Ctrip/Trip.com, 12306 and RedNote for 15 cities in China and the US, with ads and PII removed but conflicting or outdated user content kept. Agents get docSearch (Qwen3-Embedding-8B dense retrieval, top-10 snippets), imgSearch (DINOv2 ViT-S/14), getDocument, and three structured restaurant tools. Each of the 228 queries (78 simple, 76 medium, 74 complex) has text constraints plus a heavily cropped photo of a target POI. Complex queries have more travel days, more travelers and more stated preferences (2-8 vs. 1-4). Models run in a ReAct loop and must output a fixed JSON itinerary schema.","Evaluation is programmatic: Delivery Rate (DR, valid JSON), Factual Reliability (FR, fraction of transport IDs, times, POI and hotel names that match the VKB, exact or fuzzy), Preference Fulfillment Rate (PFR, rule checks for the stated transport mode, room type, must-see attractions, meal preferences and inner-city transport), micro Pass Rate on commonsense constraints (PRmi), and Average Margin over a TSP-optimal route (AM). Eleven MLLMs are evaluated (8 without thinking, 3 with thinking)."],"claim":"On VeriTrip, Preference Fulfillment Rate is lower on complex than on simple queries for all 11 evaluated MLLMs, by 9.7 to 30.0 points (e.g., Claude-4.5-Sonnet 64.10 -> 50.00, GPT-4o 45.20 -> 32.53, o3 55.38 -> 27.57, Qwen3-VL-235B 66.46 -> 36.49), while Factual Reliability is higher on complex queries for all 11 models (by 2.1 to 11.3 points).","source":"Table 4, page 7","metric":"PFR and FR in % (higher is better); single run per model and difficulty level, no error bars or seeds reported","comparison":"Same model, same toolset and prompt, evaluated on the 78 simple vs. 74 complex VeriTrip queries","falsification":"Re-running the models on the VeriTrip simple and complex splits shows a PFR drop of less than about 5 points for several models, or a PFR increase for any model, or FR no longer higher on complex queries for most models.","caption":"Table 4: Main results of different MLLMs on VeriTrip. AM is measured in units of 10 km. Tool Calls indicates the number of tool calls. The colored values indicate the performance delta (Complex score - Simple score) for each metric. Red denotes an improvement, while green denotes a degradation. All subsequent experiments followed this setting.","image":"img/2605.28683.png","kind":"table","image_w":1127,"image_h":522,"pwc":true}
|
site/p/2605.28911.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.28911","title":"Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses","authors":"Jonathan Stray; David Yang; Steven Luo; Miu N Takagi; Serina Chang","openreview_id":"BqLBsgGqSC","presentation":"spotlight","track":"evaluations_datasets","verdict":"include","background":"LLMs are now a common source of political information, and AI companies and the U.S. government want 'politically neutral' models, but there is no agreed, testable definition of neutrality. Existing evaluations mostly measure slant along a single liberal-conservative axis with proprietary methods or LLM judges. This paper defines a neutral response as one that gets maximal and equal approval from people on opposing sides of a specific issue, and measures this with a large human survey.","method":["The authors picked 20 contested U.S. policy issues, each with a canonical survey question that splits respondents into a 'for' and an 'against' side (mapped to liberal/conservative per issue in Table B.1). For each issue they took 10 Reddit-sourced user prompts (two in each of five valences from Strongly For to Strongly Against). Each prompt was answered by 8 model/stance combinations: the default response of GPT 5.4, Claude Opus 4.6, Gemini 3 Flash Preview, Grok 4.1 Non-reasoning and Llama Maverick, plus three GPT 5.4 responses written from the 'for' side, from the 'against' side, and 'balanced' (one short paragraph per side and a closing sentence acknowledging trade-offs), giving 1,600 responses.","7,434 U.S. Prolific participants each got 4 random issues; for each they answered the canonical question (which sets their side) and rated one random, unlabelled response on seven 5-point Likert statements. Ratings are mapped to 0-1 (0, 0.25, ..., 1). The main analysis uses 'I approve of this AI response': per model/stance it averages approval among participants on the conservative side (s_cons) and liberal side (s_lib) over all 20 issues (Figure 3a, with 95% CIs), and separately fits a linear regression on ratings pooled over all 20 issues and all seven Likert statements (not only 'I approve'), with fixed effects for each model x stance x participant-side cell plus prompt charge, issue, Likert statement and demographics, and SEs clustered at the question and participant level (Figure 3b; Appendix D.3). The empirical maximum-equal-approval (MEA) response per issue is the response on the Pareto frontier of (s_cons, s_lib) minimising |s_cons - s_lib|."],"claim":"Averaged over all 20 issues, GPT 5.4's balanced response gets nearly equal approval from conservative-side and liberal-side participants (about 0.68 vs 0.69; mean s_cons - s_lib = -0.01, with overlapping 95% CIs and nearly identical regression coefficients), whereas GPT 5.4's default response to the same prompts leans liberal (about 0.60 vs 0.72; s_cons - s_lib = -0.12, with non-overlapping 95% CIs); the balanced response is the empirical MEA response on 60% of issues versus at most 15% for any other model/stance.","source":"Figure 3 (a and b), page 7; numbers from Table 1, page 7","metric":"Mean approval score on 'I approve of this AI response' mapped to 0-1 (higher is better), split by participant side; error bars in Figure 3 are 95% confidence intervals (3a: of the mean; 3b: coefficients of a regression pooled over all seven Likert statements, SEs clustered at question and participant level). Table 1 values (including MEA rate) have no error bars.","comparison":"GPT 5.4 balanced vs. GPT 5.4 default (and the other four model defaults) on the same 200 prompts and participant pool, comparing the gap between conservative-side and liberal-side approval","falsification":"Recomputing from the released PARETO data (or a new survey with the released prompts and responses), either the GPT default's liberal-minus-conservative approval gap is not significantly positive (95% CIs overlap), or the GPT balanced response shows a gap of comparable size (e.g. |s_cons - s_lib| >= 0.05 with non-overlapping CIs), or the balanced response is not the most frequent empirical MEA response across the 20 issues.","caption":"Figure 3: Summary of our results over all 20 issues. (a) Approval scores per model and model stance, from participants on the more conservative (x-axis) and more liberal (y-axis) side of the issue, averaged across all issues. Shaded lines indicate 95% confidence intervals. The red/blue arrows indicate the loss of approval from one-sided to balanced responses for conservative/liberal participants. (b) Approval regression coefficients for model, model stance, and participant side interaction terms, with 95% confidence intervals. See full model specification in Appendix D.3.","image":"img/2605.28911.png","kind":"figure","image_w":1144,"image_h":772,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.28911","title":"Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses","authors":"Jonathan Stray; David Yang; Steven Luo; Miu N Takagi; Serina Chang","openreview_id":"BqLBsgGqSC","presentation":"spotlight","track":"evaluations_datasets","verdict":"include","background":"LLMs are now a common source of political information, and AI companies and the U.S. government want 'politically neutral' models, but there is no agreed, testable definition of neutrality. Existing evaluations mostly measure slant along a single liberal-conservative axis with proprietary methods or LLM judges. This paper defines a neutral response as one that gets maximal and equal approval from people on opposing sides of a specific issue, and measures this with a large human survey.","method":["The authors picked 20 contested U.S. policy issues, each with a canonical survey question that splits respondents into a 'for' and an 'against' side (mapped to liberal/conservative per issue in Table B.1). For each issue they took 10 Reddit-sourced user prompts (two in each of five valences from Strongly For to Strongly Against). Each prompt was answered by 8 model/stance combinations: the default response of GPT 5.4, Claude Opus 4.6, Gemini 3 Flash Preview, Grok 4.1 Non-reasoning and Llama Maverick, plus three GPT 5.4 responses written from the 'for' side, from the 'against' side, and 'balanced' (one short paragraph per side and a closing sentence acknowledging trade-offs), giving 1,600 responses.","7,434 U.S. Prolific participants each got 4 random issues; for each they answered the canonical question (which sets their side) and rated one random, unlabelled response on seven 5-point Likert statements. Ratings are mapped to 0-1 (0, 0.25, ..., 1). The main analysis uses 'I approve of this AI response': per model/stance it averages approval among participants on the conservative side (s_cons) and liberal side (s_lib) over all 20 issues (Figure 3a, with 95% CIs), and separately fits a linear regression on ratings pooled over all 20 issues and all seven Likert statements (not only 'I approve'), with fixed effects for each model x stance x participant-side cell plus prompt charge, issue, Likert statement and demographics, and SEs clustered at the question and participant level (Figure 3b; Appendix D.3). The empirical maximum-equal-approval (MEA) response per issue is the response on the Pareto frontier of (s_cons, s_lib) minimising |s_cons - s_lib|."],"claim":"Averaged over all 20 issues, GPT 5.4's balanced response gets nearly equal approval from conservative-side and liberal-side participants (about 0.68 vs 0.69; mean s_cons - s_lib = -0.01, with overlapping 95% CIs and nearly identical regression coefficients), whereas GPT 5.4's default response to the same prompts leans liberal (about 0.60 vs 0.72; s_cons - s_lib = -0.12, with non-overlapping 95% CIs); the balanced response is the empirical MEA response on 60% of issues versus at most 15% for any other model/stance.","source":"Figure 3 (a and b), page 7; numbers from Table 1, page 7","metric":"Mean approval score on 'I approve of this AI response' mapped to 0-1 (higher is better), split by participant side; error bars in Figure 3 are 95% confidence intervals (3a: of the mean; 3b: coefficients of a regression pooled over all seven Likert statements, SEs clustered at question and participant level). Table 1 values (including MEA rate) have no error bars.","comparison":"GPT 5.4 balanced vs. GPT 5.4 default (and the other four model defaults) on the same 200 prompts and participant pool, comparing the gap between conservative-side and liberal-side approval","falsification":"Recomputing from the released PARETO data (or a new survey with the released prompts and responses), either the GPT default's liberal-minus-conservative approval gap is not significantly positive (95% CIs overlap), or the GPT balanced response shows a gap of comparable size (e.g. |s_cons - s_lib| >= 0.05 with non-overlapping CIs), or the balanced response is not the most frequent empirical MEA response across the 20 issues.","caption":"Figure 3: Summary of our results over all 20 issues. (a) Approval scores per model and model stance, from participants on the more conservative (x-axis) and more liberal (y-axis) side of the issue, averaged across all issues. Shaded lines indicate 95% confidence intervals. The red/blue arrows indicate the loss of approval from one-sided to balanced responses for conservative/liberal participants. (b) Approval regression coefficients for model, model stance, and participant side interaction terms, with 95% confidence intervals. See full model specification in Appendix D.3.","image":"img/2605.28911.png","kind":"figure","image_w":1144,"image_h":772,"pwc":true}
|
site/p/2605.29059.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.29059","title":"SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers","authors":"Kaihua Qin; Dawn Song; Arthur Gervais","openreview_id":"M1QDaACV7f","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"More than 99% of Ethereum smart contracts are deployed without verified source code, so auditors and users see only EVM bytecode. Traditional decompilers recover function selectors and low-level pseudocode but not compilable Solidity. LLMs can produce source-like Solidity that may compile, but there has been no shared benchmark for checking whether that output compiles, recovers the right interface, and behaves like the original. SCDBench provides 600 verified Ethereum contracts (200 each in easy, medium and hard complexity strata) with bytecode, ground-truth source and replayable semantic test cases, and it scores decompilers in four cumulative stages: format completeness, compilability, ABI recovery and semantic consistency.","method":["Each model gets the contract bytecode as EVM assembly text, plus non-unique function-selector and event-signature candidate mappings (from 4byte.sourcify.dev) when available. It must return three XML-style blocks: a compiler JSON (version and settings), self-contained Solidity with no imports, and a target contract name (exact prompts in Appendix E.1). An output passes compilability if it has the required format and compiles with solc, using its own stated compiler settings, into a valid artifact for the named contract. Scoring is cumulative over all 600 contracts, so format failures also count as compile failures.","Zero-shot compilation repair (marked †): if an output has the required format but fails to compile, the same model is called once with the generated output and the solc error message (prompt in Appendix E.2). The repaired output is used if one is available; otherwise the original is kept. Models, all via OpenRouter: Claude Opus 4.7 and GPT-5.3-Codex with high reasoning effort, GLM-5 with high thinking, and GLM-5 (instruct) without thinking. Each setting is a single run."],"claim":"On the 600 SCDBench contracts, one same-model compilation-repair call raises the share of decompiled outputs that compile for every model: GPT-5.3-Codex from 421/600 (70.2%) to 542/600 (90.3%), Opus 4.7 from 56.8% to 69.2%, GLM-5 from 36.8% to 67.2%, and GLM-5 (instruct) from 24.0% to 64.3%.","source":"Table 2, page 7 (replicated on a second 600-contract sample in Table 23, page 22)","metric":"Compilation success rate (higher is better): contracts whose output has the required format and compiles to a valid target artifact, over all 600 contracts (200 per stratum). Single run, no error bars; the binomial SE at n = 600 is about 2 percentage points.","comparison":"For each model, the initial zero-shot decompilation vs. the same output after one same-model repair call given the solc error, on the same contracts","falsification":"With the released prompts and data, one repair call raises a model's total compile rate by less than about 5 percentage points (e.g. GPT-5.3-Codex stays under 75%, or GLM-5 stays under 45%), or the gain disappears on the hard stratum.","caption":"Table 2: Compilation success by complexity. A contract passes if the recovered Solidity compiles to a valid target artifact. Each complexity stratum contains 200 contracts. Blue highlights the best unrepaired result, and green highlights the best result after zero-shot repair.","image":"img/2605.29059.png","kind":"table","image_w":1144,"image_h":444,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.29059","title":"SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers","authors":"Kaihua Qin; Dawn Song; Arthur Gervais","openreview_id":"M1QDaACV7f","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"More than 99% of Ethereum smart contracts are deployed without verified source code, so auditors and users see only EVM bytecode. Traditional decompilers recover function selectors and low-level pseudocode but not compilable Solidity. LLMs can produce source-like Solidity that may compile, but there has been no shared benchmark for checking whether that output compiles, recovers the right interface, and behaves like the original. SCDBench provides 600 verified Ethereum contracts (200 each in easy, medium and hard complexity strata) with bytecode, ground-truth source and replayable semantic test cases, and it scores decompilers in four cumulative stages: format completeness, compilability, ABI recovery and semantic consistency.","method":["Each model gets the contract bytecode as EVM assembly text, plus non-unique function-selector and event-signature candidate mappings (from 4byte.sourcify.dev) when available. It must return three XML-style blocks: a compiler JSON (version and settings), self-contained Solidity with no imports, and a target contract name (exact prompts in Appendix E.1). An output passes compilability if it has the required format and compiles with solc, using its own stated compiler settings, into a valid artifact for the named contract. Scoring is cumulative over all 600 contracts, so format failures also count as compile failures.","Zero-shot compilation repair (marked †): if an output has the required format but fails to compile, the same model is called once with the generated output and the solc error message (prompt in Appendix E.2). The repaired output is used if one is available; otherwise the original is kept. Models, all via OpenRouter: Claude Opus 4.7 and GPT-5.3-Codex with high reasoning effort, GLM-5 with high thinking, and GLM-5 (instruct) without thinking. Each setting is a single run."],"claim":"On the 600 SCDBench contracts, one same-model compilation-repair call raises the share of decompiled outputs that compile for every model: GPT-5.3-Codex from 421/600 (70.2%) to 542/600 (90.3%), Opus 4.7 from 56.8% to 69.2%, GLM-5 from 36.8% to 67.2%, and GLM-5 (instruct) from 24.0% to 64.3%.","source":"Table 2, page 7 (replicated on a second 600-contract sample in Table 23, page 22)","metric":"Compilation success rate (higher is better): contracts whose output has the required format and compiles to a valid target artifact, over all 600 contracts (200 per stratum). Single run, no error bars; the binomial SE at n = 600 is about 2 percentage points.","comparison":"For each model, the initial zero-shot decompilation vs. the same output after one same-model repair call given the solc error, on the same contracts","falsification":"With the released prompts and data, one repair call raises a model's total compile rate by less than about 5 percentage points (e.g. GPT-5.3-Codex stays under 75%, or GLM-5 stays under 45%), or the gain disappears on the hard stratum.","caption":"Table 2: Compilation success by complexity. A contract passes if the recovered Solidity compiles to a valid target artifact. Each complexity stratum contains 200 contracts. Blue highlights the best unrepaired result, and green highlights the best result after zero-shot repair.","image":"img/2605.29059.png","kind":"table","image_w":1144,"image_h":444,"pwc":true}
|
site/p/2605.29253.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.29253","title":"OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories","authors":"Yibing Liu; Yangze Liu; Xiao-Long Yin; Bin Wang; Chong Zhang; Hao Yin; Zhongyi Han","openreview_id":"DWMn45BwZe","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Agent benchmarks usually score whether a task was completed, but an agent can pass the task oracle while still behaving unreliably along the way, for example writing to external state before resolving ambiguity or ignoring tool errors. The paper calls this the Outcome-Process Gap and argues that detecting such process-side anomalies needs trajectory-level supervision that is separate from task success. It releases OpenClawBench, a dataset of real OpenClaw agent executions on BFCL tasks with anomaly labels, and uses it to train a small detector that flags anomalous trajectories.","method":["OpenClawBench collects 31,264 OpenClaw agent runs on BFCL tasks from 6 open-weight source models (DeepSeek-R1, GLM-4.5-Air, gpt-oss-20B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B), normalizes each into ReAct-style steps (thinking, action, feedback), and labels each trajectory with the FullTax LLM-annotation protocol as normal or one of 5 anomaly subtypes. Final released labels come from a single-call relabel by DeepSeek-V4-Pro, keeping only high-confidence labels: 23,857 train and 2,646 test trajectories (test anomaly rate 14.8%). Silver labels agree with the human binary label on 280/300 Gold-300 trajectories.","The detector is Gemma 3 12B Instruct with rank-32 LoRA (alpha 64, dropout 0.05, all attention and MLP projections, AdamW, peak lr 1.5e-4, cosine with 3% warmup, global batch 32, max length 8192, bf16) trained to emit one token from a 6-symbol vocabulary (normal + 5 subtypes); binary anomaly = any non-normal output. Normals are downsampled once (seed 42) to about 1:2 anomaly:normal, giving about 9,809 training examples; training runs in two stages: an initial 3-epoch run to step 921, then a resume from that checkpoint under a 10-epoch (3,070-step) schedule that was stopped manually at step 1,520 (about 4.95 epochs; lowest dev loss). Baselines on the same test split and prompt template: base Gemma 3 12B (no fine-tuning, greedy) and GPT-5.4 zero-shot (reasoning_effort=medium, JSON output mapped to the same labels)."],"claim":"On the OpenClawBench silver test split (n=2,646, 14.8% anomalous), the LoRA-fine-tuned Gemma 3 12B detector reaches binary anomaly F1 0.729 (precision 0.660, recall 0.813), versus 0.426 for zero-shot GPT-5.4 and 0.372 for the un-fine-tuned Gemma 3 12B, with the gain coming from precision (predicted-anomaly rate 18.2% vs. 41.6% for GPT-5.4).","source":"Table 2, page 8","metric":"Binary F1 for trajectory-level anomaly detection against silver (LLM-generated) labels, higher is better; single run, no error bars or confidence intervals reported for this split","comparison":"Gemma 3 12B + LoRA vs. GPT-5.4 zero-shot (reasoning) and base Gemma 3 12B, same test split and prompt template","falsification":"A LoRA fine-tune of Gemma 3 12B with the stated recipe on the released train split fails to reach a binary F1 clearly above GPT-5.4 zero-shot on the test split (e.g. below about 0.55), or its F1 is far below 0.729 (e.g. under 0.68) across reasonable seeds.","caption":"Table 2: Trajectory-level anomaly detection on the cleaner-labels test split (n=2,646, label-anomaly rate 14.8%). PredAnom% reports the rate at which each detector predicts an anomaly label; agreement with the 14.8% label rate is only a diagnostic, not a measure of probabilistic calibration.","image":"img/2605.29253.png","kind":"table","image_w":1144,"image_h":316,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.29253","title":"OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories","authors":"Yibing Liu; Yangze Liu; Xiao-Long Yin; Bin Wang; Chong Zhang; Hao Yin; Zhongyi Han","openreview_id":"DWMn45BwZe","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Agent benchmarks usually score whether a task was completed, but an agent can pass the task oracle while still behaving unreliably along the way, for example writing to external state before resolving ambiguity or ignoring tool errors. The paper calls this the Outcome-Process Gap and argues that detecting such process-side anomalies needs trajectory-level supervision that is separate from task success. It releases OpenClawBench, a dataset of real OpenClaw agent executions on BFCL tasks with anomaly labels, and uses it to train a small detector that flags anomalous trajectories.","method":["OpenClawBench collects 31,264 OpenClaw agent runs on BFCL tasks from 6 open-weight source models (DeepSeek-R1, GLM-4.5-Air, gpt-oss-20B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B), normalizes each into ReAct-style steps (thinking, action, feedback), and labels each trajectory with the FullTax LLM-annotation protocol as normal or one of 5 anomaly subtypes. Final released labels come from a single-call relabel by DeepSeek-V4-Pro, keeping only high-confidence labels: 23,857 train and 2,646 test trajectories (test anomaly rate 14.8%). Silver labels agree with the human binary label on 280/300 Gold-300 trajectories.","The detector is Gemma 3 12B Instruct with rank-32 LoRA (alpha 64, dropout 0.05, all attention and MLP projections, AdamW, peak lr 1.5e-4, cosine with 3% warmup, global batch 32, max length 8192, bf16) trained to emit one token from a 6-symbol vocabulary (normal + 5 subtypes); binary anomaly = any non-normal output. Normals are downsampled once (seed 42) to about 1:2 anomaly:normal, giving about 9,809 training examples; training runs in two stages: an initial 3-epoch run to step 921, then a resume from that checkpoint under a 10-epoch (3,070-step) schedule that was stopped manually at step 1,520 (about 4.95 epochs; lowest dev loss). Baselines on the same test split and prompt template: base Gemma 3 12B (no fine-tuning, greedy) and GPT-5.4 zero-shot (reasoning_effort=medium, JSON output mapped to the same labels)."],"claim":"On the OpenClawBench silver test split (n=2,646, 14.8% anomalous), the LoRA-fine-tuned Gemma 3 12B detector reaches binary anomaly F1 0.729 (precision 0.660, recall 0.813), versus 0.426 for zero-shot GPT-5.4 and 0.372 for the un-fine-tuned Gemma 3 12B, with the gain coming from precision (predicted-anomaly rate 18.2% vs. 41.6% for GPT-5.4).","source":"Table 2, page 8","metric":"Binary F1 for trajectory-level anomaly detection against silver (LLM-generated) labels, higher is better; single run, no error bars or confidence intervals reported for this split","comparison":"Gemma 3 12B + LoRA vs. GPT-5.4 zero-shot (reasoning) and base Gemma 3 12B, same test split and prompt template","falsification":"A LoRA fine-tune of Gemma 3 12B with the stated recipe on the released train split fails to reach a binary F1 clearly above GPT-5.4 zero-shot on the test split (e.g. below about 0.55), or its F1 is far below 0.729 (e.g. under 0.68) across reasonable seeds.","caption":"Table 2: Trajectory-level anomaly detection on the cleaner-labels test split (n=2,646, label-anomaly rate 14.8%). PredAnom% reports the rate at which each detector predicts an anomaly label; agreement with the 14.8% label rate is only a diagnostic, not a measure of probabilistic calibration.","image":"img/2605.29253.png","kind":"table","image_w":1144,"image_h":316,"pwc":true}
|
site/p/2605.30326.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.30326","title":"RoboWits: Unexpected Challenges for Robotic Creative Problem Solving","authors":"Chunru Lin; Hongxin Zhang; Fenghao Yu; Zhehuan Chen; Tom Griffiths; Yejin Choi; David Held; Chuang Gan","openreview_id":"0dyy3qodV0","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Robot manipulation benchmarks mostly test skill execution under well-specified conditions, so they say little about whether a policy can reason about geometry, materials or tool use and change strategy when the obvious solution is blocked. RoboWits is a simulated bi-manual benchmark built to probe this: each 'seed' task (e.g. retrieve a cube from a container) is expanded into 'mutated' variants in which small scene changes block the original solution, add deceptive trap objects, or add distractors, while the goal and success metric stay the same. If a policy has learned the physics of the task rather than surface correlations, its success should transfer from seeds to mutations.","method":["An LLM-agent pipeline (Seed Task Generator, Task Verifier, Metric Generator, Task Mutator, Scene Generator) produces 30 seed tasks and 200 tasks in total in the Genesis simulator, with a dual-arm robot (two 7-DoF arms, parallel grippers, two wrist cameras and one base camera, Cartesian control). Mutations are of three kinds: pivot (remove or modify a key object so a different solution is needed), trap (add objects that look useful but are not, e.g. fixed to the table or made of a soft material), and add (distractor objects). Mutations share the seed task's success metric. The main experiments use 10 seed task families (RoboWits-10: Align Blocks, Retrieve Cube, Dominos, Gap Retrieve, Stand Bulb, Cover Lid, Ball Tower, Hold Cup, Pinch Card, Water Mug).","ACT (trained from scratch), π0 and π0.5 are fine-tuned in a multi-task setting on 50 human teleoperated demonstrations per seed task, using seed tasks only; they never see a mutated task. Each policy is evaluated on the seed tasks and on every mutation of those tasks with randomized object positions and rotations. In the repeated evaluation (Appendix E.2) the whole evaluation is run three times with different random seeds, with 20 trials per mutation task, and the paper reports the mean and standard error of the success rate across the three runs, averaged over the 10 task families."],"claim":"On RoboWits-10, π0 fine-tuned on 50 seed-task demonstrations per task drops from 11.9 ± 0.6% average success on seed tasks to 7.2 ± 0.7% on their mutations (π0.5: 12.3 ± 0.3% to 8.1 ± 1.5%; ACT: 10.3 ± 0.7% to 4.5 ± 0.3%), i.e. every fine-tuned policy (both VLAs and ACT) loses success when the task is mutated, by several standard errors.","source":"Table 9, page 24 (Appendix E.2); single-run version in Table 3, page 9","metric":"Binary task success rate (%), higher is better, averaged over the 10 task families; ± is the standard error over three independent evaluation runs with different random seeds (20 trials per mutation task)","comparison":"The same fine-tuned policy (π0, also π0.5 and ACT) evaluated on the 10 seed tasks vs. the mutated variants of those tasks, which share the goal and success metric","falsification":"After fine-tuning π0 on 50 seed demonstrations per task and evaluating over several runs, the average mutation success rate is not clearly below the seed success rate (e.g. the gap is smaller than about 2 percentage points or within the run-to-run standard errors).","caption":"Table 9: Success rate (%) over three independent evaluation runs, reported as mean±SE. Mutation tasks are evaluated with 20 trials each. * Models assume oracle state. ** Models assume perfect action primitives.","image":"img/2605.30326.png","kind":"table","image_w":1144,"image_h":338,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.30326","title":"RoboWits: Unexpected Challenges for Robotic Creative Problem Solving","authors":"Chunru Lin; Hongxin Zhang; Fenghao Yu; Zhehuan Chen; Tom Griffiths; Yejin Choi; David Held; Chuang Gan","openreview_id":"0dyy3qodV0","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Robot manipulation benchmarks mostly test skill execution under well-specified conditions, so they say little about whether a policy can reason about geometry, materials or tool use and change strategy when the obvious solution is blocked. RoboWits is a simulated bi-manual benchmark built to probe this: each 'seed' task (e.g. retrieve a cube from a container) is expanded into 'mutated' variants in which small scene changes block the original solution, add deceptive trap objects, or add distractors, while the goal and success metric stay the same. If a policy has learned the physics of the task rather than surface correlations, its success should transfer from seeds to mutations.","method":["An LLM-agent pipeline (Seed Task Generator, Task Verifier, Metric Generator, Task Mutator, Scene Generator) produces 30 seed tasks and 200 tasks in total in the Genesis simulator, with a dual-arm robot (two 7-DoF arms, parallel grippers, two wrist cameras and one base camera, Cartesian control). Mutations are of three kinds: pivot (remove or modify a key object so a different solution is needed), trap (add objects that look useful but are not, e.g. fixed to the table or made of a soft material), and add (distractor objects). Mutations share the seed task's success metric. The main experiments use 10 seed task families (RoboWits-10: Align Blocks, Retrieve Cube, Dominos, Gap Retrieve, Stand Bulb, Cover Lid, Ball Tower, Hold Cup, Pinch Card, Water Mug).","ACT (trained from scratch), π0 and π0.5 are fine-tuned in a multi-task setting on 50 human teleoperated demonstrations per seed task, using seed tasks only; they never see a mutated task. Each policy is evaluated on the seed tasks and on every mutation of those tasks with randomized object positions and rotations. In the repeated evaluation (Appendix E.2) the whole evaluation is run three times with different random seeds, with 20 trials per mutation task, and the paper reports the mean and standard error of the success rate across the three runs, averaged over the 10 task families."],"claim":"On RoboWits-10, π0 fine-tuned on 50 seed-task demonstrations per task drops from 11.9 ± 0.6% average success on seed tasks to 7.2 ± 0.7% on their mutations (π0.5: 12.3 ± 0.3% to 8.1 ± 1.5%; ACT: 10.3 ± 0.7% to 4.5 ± 0.3%), i.e. every fine-tuned policy (both VLAs and ACT) loses success when the task is mutated, by several standard errors.","source":"Table 9, page 24 (Appendix E.2); single-run version in Table 3, page 9","metric":"Binary task success rate (%), higher is better, averaged over the 10 task families; ± is the standard error over three independent evaluation runs with different random seeds (20 trials per mutation task)","comparison":"The same fine-tuned policy (π0, also π0.5 and ACT) evaluated on the 10 seed tasks vs. the mutated variants of those tasks, which share the goal and success metric","falsification":"After fine-tuning π0 on 50 seed demonstrations per task and evaluating over several runs, the average mutation success rate is not clearly below the seed success rate (e.g. the gap is smaller than about 2 percentage points or within the run-to-run standard errors).","caption":"Table 9: Success rate (%) over three independent evaluation runs, reported as mean±SE. Mutation tasks are evaluated with 20 trials each. * Models assume oracle state. ** Models assume perfect action primitives.","image":"img/2605.30326.png","kind":"table","image_w":1144,"image_h":338,"pwc":true}
|
site/p/2605.31572.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.31572","title":"nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving","authors":"Zhiyu Huang; Johnson Liu; Rui Song; Zewei Zhou; Ruining Yang; Yun Zhang; Tianhui Cai; Hanyin Zhang; Mingxuan Gao; Valeria Xu; Jiali Chen; Yishan Shen; Yiluan Guo; Tony Qi; Jiaqi Ma","openreview_id":"WDTaPI01Fw","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"End-to-end driving models are mostly trained to imitate expert trajectories, which may not generalize to rare long-tail scenarios (work zones, emergency vehicles, animals on the road) that need scene understanding and reasoning. Vision-language-action (VLA) models promise reasoning, but existing driving datasets either evaluate question answering without planning, or contain routine driving with little reasoning supervision. nuReasoning is a 20K-clip (about 105 hours) real-world long-tail dataset from Motional fleet logs with multi-camera images, LiDAR, HD maps, object boxes, and human-verified spatial, decision, and counterfactual reasoning annotations. Its benchmark evaluates both VLM question answering and trajectory planning, so it can test whether reasoning supervision improves driving itself.","method":["nuVLA is a Qwen3-VL-2B-Instruct backbone (fine-tuned with LoRA, rank 32) plus a flow-matching Diffusion Transformer action expert (12 layers, hidden 512). Inputs are 8 surround camera views at t=0 and t=-1 s (448x448) and a prompt with the mission command. The action expert cross-attends to the prompt-side last-layer VLM hidden states and also gets ego dynamics and ego history; it predicts a 5 s, 10-waypoint ego trajectory. Training jointly minimizes the language loss on reasoning-target text and the flow-matching action loss (lambda = 1), for 10 epochs, effective batch 64, lr 5e-5 (VLM) and 1e-4 (action expert). Reasoning generation is disabled at inference, so reasoning supervision only shapes the backbone features.","The controlled variants share architecture, training clip pool, planning targets, optimizer and inference, and differ only in which reasoning text supervises the VLM: none (planning only), spatial (S), decision (D), counterfactual (CF), and combinations. Planning is scored open-loop on the 1K-clip private test split with the nuReasoning Planning Score, NPS = s_NC * s_DA * (0.3 s_EP + 0.2 s_Comf + 0.5 s_HL), from no-at-fault collision, drivable-area compliance, ego progress, comfort, and human-likeness, on a 0-100 scale."],"claim":"On the nuReasoning test set, nuVLA trained with spatial + decision + counterfactual reasoning supervision reaches NPS 72.86 ± 0.07 versus 64.52 ± 0.31 for the identical planning-only nuVLA (+8.34), and every reasoning-supervised variant (S 70.18, D 71.51, S+D 71.19, D+CF 72.33) also beats planning-only by more than 5 points.","source":"Table S8 (Appendix F.4), page 32; single-seed version in Table 4, page 9","metric":"nuReasoning Planning Score (NPS, 0-100, higher is better), open-loop over a 5 s horizon; mean ± std over 3 training seeds","comparison":"nuVLA with reasoning supervision (S, D, D+CF, S+D, S+D+CF) vs. the same nuVLA with planning supervision only, same architecture, data pool and training protocol","falsification":"Retrained under the same protocol, the S+D+CF nuVLA's NPS does not exceed the planning-only nuVLA by clearly more than the seed std (e.g. the gap shrinks to around 1-2 NPS or reverses).","caption":"Table S8: nuVLA planning ablation across three random seeds (mean ± standard deviation). All variants receive planning supervision; Comf denotes comfort.","image":"img/2605.31572.png","kind":"table","image_w":1144,"image_h":294,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.31572","title":"nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving","authors":"Zhiyu Huang; Johnson Liu; Rui Song; Zewei Zhou; Ruining Yang; Yun Zhang; Tianhui Cai; Hanyin Zhang; Mingxuan Gao; Valeria Xu; Jiali Chen; Yishan Shen; Yiluan Guo; Tony Qi; Jiaqi Ma","openreview_id":"WDTaPI01Fw","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"End-to-end driving models are mostly trained to imitate expert trajectories, which may not generalize to rare long-tail scenarios (work zones, emergency vehicles, animals on the road) that need scene understanding and reasoning. Vision-language-action (VLA) models promise reasoning, but existing driving datasets either evaluate question answering without planning, or contain routine driving with little reasoning supervision. nuReasoning is a 20K-clip (about 105 hours) real-world long-tail dataset from Motional fleet logs with multi-camera images, LiDAR, HD maps, object boxes, and human-verified spatial, decision, and counterfactual reasoning annotations. Its benchmark evaluates both VLM question answering and trajectory planning, so it can test whether reasoning supervision improves driving itself.","method":["nuVLA is a Qwen3-VL-2B-Instruct backbone (fine-tuned with LoRA, rank 32) plus a flow-matching Diffusion Transformer action expert (12 layers, hidden 512). Inputs are 8 surround camera views at t=0 and t=-1 s (448x448) and a prompt with the mission command. The action expert cross-attends to the prompt-side last-layer VLM hidden states and also gets ego dynamics and ego history; it predicts a 5 s, 10-waypoint ego trajectory. Training jointly minimizes the language loss on reasoning-target text and the flow-matching action loss (lambda = 1), for 10 epochs, effective batch 64, lr 5e-5 (VLM) and 1e-4 (action expert). Reasoning generation is disabled at inference, so reasoning supervision only shapes the backbone features.","The controlled variants share architecture, training clip pool, planning targets, optimizer and inference, and differ only in which reasoning text supervises the VLM: none (planning only), spatial (S), decision (D), counterfactual (CF), and combinations. Planning is scored open-loop on the 1K-clip private test split with the nuReasoning Planning Score, NPS = s_NC * s_DA * (0.3 s_EP + 0.2 s_Comf + 0.5 s_HL), from no-at-fault collision, drivable-area compliance, ego progress, comfort, and human-likeness, on a 0-100 scale."],"claim":"On the nuReasoning test set, nuVLA trained with spatial + decision + counterfactual reasoning supervision reaches NPS 72.86 ± 0.07 versus 64.52 ± 0.31 for the identical planning-only nuVLA (+8.34), and every reasoning-supervised variant (S 70.18, D 71.51, S+D 71.19, D+CF 72.33) also beats planning-only by more than 5 points.","source":"Table S8 (Appendix F.4), page 32; single-seed version in Table 4, page 9","metric":"nuReasoning Planning Score (NPS, 0-100, higher is better), open-loop over a 5 s horizon; mean ± std over 3 training seeds","comparison":"nuVLA with reasoning supervision (S, D, D+CF, S+D, S+D+CF) vs. the same nuVLA with planning supervision only, same architecture, data pool and training protocol","falsification":"Retrained under the same protocol, the S+D+CF nuVLA's NPS does not exceed the planning-only nuVLA by clearly more than the seed std (e.g. the gap shrinks to around 1-2 NPS or reverses).","caption":"Table S8: nuVLA planning ablation across three random seeds (mean ± standard deviation). All variants receive planning supervision; Comf denotes comfort.","image":"img/2605.31572.png","kind":"table","image_w":1144,"image_h":294,"pwc":true}
|
site/p/2606.01848.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.01848","title":"RescueBench: Can Embodied Agents Save Lives in the Wild?","authors":"Kui Wu; Beiyu Guo; Hao Chen; Shuhang Xu; Yuling Li; Yongdan Zeng; Zhoujun Li; Yizhou Wang; Fangwei Zhong","openreview_id":"d1QPUUrYxZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Search-and-rescue is a long-horizon embodied task. An agent must explore an unfamiliar environment from vague clues, find and pick up a victim, remember the way back, and hand the victim off at an ambulance. Existing embodied benchmarks test these abilities one at a time: instruction-following navigation, tabletop manipulation, or goal-reaching with a known destination. So it is unclear how current navigation architectures behave when exploration, interaction and spatial memory have to be chained, and where in that chain they break.","method":["RescueBench runs in photo-realistic UnrealZoo/UE5 scenes (7 environments, 213 test episodes). Each episode has four sequential stages: S1 explore from a text+image clue, S2 reach and pick up the injured person, S3 return to the ambulance from memory, S4 hand off at the stretcher. Five difficulty levels go from close-range targets in a clean room (L1, n=25) and in cluttered scenes (L2, n=59), to mid-range search (L3, n=59), cross-region indoor/outdoor search through doors (L4, n=39), and multi-floor search (L5, n=31). Time limits are 180 s (L1-L2), 240 s (L3) and 300 s (L4-L5). Except for ROCKET-2, which has native pick/drop/open-door actions, interactions are triggered automatically once the agent is within 1.5 m.","Seven baselines are evaluated zero-shot with their pretrained weights: LLM-YOLO Planner (YOLO + GPT-4o), Uni-NaVid, ViNT, NoMaD, SG-Nav, OmniNav and ROCKET-2. Four of them (Uni-NaVid, ViNT, NoMaD, ROCKET-2) are also fine-tuned on about 400K automatically collected expert steps. Human players and a ground-truth Oracle serve as references. Metrics are Task Completion Rate (TCR, the fraction of episodes in which all four stages succeed) and Task Score (TS, 0-100: four stages × 25 points, each proportional to the distance closed toward that stage's goal)."],"claim":"Under zero-shot evaluation, ROCKET-2 reaches Task Score 42.6 at L3 and 32.2 at L4, about 2.5x or more any other zero-shot baseline (next best: Uni-NaVid 14.2 at L3 and 12.9 at L4). No learned method, zero-shot or fine-tuned, completes any L5 episode (TCR 0%, n=31), while human players reach 93.5% TCR at L5.","source":"Table 3, page 14 (appendix B.1); summarized in Figure 5, page 7","metric":"TCR (%) and Task Score (0-100), higher is better; single evaluation pass per method and level, no ± reported","comparison":"ROCKET-2 (zero-shot) vs. LLM-YOLO Planner, Uni-NaVid, ViNT, NoMaD, SG-Nav and OmniNav (zero-shot) at L3/L4; all learned methods vs. human players at L5","falsification":"Refuted if, on the L3 or L4 test episodes, any other zero-shot baseline gets within about 1.5x of ROCKET-2's Task Score, or ROCKET-2's L3 TS falls below about 25. Also refuted if any of the evaluated learned methods completes a meaningful fraction (e.g., ≥5%) of L5 episodes.","caption":"Table 3: Overall performance across five difficulty levels, covering all baseline methods and their fine-tuned version, followed by \"(fine-tuned)\". Each cell reports TCR (%) / TS (0–100) / Avg Time Cost (s) / Avg Collision Counts.","image":"img/2606.01848.png","kind":"table","image_w":1144,"image_h":427,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.01848","title":"RescueBench: Can Embodied Agents Save Lives in the Wild?","authors":"Kui Wu; Beiyu Guo; Hao Chen; Shuhang Xu; Yuling Li; Yongdan Zeng; Zhoujun Li; Yizhou Wang; Fangwei Zhong","openreview_id":"d1QPUUrYxZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Search-and-rescue is a long-horizon embodied task. An agent must explore an unfamiliar environment from vague clues, find and pick up a victim, remember the way back, and hand the victim off at an ambulance. Existing embodied benchmarks test these abilities one at a time: instruction-following navigation, tabletop manipulation, or goal-reaching with a known destination. So it is unclear how current navigation architectures behave when exploration, interaction and spatial memory have to be chained, and where in that chain they break.","method":["RescueBench runs in photo-realistic UnrealZoo/UE5 scenes (7 environments, 213 test episodes). Each episode has four sequential stages: S1 explore from a text+image clue, S2 reach and pick up the injured person, S3 return to the ambulance from memory, S4 hand off at the stretcher. Five difficulty levels go from close-range targets in a clean room (L1, n=25) and in cluttered scenes (L2, n=59), to mid-range search (L3, n=59), cross-region indoor/outdoor search through doors (L4, n=39), and multi-floor search (L5, n=31). Time limits are 180 s (L1-L2), 240 s (L3) and 300 s (L4-L5). Except for ROCKET-2, which has native pick/drop/open-door actions, interactions are triggered automatically once the agent is within 1.5 m.","Seven baselines are evaluated zero-shot with their pretrained weights: LLM-YOLO Planner (YOLO + GPT-4o), Uni-NaVid, ViNT, NoMaD, SG-Nav, OmniNav and ROCKET-2. Four of them (Uni-NaVid, ViNT, NoMaD, ROCKET-2) are also fine-tuned on about 400K automatically collected expert steps. Human players and a ground-truth Oracle serve as references. Metrics are Task Completion Rate (TCR, the fraction of episodes in which all four stages succeed) and Task Score (TS, 0-100: four stages × 25 points, each proportional to the distance closed toward that stage's goal)."],"claim":"Under zero-shot evaluation, ROCKET-2 reaches Task Score 42.6 at L3 and 32.2 at L4, about 2.5x or more any other zero-shot baseline (next best: Uni-NaVid 14.2 at L3 and 12.9 at L4). No learned method, zero-shot or fine-tuned, completes any L5 episode (TCR 0%, n=31), while human players reach 93.5% TCR at L5.","source":"Table 3, page 14 (appendix B.1); summarized in Figure 5, page 7","metric":"TCR (%) and Task Score (0-100), higher is better; single evaluation pass per method and level, no ± reported","comparison":"ROCKET-2 (zero-shot) vs. LLM-YOLO Planner, Uni-NaVid, ViNT, NoMaD, SG-Nav and OmniNav (zero-shot) at L3/L4; all learned methods vs. human players at L5","falsification":"Refuted if, on the L3 or L4 test episodes, any other zero-shot baseline gets within about 1.5x of ROCKET-2's Task Score, or ROCKET-2's L3 TS falls below about 25. Also refuted if any of the evaluated learned methods completes a meaningful fraction (e.g., ≥5%) of L5 episodes.","caption":"Table 3: Overall performance across five difficulty levels, covering all baseline methods and their fine-tuned version, followed by \"(fine-tuned)\". Each cell reports TCR (%) / TS (0–100) / Avg Time Cost (s) / Avg Collision Counts.","image":"img/2606.01848.png","kind":"table","image_w":1144,"image_h":427,"pwc":true}
|
site/p/2606.08340.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.08340","title":"Benchmarking Open-Ended Multi-Agent Coordination in Language Agents","authors":"Kale-ab Tessera; Andras Szecsenyi; Cameron Barker; Alexander Rutherford; Davide Paglieri; Aidan Scannell; Henry Gouk; Elliot Crowley; Tim Rocktäschel; Amos Storkey","openreview_id":"GjoyYpGTqL","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLM agents are increasingly expected to work alongside other agents over long horizons, but most agent benchmarks test single agents, short interactions, or rigidly structured multi-agent tasks. ALEM is a JAX benchmark built on Craftax-Coop dynamics (a procedurally generated survival/crafting world). Each episode it samples coordination tasks (synchronous joint actions, timed handovers), soft role specialisation and an explicit communication channel, and a single scalar alpha sets coordination difficulty. With the benchmark, the authors ask whether zero-shot LLM teams can coordinate and which harness components (communication, scratchpad memory, reasoning) make coordination possible.","method":["Homogeneous teams of LLM agents (the prompts describe 3-agent tasks) play ALEM on the Hard setting (alpha = 0.9) through a text interface. At every step, each agent gets the game rules, a text rendering of its local observation, the last eight observations and actions, teammate messages, its private scratchpad, the visible coordination constraints, and the legal actions. It must return a structured response with exactly one action, plus optional reasoning, a broadcast message and a scratchpad update. Rewards are normalised per category: Base% (ordinary achievements and health) and Coord.% (27 coordination achievements), each a percentage of the maximum achievable reward, averaged over agents.","In the ablation, the default harness (communication, scratchpad memory and reasoning all on) is compared with the same harness with the broadcast communication channel removed. Everything else stays the same: Gemma-4-31B-it served with vLLM, reasoning enabled, Hard difficulty, 20 seeds. Results are cross-seed means with 95% stratified bootstrap CIs (rliable). The same ablation was also run on Gemini-3.1-Pro-High (10 seeds, API)."],"claim":"On ALEM Hard (alpha = 0.9), removing inter-agent communication from Gemma-4-31B-it teams cuts coordination reward from 8.8% (95% CI 6.9-10.9) to 3.8% (CI roughly 2.5-5.2) of the maximum, with non-overlapping CIs, while Base% does not fall (7.3% -> 8.3%).","source":"Figure 5(b), page 8 (baseline CI from Table 1, page 6)","metric":"Coord.%: coordination reward as a % of the maximum achievable (higher is better). Error bars are 95% stratified bootstrap CIs over 20 seeds. Base% is reported alongside.","comparison":"Gemma-4-31B-it, default harness (communication + memory + reasoning) vs. the same harness without communication, both on Hard. Secondary: the same ablation on Gemini-3.1-Pro-High (17.5 -> 5.3 Coord.%).","falsification":"Over 20 seeds on Hard with Gemma-4-31B-it, the no-communication Coord.% has a 95% CI that overlaps the full-harness CI, or the drop is much smaller than the reported ~5 points (e.g. under 2 points).","caption":"Figure 5: Harness and team-composition ablations on Hard. Left: we ablate communication, scratchpad memory, and reasoning for Gemini-3.1-Pro-High and Gemma-4-31B-it. Bars show Coord.% and Base%, each normalised by the maximum achievable reward in that category, with 95% bootstrap confidence intervals. Right: we compare heterogeneous teams against their constituent homogeneous baselines. Bars show mean Total% with 95% stratified bootstrap confidence intervals; the dashed line marks the average homogeneous-team baseline, and the text annotation reports the difference relative to that baseline.","image":"img/2606.08340.png","kind":"figure","image_w":1116,"image_h":677,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.08340","title":"Benchmarking Open-Ended Multi-Agent Coordination in Language Agents","authors":"Kale-ab Tessera; Andras Szecsenyi; Cameron Barker; Alexander Rutherford; Davide Paglieri; Aidan Scannell; Henry Gouk; Elliot Crowley; Tim Rocktäschel; Amos Storkey","openreview_id":"GjoyYpGTqL","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLM agents are increasingly expected to work alongside other agents over long horizons, but most agent benchmarks test single agents, short interactions, or rigidly structured multi-agent tasks. ALEM is a JAX benchmark built on Craftax-Coop dynamics (a procedurally generated survival/crafting world). Each episode it samples coordination tasks (synchronous joint actions, timed handovers), soft role specialisation and an explicit communication channel, and a single scalar alpha sets coordination difficulty. With the benchmark, the authors ask whether zero-shot LLM teams can coordinate and which harness components (communication, scratchpad memory, reasoning) make coordination possible.","method":["Homogeneous teams of LLM agents (the prompts describe 3-agent tasks) play ALEM on the Hard setting (alpha = 0.9) through a text interface. At every step, each agent gets the game rules, a text rendering of its local observation, the last eight observations and actions, teammate messages, its private scratchpad, the visible coordination constraints, and the legal actions. It must return a structured response with exactly one action, plus optional reasoning, a broadcast message and a scratchpad update. Rewards are normalised per category: Base% (ordinary achievements and health) and Coord.% (27 coordination achievements), each a percentage of the maximum achievable reward, averaged over agents.","In the ablation, the default harness (communication, scratchpad memory and reasoning all on) is compared with the same harness with the broadcast communication channel removed. Everything else stays the same: Gemma-4-31B-it served with vLLM, reasoning enabled, Hard difficulty, 20 seeds. Results are cross-seed means with 95% stratified bootstrap CIs (rliable). The same ablation was also run on Gemini-3.1-Pro-High (10 seeds, API)."],"claim":"On ALEM Hard (alpha = 0.9), removing inter-agent communication from Gemma-4-31B-it teams cuts coordination reward from 8.8% (95% CI 6.9-10.9) to 3.8% (CI roughly 2.5-5.2) of the maximum, with non-overlapping CIs, while Base% does not fall (7.3% -> 8.3%).","source":"Figure 5(b), page 8 (baseline CI from Table 1, page 6)","metric":"Coord.%: coordination reward as a % of the maximum achievable (higher is better). Error bars are 95% stratified bootstrap CIs over 20 seeds. Base% is reported alongside.","comparison":"Gemma-4-31B-it, default harness (communication + memory + reasoning) vs. the same harness without communication, both on Hard. Secondary: the same ablation on Gemini-3.1-Pro-High (17.5 -> 5.3 Coord.%).","falsification":"Over 20 seeds on Hard with Gemma-4-31B-it, the no-communication Coord.% has a 95% CI that overlaps the full-harness CI, or the drop is much smaller than the reported ~5 points (e.g. under 2 points).","caption":"Figure 5: Harness and team-composition ablations on Hard. Left: we ablate communication, scratchpad memory, and reasoning for Gemini-3.1-Pro-High and Gemma-4-31B-it. Bars show Coord.% and Base%, each normalised by the maximum achievable reward in that category, with 95% bootstrap confidence intervals. Right: we compare heterogeneous teams against their constituent homogeneous baselines. Bars show mean Total% with 95% stratified bootstrap confidence intervals; the dashed line marks the average homogeneous-team baseline, and the text annotation reports the difference relative to that baseline.","image":"img/2606.08340.png","kind":"figure","image_w":1116,"image_h":677,"pwc":true}
|
site/p/2606.14299.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.14299","title":"What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from an Update Perspective","authors":"Jiazhen Huang; Xiao Chen; Zhiming Liu; Yaru Sun; Jingyan Jiang; Zhi Wang","openreview_id":"jUrzOPgzZq","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"CLIP's zero-shot accuracy drops under distribution shift, and many test-time adaptation (TTA) methods for CLIP have been proposed that adapt on the unlabeled test stream. A common family, parameter-based TTA, is often credited to the gradient update itself: for each test image it builds augmented views, keeps the most confident ones, and takes a gradient step on an unsupervised entropy objective. The paper re-implements more than 20 methods in one codebase (TTABC) and asks which part of this pipeline actually produces the gain. Its answer is that the quantity and quality of test-time evidence (how many augmented views there are, and which of them are kept) matter more than how strongly or how long the model is optimized.","method":["Two representative parameter-based methods are used with CLIP ViT-B/16 on ImageNet-A (7,500 images, 200 classes): TPT, which tunes 4 learnable text-prompt tokens initialized as 'a photo of a', and TPS, which learns per-class shift vectors added to the text prototypes (prototypes built from templates plus GPT-4 descriptions). For each test image, AugMix produces 63 augmented views plus the original (N = 64). The views are ranked by prediction entropy, the top fraction p is kept, and one SGD step (lr 5e-3) minimizes the marginal entropy of the averaged prediction over the kept views. The parameters are reset after every sample (episodic), and the prediction is made on the original image after the update.","The evidence-quality ablation (Table 12) varies only the selection ratio p in {1.0, 0.75, 0.5, 0.25, 0.1}, with entropy or maximum-softmax-probability (MSP) as the ranking score. The optimization-depth ablation (Table 11) keeps p = 0.1 and varies only the number of gradient steps per sample from 1 to 20, recording wall-clock time."],"claim":"On ImageNet-A with CLIP ViT-B/16 (64 AugMix views, 1 step), keeping only the top 10% lowest-entropy views instead of all views raises accuracy from 50.75% to 58.47% for TPS (+7.7) and from 49.07% to 52.03% for TPT (+3.0). For TPS this exceeds the gain from going from 1 to 20 gradient steps (58.47% to 59.21%, +0.7).","source":"Table 12 and Table 11 (Appendix F.4-F.5), page 23; summarized in Figure 4(b) and Figure 3(b) on pages 4-5","metric":"Top-1 accuracy (%) on ImageNet-A, higher is better; single run with a fixed seed and no error bars (the binomial SE over 7,500 images is about 0.6 points)","comparison":"Same method and backbone, with the view-selection ratio p = 1.0 (all 64 views) vs. p = 0.1 (top 10% by entropy), 1 step; secondarily, 1 vs. 20 adaptation steps at p = 0.1","falsification":"With the same setup, entropy-based top-10% filtering improves TPS by less than about 3 points over using all views (or improves TPT by less than about 1.5 points), or going from 1 to 20 steps improves TPS by more than filtering does.","caption":"Table 11: Scaling Adaptation Steps. Accuracy (%) and total inference time (seconds) on ImageNet-A as the number of per-sample gradient steps increases. Table 12: Impact of Confidence Filtering Strategies. Top-1 Accuracy (%) of TPT and TPS on ImageNet-A across different view selection ratios (p).","image":"img/2606.14299.png","kind":"table","image_w":1133,"image_h":650,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.14299","title":"What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from an Update Perspective","authors":"Jiazhen Huang; Xiao Chen; Zhiming Liu; Yaru Sun; Jingyan Jiang; Zhi Wang","openreview_id":"jUrzOPgzZq","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"CLIP's zero-shot accuracy drops under distribution shift, and many test-time adaptation (TTA) methods for CLIP have been proposed that adapt on the unlabeled test stream. A common family, parameter-based TTA, is often credited to the gradient update itself: for each test image it builds augmented views, keeps the most confident ones, and takes a gradient step on an unsupervised entropy objective. The paper re-implements more than 20 methods in one codebase (TTABC) and asks which part of this pipeline actually produces the gain. Its answer is that the quantity and quality of test-time evidence (how many augmented views there are, and which of them are kept) matter more than how strongly or how long the model is optimized.","method":["Two representative parameter-based methods are used with CLIP ViT-B/16 on ImageNet-A (7,500 images, 200 classes): TPT, which tunes 4 learnable text-prompt tokens initialized as 'a photo of a', and TPS, which learns per-class shift vectors added to the text prototypes (prototypes built from templates plus GPT-4 descriptions). For each test image, AugMix produces 63 augmented views plus the original (N = 64). The views are ranked by prediction entropy, the top fraction p is kept, and one SGD step (lr 5e-3) minimizes the marginal entropy of the averaged prediction over the kept views. The parameters are reset after every sample (episodic), and the prediction is made on the original image after the update.","The evidence-quality ablation (Table 12) varies only the selection ratio p in {1.0, 0.75, 0.5, 0.25, 0.1}, with entropy or maximum-softmax-probability (MSP) as the ranking score. The optimization-depth ablation (Table 11) keeps p = 0.1 and varies only the number of gradient steps per sample from 1 to 20, recording wall-clock time."],"claim":"On ImageNet-A with CLIP ViT-B/16 (64 AugMix views, 1 step), keeping only the top 10% lowest-entropy views instead of all views raises accuracy from 50.75% to 58.47% for TPS (+7.7) and from 49.07% to 52.03% for TPT (+3.0). For TPS this exceeds the gain from going from 1 to 20 gradient steps (58.47% to 59.21%, +0.7).","source":"Table 12 and Table 11 (Appendix F.4-F.5), page 23; summarized in Figure 4(b) and Figure 3(b) on pages 4-5","metric":"Top-1 accuracy (%) on ImageNet-A, higher is better; single run with a fixed seed and no error bars (the binomial SE over 7,500 images is about 0.6 points)","comparison":"Same method and backbone, with the view-selection ratio p = 1.0 (all 64 views) vs. p = 0.1 (top 10% by entropy), 1 step; secondarily, 1 vs. 20 adaptation steps at p = 0.1","falsification":"With the same setup, entropy-based top-10% filtering improves TPS by less than about 3 points over using all views (or improves TPT by less than about 1.5 points), or going from 1 to 20 steps improves TPS by more than filtering does.","caption":"Table 11: Scaling Adaptation Steps. Accuracy (%) and total inference time (seconds) on ImageNet-A as the number of per-sample gradient steps increases. Table 12: Impact of Confidence Filtering Strategies. Top-1 Accuracy (%) of TPT and TPS on ImageNet-A across different view selection ratios (p).","image":"img/2606.14299.png","kind":"table","image_w":1133,"image_h":650,"pwc":true}
|
site/p/2606.15320.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.15320","title":"Conditional Multi-Event Temporal Grounding in Long-Form Video","authors":"Yuanhao Zou; Arthad Kulkarni; Lucas Toñanez; Lincoln Spencer; Guangyu Sun; Tianxingjian Ding; Andong Deng; Yi Li; Shuangjun Liu; Yuan Li; Dashan Gao; Ning Bi; TAOTAO JING; Shuai Zhang; Chen Chen","openreview_id":"i6iM7B1Koa","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Video temporal grounding asks a model to find the time interval in a video that matches a text query. Real uses such as video editing or sports analysis often need every occurrence of an event that meets temporal conditions (after X, while Y, within a time window) and spatial conditions (a specific person, a motion), along with an exact count, and sometimes the correct answer is that no such event exists. Existing benchmarks handle single-moment grounding, unconditioned counting, or multiple-choice QA separately. The paper introduces CoMET-Bench (2,789 queries over 600 long videos averaging 33.8 minutes, 26.6% negative queries) and a training-free agent, CoMET-Agent, for this task.","method":["Baselines are video MLLMs run single-pass with a shared evaluation prompt and 128 uniformly sampled frames per video. They output a set of [start, end] intervals per query, or an empty set. Grounding is scored on positive queries with mIoU, Recall@0.5 and F1@0.5 (intervals matched at tIoU 0.5). Counting is scored with MAE, off-by-one accuracy and Pearson correlation. Negative queries are scored with Rejection-F1 (the harmonic mean of the rejection rate on negatives and non-empty coverage on positives) and FPR.","CoMET-Agent wraps the same MLLM backbone in four steps. (1) A Planner agent picks a hyperparameter profile and trims the video if the query names a time window. (2) The video is split into coarse event nodes by change-point detection on DINOv2 frame-similarity features. A Filter agent drops irrelevant nodes, and the remaining nodes are split into action nodes by change-point detection on RAFT optical-flow similarity. Edges between nodes are weighted with Qwen3-VL-Embedding-8B. (3) A Verifier agent goes through the candidate action nodes, adds neighbours whose edge weight exceeds a threshold to give more context, checks the query conditions, and writes verified nodes to a Global Memory Bank (summary, timestamps, confidence) with deduplication. (4) An Aggregator agent turns the memory bank into the final intervals and count. Nothing is trained."],"claim":"On CoMET-Bench, CoMET-Agent with a Qwen3-VL-8B backbone raises grounding F1@0.5 from 3.4% (Qwen3-VL-8B single-pass) to 8.8% and Rejection-F1 from 52.1% to 57.4%. With a GPT-5 backbone it raises F1@0.5 from 10.1% to 16.2%.","source":"Table 3, page 9 (component ablation in Table 8, page 31)","metric":"F1@0.5 on positive queries (tIoU threshold 0.5, higher is better), plus Rejection-F1 on negative/positive coverage (higher is better). Single run, no error bars. The authors say open-source inference uses deterministic greedy decoding.","comparison":"CoMET-Agent vs. the same backbone used single-pass with 128 uniform frames (Qwen3-VL-8B; also GPT-5 and Gemini 3 Flash). The strongest non-ours baseline overall is Gemini 3 Flash single-pass (F1@0.5 14.6%), and the best other agent is VideoARM (13.0%).","falsification":"A faithful reimplementation of CoMET-Agent on Qwen3-VL-8B, evaluated with the released CoMET-Bench toolkit, does not clearly beat single-pass Qwen3-VL-8B on F1@0.5 (for example, a gain under about 2 points, or below 6% absolute), or the GPT-5 version does not exceed GPT-5 single-pass F1@0.5 of 10.1%.","caption":"Table 3: Main results on CoMET-Bench across three evaluation aspects. Counting: MAE (↓, mean absolute error), OBO (↑, off-by-one accuracy), Pearson (↑, Pearson correlation). Grounding: mIoU (↑), R@0.5 (↑, Recall at tIoU 0.5), F1@0.5 (↑). Negative: Rej.-F1 (↑, harmonic mean of negative-rejection rate and positive-coverage rate), FPR (↓, fraction of negative queries with non-empty prediction). Bold marks the best score per metric; underline marks the second best.","image":"img/2606.15320.png","kind":"table","image_w":1125,"image_h":872,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.15320","title":"Conditional Multi-Event Temporal Grounding in Long-Form Video","authors":"Yuanhao Zou; Arthad Kulkarni; Lucas Toñanez; Lincoln Spencer; Guangyu Sun; Tianxingjian Ding; Andong Deng; Yi Li; Shuangjun Liu; Yuan Li; Dashan Gao; Ning Bi; TAOTAO JING; Shuai Zhang; Chen Chen","openreview_id":"i6iM7B1Koa","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Video temporal grounding asks a model to find the time interval in a video that matches a text query. Real uses such as video editing or sports analysis often need every occurrence of an event that meets temporal conditions (after X, while Y, within a time window) and spatial conditions (a specific person, a motion), along with an exact count, and sometimes the correct answer is that no such event exists. Existing benchmarks handle single-moment grounding, unconditioned counting, or multiple-choice QA separately. The paper introduces CoMET-Bench (2,789 queries over 600 long videos averaging 33.8 minutes, 26.6% negative queries) and a training-free agent, CoMET-Agent, for this task.","method":["Baselines are video MLLMs run single-pass with a shared evaluation prompt and 128 uniformly sampled frames per video. They output a set of [start, end] intervals per query, or an empty set. Grounding is scored on positive queries with mIoU, Recall@0.5 and F1@0.5 (intervals matched at tIoU 0.5). Counting is scored with MAE, off-by-one accuracy and Pearson correlation. Negative queries are scored with Rejection-F1 (the harmonic mean of the rejection rate on negatives and non-empty coverage on positives) and FPR.","CoMET-Agent wraps the same MLLM backbone in four steps. (1) A Planner agent picks a hyperparameter profile and trims the video if the query names a time window. (2) The video is split into coarse event nodes by change-point detection on DINOv2 frame-similarity features. A Filter agent drops irrelevant nodes, and the remaining nodes are split into action nodes by change-point detection on RAFT optical-flow similarity. Edges between nodes are weighted with Qwen3-VL-Embedding-8B. (3) A Verifier agent goes through the candidate action nodes, adds neighbours whose edge weight exceeds a threshold to give more context, checks the query conditions, and writes verified nodes to a Global Memory Bank (summary, timestamps, confidence) with deduplication. (4) An Aggregator agent turns the memory bank into the final intervals and count. Nothing is trained."],"claim":"On CoMET-Bench, CoMET-Agent with a Qwen3-VL-8B backbone raises grounding F1@0.5 from 3.4% (Qwen3-VL-8B single-pass) to 8.8% and Rejection-F1 from 52.1% to 57.4%. With a GPT-5 backbone it raises F1@0.5 from 10.1% to 16.2%.","source":"Table 3, page 9 (component ablation in Table 8, page 31)","metric":"F1@0.5 on positive queries (tIoU threshold 0.5, higher is better), plus Rejection-F1 on negative/positive coverage (higher is better). Single run, no error bars. The authors say open-source inference uses deterministic greedy decoding.","comparison":"CoMET-Agent vs. the same backbone used single-pass with 128 uniform frames (Qwen3-VL-8B; also GPT-5 and Gemini 3 Flash). The strongest non-ours baseline overall is Gemini 3 Flash single-pass (F1@0.5 14.6%), and the best other agent is VideoARM (13.0%).","falsification":"A faithful reimplementation of CoMET-Agent on Qwen3-VL-8B, evaluated with the released CoMET-Bench toolkit, does not clearly beat single-pass Qwen3-VL-8B on F1@0.5 (for example, a gain under about 2 points, or below 6% absolute), or the GPT-5 version does not exceed GPT-5 single-pass F1@0.5 of 10.1%.","caption":"Table 3: Main results on CoMET-Bench across three evaluation aspects. Counting: MAE (↓, mean absolute error), OBO (↑, off-by-one accuracy), Pearson (↑, Pearson correlation). Grounding: mIoU (↑), R@0.5 (↑, Recall at tIoU 0.5), F1@0.5 (↑). Negative: Rej.-F1 (↑, harmonic mean of negative-rejection rate and positive-coverage rate), FPR (↓, fraction of negative queries with non-empty prediction). Bold marks the best score per metric; underline marks the second best.","image":"img/2606.15320.png","kind":"table","image_w":1125,"image_h":872,"pwc":true}
|
site/p/2606.15749.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.15749","title":"OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning","authors":"Maonan Wang; Zhengyan Huang; Kemou Jiang; Yuhang Fu; Jiayue Zhu; Yuxin Cai; Xingchen Zou; Qiaosheng Zhang; Yi Yu; Ding Wang; XI CHEN; Ben Chen; Yuxuan Liang; Zhiyong Cui; Man On Pun; Yirong Chen","openreview_id":"5fv0JAYAwL","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Roadside traffic cameras record lane usage, queues, special vehicles, incidents and signal phases, and multimodal LLMs are increasingly proposed as a natural-language interface to such footage. Existing traffic VQA benchmarks mostly test passive recognition on single images or videos from real cameras, which offer no control over intersection layout, viewpoints or signal phases and are expensive to annotate with lane-level structure. As a result, it is unclear whether MLLMs can reason about intersection topology, align views with a bird's-eye-view (BEV) map, or track how traffic evolves over time.","method":["OmniTraffic rebuilds 12 real intersections as controllable 3D simulation scenes and renders synchronized multi-view, multi-timestep images together with structured metadata (lane topology, vehicle-to-lane assignment, signal phases, events). Rule-based templates turn the metadata into VQA pairs, which are then converted to four-option multiple-choice questions with distractors drawn from the same scene. Real surveillance footage from South Korea and Tianjin adds human-annotated questions for perception and temporal tasks. A 3,204-item human-verified test set covers 11 categories at three levels: L1 perception, L2 spatio-temporal reasoning (including View-BEV Mapping, which asks a model to match directional camera views to a BEV layout), and L3 signal-phase decision support.","Eleven MLLMs (e.g. Gemini-3-Pro-Thinking, GPT-5.2, Grok-4, Claude-Sonnet-4.5-Thinking, Qwen3-VL-235B-A22B-Thinking) are evaluated zero-shot. They use greedy decoding (temperature 0, at most 500 output tokens), the prompt template in Fig. 22, and a regex extraction of the A/B/C/D label, with up to three retries when no label can be parsed; responses that are still unparseable are scored as wrong. The metric is exact-match accuracy. Ten non-expert humans answer the same questions on a web platform without external aids."],"claim":"On the 3,204-item OmniTraffic test set, all 11 evaluated MLLMs score between 9.2% and 39.2% on View-BEV Mapping, against 100.0% for non-expert humans, and the best model overall (Gemini-3-Pro) reaches 72.4% average accuracy, against 92.9% for humans.","source":"Table 3, page 8","metric":"Exact-match multiple-choice accuracy (%), higher is better; 4 options (chance about 25%); single run with greedy decoding, no error bars or seeds reported","comparison":"11 proprietary and open-weight MLLMs vs. a human baseline of 10 non-expert evaluators, per task category and overall","falsification":"A re-run of any of the listed models (or a comparable frontier MLLM) with the paper's protocol scores well above about 40% on View-BEV Mapping (e.g. above 60%), or the best model's overall accuracy comes within a few points of 92.9%. A check of the View-BEV items finding that they have wrong or ambiguous answers would also undermine the claim.","caption":"Table 3: Performance comparison on the OmniTraffic benchmark (%). *, †, and ‡ denote the first, second, and third best model results.","image":"img/2606.15749.png","kind":"table","image_w":1330,"image_h":672,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.15749","title":"OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning","authors":"Maonan Wang; Zhengyan Huang; Kemou Jiang; Yuhang Fu; Jiayue Zhu; Yuxin Cai; Xingchen Zou; Qiaosheng Zhang; Yi Yu; Ding Wang; XI CHEN; Ben Chen; Yuxuan Liang; Zhiyong Cui; Man On Pun; Yirong Chen","openreview_id":"5fv0JAYAwL","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Roadside traffic cameras record lane usage, queues, special vehicles, incidents and signal phases, and multimodal LLMs are increasingly proposed as a natural-language interface to such footage. Existing traffic VQA benchmarks mostly test passive recognition on single images or videos from real cameras, which offer no control over intersection layout, viewpoints or signal phases and are expensive to annotate with lane-level structure. As a result, it is unclear whether MLLMs can reason about intersection topology, align views with a bird's-eye-view (BEV) map, or track how traffic evolves over time.","method":["OmniTraffic rebuilds 12 real intersections as controllable 3D simulation scenes and renders synchronized multi-view, multi-timestep images together with structured metadata (lane topology, vehicle-to-lane assignment, signal phases, events). Rule-based templates turn the metadata into VQA pairs, which are then converted to four-option multiple-choice questions with distractors drawn from the same scene. Real surveillance footage from South Korea and Tianjin adds human-annotated questions for perception and temporal tasks. A 3,204-item human-verified test set covers 11 categories at three levels: L1 perception, L2 spatio-temporal reasoning (including View-BEV Mapping, which asks a model to match directional camera views to a BEV layout), and L3 signal-phase decision support.","Eleven MLLMs (e.g. Gemini-3-Pro-Thinking, GPT-5.2, Grok-4, Claude-Sonnet-4.5-Thinking, Qwen3-VL-235B-A22B-Thinking) are evaluated zero-shot. They use greedy decoding (temperature 0, at most 500 output tokens), the prompt template in Fig. 22, and a regex extraction of the A/B/C/D label, with up to three retries when no label can be parsed; responses that are still unparseable are scored as wrong. The metric is exact-match accuracy. Ten non-expert humans answer the same questions on a web platform without external aids."],"claim":"On the 3,204-item OmniTraffic test set, all 11 evaluated MLLMs score between 9.2% and 39.2% on View-BEV Mapping, against 100.0% for non-expert humans, and the best model overall (Gemini-3-Pro) reaches 72.4% average accuracy, against 92.9% for humans.","source":"Table 3, page 8","metric":"Exact-match multiple-choice accuracy (%), higher is better; 4 options (chance about 25%); single run with greedy decoding, no error bars or seeds reported","comparison":"11 proprietary and open-weight MLLMs vs. a human baseline of 10 non-expert evaluators, per task category and overall","falsification":"A re-run of any of the listed models (or a comparable frontier MLLM) with the paper's protocol scores well above about 40% on View-BEV Mapping (e.g. above 60%), or the best model's overall accuracy comes within a few points of 92.9%. A check of the View-BEV items finding that they have wrong or ambiguous answers would also undermine the claim.","caption":"Table 3: Performance comparison on the OmniTraffic benchmark (%). *, †, and ‡ denote the first, second, and third best model results.","image":"img/2606.15749.png","kind":"table","image_w":1330,"image_h":672,"pwc":true}
|
site/p/2606.17541.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.17541","title":"Offline Preference-Based Trajectory Evaluation","authors":"Fernando Diaz","openreview_id":"M1eNDOOTeU","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Agent benchmarks usually score each trajectory by binary terminal success, and compare systems by success rate. This throws away partial progress and how quickly the agent got there, so many instance-level comparisons between two systems are ties. Ties shrink the effective sample size, make it harder to separate systems, and may contribute to apparent benchmark saturation. The paper proposes comparing pairs of trajectories directly through temporal preferences over their return-versus-time profiles. It also measures how much signal this recovers.","method":["A trajectory is a nondecreasing normalized return curve f(t) in [0, 1], and g(R) is the first step at which return R is reached (infinity if never). Metric baselines are success rate SR = 1[g(1) < inf], partial return PR = max_t f(t), and reference-free SPL = success / trajectory length. Each metric gives a pairwise preference Δ = μ(f) − μ(f′). The proposed preferences are computed directly on a pair of trajectories. Lexicographic Return (LR) is sgn[g′(R*) − g(R*)] at the highest return level R* where the two times-to-return differ. Return-Paired Preference (RPP) sums, over the segments of the sorted union of both trajectories' return levels (plus 0 and 1), the segment width times sgn[g′(R̂k) − g(R̂k)], with ∞ − ∞ = 0. Interval-Paired Preference (IPP) does the same with the time increments between consecutive levels.","Tie rate is the percentage of instance-level comparisons (pairs of systems on the same task instance) with Δ = 0. It is averaged over tasks within each benchmark and then over benchmarks. The benchmarks are AgentBoard (6 tasks, 12 models, with subgoal progress scores), OpenHands-Index (4 tasks, success and step count only), TheAgentCompany (partial terminal score and step count), TALES (Jericho and ScienceWorld, about 50 models), and a self-generated sub-goal RL set (Taxi, DoorKey, FourRooms). Tasks with fewer than 30 instances are removed."],"claim":"On AgentBoard trajectories (6 tasks, 12 models), the trajectory preferences LR, RPP and IPP tie on 33.3%, 34.9% and 36.1% of instance-level comparisons. Success rate ties on 78.9% and partial return, the strongest scalar baseline, on 43.5%. Averaged over all five benchmarks, ties fall from 74.9% (SR) and 49.7% (PR) to 33.9-35.1%.","source":"Table 8, page 19 (appendix); aggregate in Table 2d, page 8","metric":"Tie rate: % of instance-level system-pair comparisons with zero preference, lower is better. Computed exactly with no error bars, averaged over tasks within a benchmark.","comparison":"LR, RPP, IPP (trajectory preferences) vs. SR, PR, SPL (scalar metrics turned into preferences), on the same logged trajectories","falsification":"Recomputing tie rates on the public AgentBoard logs, RPP (or LR) ties on about as many comparisons as PR (e.g. a gap under 5 percentage points), or SR's tie rate is far below about 79%.","caption":"Table 8: Tie rate (% of instance comparisons with zero difference), averaged within benchmark.","image":"img/2606.17541.png","kind":"table","image_w":1119,"image_h":291,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.17541","title":"Offline Preference-Based Trajectory Evaluation","authors":"Fernando Diaz","openreview_id":"M1eNDOOTeU","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Agent benchmarks usually score each trajectory by binary terminal success, and compare systems by success rate. This throws away partial progress and how quickly the agent got there, so many instance-level comparisons between two systems are ties. Ties shrink the effective sample size, make it harder to separate systems, and may contribute to apparent benchmark saturation. The paper proposes comparing pairs of trajectories directly through temporal preferences over their return-versus-time profiles. It also measures how much signal this recovers.","method":["A trajectory is a nondecreasing normalized return curve f(t) in [0, 1], and g(R) is the first step at which return R is reached (infinity if never). Metric baselines are success rate SR = 1[g(1) < inf], partial return PR = max_t f(t), and reference-free SPL = success / trajectory length. Each metric gives a pairwise preference Δ = μ(f) − μ(f′). The proposed preferences are computed directly on a pair of trajectories. Lexicographic Return (LR) is sgn[g′(R*) − g(R*)] at the highest return level R* where the two times-to-return differ. Return-Paired Preference (RPP) sums, over the segments of the sorted union of both trajectories' return levels (plus 0 and 1), the segment width times sgn[g′(R̂k) − g(R̂k)], with ∞ − ∞ = 0. Interval-Paired Preference (IPP) does the same with the time increments between consecutive levels.","Tie rate is the percentage of instance-level comparisons (pairs of systems on the same task instance) with Δ = 0. It is averaged over tasks within each benchmark and then over benchmarks. The benchmarks are AgentBoard (6 tasks, 12 models, with subgoal progress scores), OpenHands-Index (4 tasks, success and step count only), TheAgentCompany (partial terminal score and step count), TALES (Jericho and ScienceWorld, about 50 models), and a self-generated sub-goal RL set (Taxi, DoorKey, FourRooms). Tasks with fewer than 30 instances are removed."],"claim":"On AgentBoard trajectories (6 tasks, 12 models), the trajectory preferences LR, RPP and IPP tie on 33.3%, 34.9% and 36.1% of instance-level comparisons. Success rate ties on 78.9% and partial return, the strongest scalar baseline, on 43.5%. Averaged over all five benchmarks, ties fall from 74.9% (SR) and 49.7% (PR) to 33.9-35.1%.","source":"Table 8, page 19 (appendix); aggregate in Table 2d, page 8","metric":"Tie rate: % of instance-level system-pair comparisons with zero preference, lower is better. Computed exactly with no error bars, averaged over tasks within a benchmark.","comparison":"LR, RPP, IPP (trajectory preferences) vs. SR, PR, SPL (scalar metrics turned into preferences), on the same logged trajectories","falsification":"Recomputing tie rates on the public AgentBoard logs, RPP (or LR) ties on about as many comparisons as PR (e.g. a gap under 5 percentage points), or SR's tie rate is far below about 79%.","caption":"Table 8: Tie rate (% of instance comparisons with zero difference), averaged within benchmark.","image":"img/2606.17541.png","kind":"table","image_w":1119,"image_h":291,"pwc":true}
|
site/p/2606.17833.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.17833","title":"HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning","authors":"Taowen Wang; zikang xie; Bin Yang; Yunheng Wang; Zizhao Yuan; Yuetong Fang; Yixiao Feng; Yichi Wang; Xingyu Chen; Haodong Chen; Qiwei Wu; Weisheng Xu; Lihan Chen; Lusong Li; zeng zecui; Renjing Xu","openreview_id":"r3Vj07V062","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Humanoid robots doing whole-body interaction (kicking a ball, sitting on a sofa, stepping between desks) are often controlled hierarchically. A high-level visuomotor policy predicts an intermediate whole-body action, and a low-level general motion tracker (GMT), an RL-trained controller, turns it into stable joint targets. If the intermediate action is a shared interface, a policy trained with one tracker could in principle be deployed with another. Existing benchmarks do not test whether this works. HumanoidArena is a simulated benchmark of 7 leg-critical tasks built to measure that policy-tracker interface directly.","method":["In Isaac Lab, a Unitree G1 humanoid is teleoperated through VR, from an egocentric view, to collect 100 successful demonstrations per task per GMT backend (TWIST2 and SONIC). Both backends receive the same 40D canonical action: planar root motion, root height, root orientation, 29 joint targets and binary hand open/close. A GMT-specific adapter maps this action into each tracker's input format. High-level policies (ACT, Diffusion Policy, Flow Matching, pi0.5; LeRobot implementations, 100k gradient steps each) take the egocentric RGB image, a 64D proprioceptive state and the task instruction, and predict the 40D action. A separate policy is trained per task and per GMT.","In-GMT evaluation deploys a policy with the tracker that produced its training data. Cross-GMT evaluation deploys the same trained policy with the other tracker. Every configuration is run with 3 seeds x 20 rollouts = 60 episodes on identical held-out initializations. The paper reports the absolute drop Delta_cross = SR_in - SR_cross and the retention Ret = SR_cross / SR_in."],"claim":"Diffusion Policy trained on TWIST2 demonstrations and deployed with the SONIC tracker loses most of its success rate on all 7 tasks: Football 46.7% -> ~21.7%, DoubleDesk 10.0% -> 0%, P&PBox 50.0% -> 0%, OpenDoor 68.3% -> 0%, SitSofa 73.3% -> 0%, Boxing 51.7% -> ~1.7%, VisNavi 30.0% -> ~16.7% (retention 0-56%). Across all 4 policies, 22 of the 28 TWIST2->SONIC entries reach 0% retention.","source":"Table 2, page 9 (in-GMT success rates with seed std in Table 1, page 8)","metric":"Task success rate over 60 rollouts (3 seeds x 20), higher is better. The table reports absolute drop Delta_cross (points) and retention Ret = SR_cross/SR_in. In-GMT SRs come with mean +/- std over 3 seeds (Table 1). Cross-GMT entries have no error bars.","comparison":"The same trained high-level policy deployed with its training GMT (in-GMT) vs. with the other GMT (cross-GMT). The headline case is DP trained on TWIST2 data and deployed with TWIST2 vs. with SONIC.","falsification":"Using the released TWIST2 demonstrations and the paper's evaluation protocol, a Diffusion Policy deployed with SONIC keeps SR within about one in-GMT seed std (<=12 points) of its TWIST2 in-GMT SR on most of the 7 tasks. Equivalently, retention is above roughly 80% on most tasks rather than near 0%.","caption":"Table 2: Cross-GMT evaluation. T->S denotes policies trained on TWIST2 demonstrations and deployed with the SONIC backend, while S->T denotes the reverse direction. Each policy entry reports absolute cross-GMT drop Delta_cross (lower is better) / relative transfer retention Ret (higher is better), computed with respect to the corresponding matched in-GMT performance. AFR (lower is better) denotes the task-level average fall rate under each cross-GMT deployment direction, measuring tracker-induced stability failures.","image":"img/2606.17833.png","kind":"table","image_w":1144,"image_h":394,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.17833","title":"HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning","authors":"Taowen Wang; zikang xie; Bin Yang; Yunheng Wang; Zizhao Yuan; Yuetong Fang; Yixiao Feng; Yichi Wang; Xingyu Chen; Haodong Chen; Qiwei Wu; Weisheng Xu; Lihan Chen; Lusong Li; zeng zecui; Renjing Xu","openreview_id":"r3Vj07V062","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Humanoid robots doing whole-body interaction (kicking a ball, sitting on a sofa, stepping between desks) are often controlled hierarchically. A high-level visuomotor policy predicts an intermediate whole-body action, and a low-level general motion tracker (GMT), an RL-trained controller, turns it into stable joint targets. If the intermediate action is a shared interface, a policy trained with one tracker could in principle be deployed with another. Existing benchmarks do not test whether this works. HumanoidArena is a simulated benchmark of 7 leg-critical tasks built to measure that policy-tracker interface directly.","method":["In Isaac Lab, a Unitree G1 humanoid is teleoperated through VR, from an egocentric view, to collect 100 successful demonstrations per task per GMT backend (TWIST2 and SONIC). Both backends receive the same 40D canonical action: planar root motion, root height, root orientation, 29 joint targets and binary hand open/close. A GMT-specific adapter maps this action into each tracker's input format. High-level policies (ACT, Diffusion Policy, Flow Matching, pi0.5; LeRobot implementations, 100k gradient steps each) take the egocentric RGB image, a 64D proprioceptive state and the task instruction, and predict the 40D action. A separate policy is trained per task and per GMT.","In-GMT evaluation deploys a policy with the tracker that produced its training data. Cross-GMT evaluation deploys the same trained policy with the other tracker. Every configuration is run with 3 seeds x 20 rollouts = 60 episodes on identical held-out initializations. The paper reports the absolute drop Delta_cross = SR_in - SR_cross and the retention Ret = SR_cross / SR_in."],"claim":"Diffusion Policy trained on TWIST2 demonstrations and deployed with the SONIC tracker loses most of its success rate on all 7 tasks: Football 46.7% -> ~21.7%, DoubleDesk 10.0% -> 0%, P&PBox 50.0% -> 0%, OpenDoor 68.3% -> 0%, SitSofa 73.3% -> 0%, Boxing 51.7% -> ~1.7%, VisNavi 30.0% -> ~16.7% (retention 0-56%). Across all 4 policies, 22 of the 28 TWIST2->SONIC entries reach 0% retention.","source":"Table 2, page 9 (in-GMT success rates with seed std in Table 1, page 8)","metric":"Task success rate over 60 rollouts (3 seeds x 20), higher is better. The table reports absolute drop Delta_cross (points) and retention Ret = SR_cross/SR_in. In-GMT SRs come with mean +/- std over 3 seeds (Table 1). Cross-GMT entries have no error bars.","comparison":"The same trained high-level policy deployed with its training GMT (in-GMT) vs. with the other GMT (cross-GMT). The headline case is DP trained on TWIST2 data and deployed with TWIST2 vs. with SONIC.","falsification":"Using the released TWIST2 demonstrations and the paper's evaluation protocol, a Diffusion Policy deployed with SONIC keeps SR within about one in-GMT seed std (<=12 points) of its TWIST2 in-GMT SR on most of the 7 tasks. Equivalently, retention is above roughly 80% on most tasks rather than near 0%.","caption":"Table 2: Cross-GMT evaluation. T->S denotes policies trained on TWIST2 demonstrations and deployed with the SONIC backend, while S->T denotes the reverse direction. Each policy entry reports absolute cross-GMT drop Delta_cross (lower is better) / relative transfer retention Ret (higher is better), computed with respect to the corresponding matched in-GMT performance. AFR (lower is better) denotes the task-level average fall rate under each cross-GMT deployment direction, measuring tracker-induced stability failures.","image":"img/2606.17833.png","kind":"table","image_w":1144,"image_h":394,"pwc":true}
|
site/p/2606.21116.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.21116","title":"ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading","authors":"Jeff Brown; Tim Farkas; Gleb Razgar; Edward Boyden","openreview_id":"ceOs8fK1Ev","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Connectomics reconstructs neurons and their synapses from 3D electron-microscopy volumes. Automatic segmentation leaves two kinds of errors: split errors, where one neuron is broken into pieces, and merge errors, where several neurons are fused. Correcting them by hand ('proofreading') is the main bottleneck in building whole-brain wiring maps. Earlier automated proofreading systems were usually trained and evaluated on one species and one error type, and earlier benchmarks were too small to train on. This paper releases a large multi-species benchmark built from expert proofreading edits and trains a single vision model on it.","method":["The dataset takes non-reverted expert proofreading edits from four public connectomes (MICrONS mouse, FlyWire fly, H01 human, Fish1 zebrafish). Positives are real merge edits (fixing split errors) and real split edits (fixing merge errors). Negatives are synapse partners and spatially adjacent segments (for split-error correction) and skeleton junctions on proofread neurons (for merge-error identification). Each sample is rendered as 3 orthographic mesh-geometry views (7 channels: normals, depth, two segment masks, silhouette) and 4 EM slice views (3 channels: EM plus two segment masks). Train/val/test splits are assigned by (80 µm)^3 spatial cubes (75/12.5/12.5).","The model is one ImageNet-initialized ViT-B/16 (224x224) with two patch embeddings, a 7-channel one for geometry and a 3-channel one for EM, feeding a shared encoder. Classification heads on the CLS token cover split-error correction, merge-error identification and an auxiliary synapse task, and a CNN decoder predicts a 2-channel mask for merge-error correction. It is trained jointly across species, modalities and tasks with masked cross-entropy (label smoothing 0.1) plus a permutation-invariant CE+Dice mask loss, using AdamW with a cosine schedule for 20 epochs. At test time, view probabilities are averaged within each operation and thresholded at 0.5. Table 1 compares the full model with geometry-only and EM-only variants. The text says modalities and channels can be ablated independently at inference time, and channel dropout is used in training."],"claim":"On the held-out spatial test split, the jointly trained mesh+EM ViT-B reaches 93.0% [92.7, 93.2] per-operation balanced accuracy on merge-error identification, versus 90.6% [90.4, 90.9] for geometry-only and 91.1% [90.9, 91.4] for EM-only. On split-error correction it reaches 97.0% [96.9, 97.2] versus 96.2% and 94.2%.","source":"Table 1, page 6","metric":"Per-sample (per-operation, views averaged) balanced accuracy at threshold 0.5, higher is better; brackets are 95% cluster-bootstrap CIs over operations (one training run, no seed variance reported)","comparison":"Full ViT-B (geometry + EM views) vs. the same ViT-B with only geometry views and with only EM views; human expert reference (93.0% split, 84.1% merge) is also reported","falsification":"On the released test split, the full mesh+EM ViT-B's merge-error bAcc is not above both single-modality variants by more than their CIs (e.g. a gap under ~1 point), or its split-error bAcc does not exceed the geometry-only variant.","caption":"Table 1: Peak performance and ablations on the held-out test set. Per-sample (one prediction per operation; views aggregated). bAcc is reported at threshold 0.5 alongside mIoU (mask GT is rendered on geometry views only, so the EM-only row's mIoU is '—'). Brackets are 95% cluster-bootstrap confidence intervals.","image":"img/2606.21116.png","kind":"table","image_w":1144,"image_h":600,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.21116","title":"ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading","authors":"Jeff Brown; Tim Farkas; Gleb Razgar; Edward Boyden","openreview_id":"ceOs8fK1Ev","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Connectomics reconstructs neurons and their synapses from 3D electron-microscopy volumes. Automatic segmentation leaves two kinds of errors: split errors, where one neuron is broken into pieces, and merge errors, where several neurons are fused. Correcting them by hand ('proofreading') is the main bottleneck in building whole-brain wiring maps. Earlier automated proofreading systems were usually trained and evaluated on one species and one error type, and earlier benchmarks were too small to train on. This paper releases a large multi-species benchmark built from expert proofreading edits and trains a single vision model on it.","method":["The dataset takes non-reverted expert proofreading edits from four public connectomes (MICrONS mouse, FlyWire fly, H01 human, Fish1 zebrafish). Positives are real merge edits (fixing split errors) and real split edits (fixing merge errors). Negatives are synapse partners and spatially adjacent segments (for split-error correction) and skeleton junctions on proofread neurons (for merge-error identification). Each sample is rendered as 3 orthographic mesh-geometry views (7 channels: normals, depth, two segment masks, silhouette) and 4 EM slice views (3 channels: EM plus two segment masks). Train/val/test splits are assigned by (80 µm)^3 spatial cubes (75/12.5/12.5).","The model is one ImageNet-initialized ViT-B/16 (224x224) with two patch embeddings, a 7-channel one for geometry and a 3-channel one for EM, feeding a shared encoder. Classification heads on the CLS token cover split-error correction, merge-error identification and an auxiliary synapse task, and a CNN decoder predicts a 2-channel mask for merge-error correction. It is trained jointly across species, modalities and tasks with masked cross-entropy (label smoothing 0.1) plus a permutation-invariant CE+Dice mask loss, using AdamW with a cosine schedule for 20 epochs. At test time, view probabilities are averaged within each operation and thresholded at 0.5. Table 1 compares the full model with geometry-only and EM-only variants. The text says modalities and channels can be ablated independently at inference time, and channel dropout is used in training."],"claim":"On the held-out spatial test split, the jointly trained mesh+EM ViT-B reaches 93.0% [92.7, 93.2] per-operation balanced accuracy on merge-error identification, versus 90.6% [90.4, 90.9] for geometry-only and 91.1% [90.9, 91.4] for EM-only. On split-error correction it reaches 97.0% [96.9, 97.2] versus 96.2% and 94.2%.","source":"Table 1, page 6","metric":"Per-sample (per-operation, views averaged) balanced accuracy at threshold 0.5, higher is better; brackets are 95% cluster-bootstrap CIs over operations (one training run, no seed variance reported)","comparison":"Full ViT-B (geometry + EM views) vs. the same ViT-B with only geometry views and with only EM views; human expert reference (93.0% split, 84.1% merge) is also reported","falsification":"On the released test split, the full mesh+EM ViT-B's merge-error bAcc is not above both single-modality variants by more than their CIs (e.g. a gap under ~1 point), or its split-error bAcc does not exceed the geometry-only variant.","caption":"Table 1: Peak performance and ablations on the held-out test set. Per-sample (one prediction per operation; views aggregated). bAcc is reported at threshold 0.5 alongside mIoU (mask GT is rendered on geometry views only, so the EM-only row's mIoU is '—'). Brackets are 95% cluster-bootstrap confidence intervals.","image":"img/2606.21116.png","kind":"table","image_w":1144,"image_h":600,"pwc":true}
|
site/p/2606.23061.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.23061","title":"MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning","authors":"Weile Guo; Shenghong He; Danying Mo; Chengdong Xu; Xuexun Liu; Chao Yu","openreview_id":"6aIVKV32ea","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Motion instruction generation asks a model to compare a query video of someone exercising with a reference video and produce corrective feedback (e.g. 'keep your torso more upright'). Large multimodal models often produce instructions that sound plausible but do not match the actual kinematic difference: they get the direction backwards, name the wrong joint, or compare mismatched phases of the movement. MotionHalluc is a benchmark of 1540 closed-form questions over 553 Fit3D video pairs (32 fitness actions) that tests these three failure modes, Directional (DH), Attributional (AH) and Temporal (TH) hallucination. The paper also asks whether giving the model explicit kinematic measurements reduces these errors.","method":["Benchmark tasks: DH is binary yes/no on whether a human-annotated instruction is accurate for (query, reference); half the items are 'Reversed', with the videos swapped so the correct answer is No. AH is a 2-way choice between a true instruction and a GPT-5-perturbed one (same joint or different joints). TH is a 3-way choice of which reference clip matches the query clip's action phase. Overall accuracy is the mean of DH, AH and TH accuracies. Videos are sampled at 4 FPS; open-source models use greedy decoding, and API models are averaged over 3 runs.","PPV (Perceive-Parse-Verify) is training-free. (1) Get 3D motion for both videos: 4D-Humans reconstruction plus DTW alignment, or, in the PPV-GT variant, Fit3D ground-truth motion capture (22-joint HumanML3D skeleton). (2) A text-only semantic parser (gemini-3-flash-preview) turns the action description and each candidate instruction into structured calls from a fixed library of 4 frame-localization functions (extremum of a joint angle, joint-axis offset, pelvis height, or bone-to-world-direction angle) and 4 measurement functions (three-joint angle, joint-axis offset, pelvis Y, bone-to-world angle). (3) These are executed on both motion sequences, and the resulting query vs. reference measurements are given as text to the evaluated model along with the original videos and question."],"claim":"Adding PPV kinematic measurements computed from ground-truth motion raises overall MotionHalluc accuracy for all five evaluated LMMs, with paired 95% bootstrap CIs on the gain excluding zero, e.g. LLaVA-OV-1.5-8B 48.37 -> 59.65 (+11.28 [9.07, 13.51]), InternVL3.5-8B +6.58 [4.16, 9.08], Qwen3-VL-8B +4.96 [2.60, 7.29], Gemini-3-Flash +18.03 [15.30, 20.62].","source":"Table 9 (Appendix H.1), page 22; reconstructed-motion version in Table 3, page 7","metric":"Overall accuracy (%) = mean of DH, AH, TH accuracies, higher is better. Brackets are percentile 95% CIs from 10,000 action-stratified cluster bootstrap resamples (video pairs as clusters, same draws for Base and PPV-GT). The LOAO range is the min-max gain over 32 leave-one-action-out evaluations.","comparison":"Each model with PPV-GT measurements injected vs. the same model on the same questions without measurements (Base)","falsification":"Re-running Base and PPV-GT on MotionHalluc for any of the open models (LLaVA-OV-1.5-8B, InternVL3.5-8B, Qwen3-VL-8B) gives an overall-accuracy gain whose 95% cluster-bootstrap CI includes zero, or a gain far below the reported interval (e.g. < +9 pp for LLaVA-OV-1.5-8B, < +4 pp for InternVL3.5-8B).","caption":"Table 9: Overall accuracy under Base and PPV-GT. Accuracy is reported in percent; gains (PPV-GT minus Base) are in percentage points. Brackets denote 95% bootstrap confidence intervals. Gains use unrounded scores. The LOAO range gives the minimum and maximum gains across 32 action exclusions.","image":"img/2606.23061.png","kind":"table","image_w":1133,"image_h":344,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.23061","title":"MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning","authors":"Weile Guo; Shenghong He; Danying Mo; Chengdong Xu; Xuexun Liu; Chao Yu","openreview_id":"6aIVKV32ea","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Motion instruction generation asks a model to compare a query video of someone exercising with a reference video and produce corrective feedback (e.g. 'keep your torso more upright'). Large multimodal models often produce instructions that sound plausible but do not match the actual kinematic difference: they get the direction backwards, name the wrong joint, or compare mismatched phases of the movement. MotionHalluc is a benchmark of 1540 closed-form questions over 553 Fit3D video pairs (32 fitness actions) that tests these three failure modes, Directional (DH), Attributional (AH) and Temporal (TH) hallucination. The paper also asks whether giving the model explicit kinematic measurements reduces these errors.","method":["Benchmark tasks: DH is binary yes/no on whether a human-annotated instruction is accurate for (query, reference); half the items are 'Reversed', with the videos swapped so the correct answer is No. AH is a 2-way choice between a true instruction and a GPT-5-perturbed one (same joint or different joints). TH is a 3-way choice of which reference clip matches the query clip's action phase. Overall accuracy is the mean of DH, AH and TH accuracies. Videos are sampled at 4 FPS; open-source models use greedy decoding, and API models are averaged over 3 runs.","PPV (Perceive-Parse-Verify) is training-free. (1) Get 3D motion for both videos: 4D-Humans reconstruction plus DTW alignment, or, in the PPV-GT variant, Fit3D ground-truth motion capture (22-joint HumanML3D skeleton). (2) A text-only semantic parser (gemini-3-flash-preview) turns the action description and each candidate instruction into structured calls from a fixed library of 4 frame-localization functions (extremum of a joint angle, joint-axis offset, pelvis height, or bone-to-world-direction angle) and 4 measurement functions (three-joint angle, joint-axis offset, pelvis Y, bone-to-world angle). (3) These are executed on both motion sequences, and the resulting query vs. reference measurements are given as text to the evaluated model along with the original videos and question."],"claim":"Adding PPV kinematic measurements computed from ground-truth motion raises overall MotionHalluc accuracy for all five evaluated LMMs, with paired 95% bootstrap CIs on the gain excluding zero, e.g. LLaVA-OV-1.5-8B 48.37 -> 59.65 (+11.28 [9.07, 13.51]), InternVL3.5-8B +6.58 [4.16, 9.08], Qwen3-VL-8B +4.96 [2.60, 7.29], Gemini-3-Flash +18.03 [15.30, 20.62].","source":"Table 9 (Appendix H.1), page 22; reconstructed-motion version in Table 3, page 7","metric":"Overall accuracy (%) = mean of DH, AH, TH accuracies, higher is better. Brackets are percentile 95% CIs from 10,000 action-stratified cluster bootstrap resamples (video pairs as clusters, same draws for Base and PPV-GT). The LOAO range is the min-max gain over 32 leave-one-action-out evaluations.","comparison":"Each model with PPV-GT measurements injected vs. the same model on the same questions without measurements (Base)","falsification":"Re-running Base and PPV-GT on MotionHalluc for any of the open models (LLaVA-OV-1.5-8B, InternVL3.5-8B, Qwen3-VL-8B) gives an overall-accuracy gain whose 95% cluster-bootstrap CI includes zero, or a gain far below the reported interval (e.g. < +9 pp for LLaVA-OV-1.5-8B, < +4 pp for InternVL3.5-8B).","caption":"Table 9: Overall accuracy under Base and PPV-GT. Accuracy is reported in percent; gains (PPV-GT minus Base) are in percentage points. Brackets denote 95% bootstrap confidence intervals. Gains use unrounded scores. The LOAO range gives the minimum and maximum gains across 32 action exclusions.","image":"img/2606.23061.png","kind":"table","image_w":1133,"image_h":344,"pwc":true}
|
site/p/2606.23669.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.23669","title":"GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity in Text-to-Image Street-View Generation","authors":"Kaizhen Tan","openreview_id":"LSACZLYAsF","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Text-to-image models can produce believable street scenes for a prompt like 'a street in New York City', but standard metrics (FID, CLIPScore, LPIPS) measure image quality or text alignment and cannot tell whether an image looks like the requested place or just a generic version of the city. GeoFidelity-Bench tests geographic fidelity at the level of named street blocks: each block (a named OpenStreetMap way, or a segment of one) has a panel of real Mapillary street-view images, so generated images can be compared with the real appearance of that street and with other streets in the same city.","method":["The benchmark has 112 named street blocks in 25 cities, each with a filtered panel of real Mapillary images (7,563 block-image assignments, at least 25 and on average 67.5 per block; four images per block are held out as real queries and the rest form the reference panel). Prompts come in three levels sharing a photorealistic, daytime, clear-weather suffix: L0 'taken in {city}, {country}'; L1 'taken on {street} in the {neighborhood} district of {city}, {country}'; L2 = L1 plus 'near GPS coordinates ({lat}, {lon})'. Six open-weight models (SDXL, SD 3.5 Large, FLUX.1-dev, FLUX.1-schnell, PixArt-Σ, HunyuanDiT) each generate four 1024×1024 images per block and prompt level, with fixed per-model settings (e.g. SDXL 30 steps, guidance 5.0, fp16) and a shared negative prompt (not used for FLUX). L0 images for a block are drawn from a pool of city-only generations for its city.","The main metric, CosSim, encodes every image with DINOv2 ViT-B/14, averages the embeddings within the four-image generated panel and within the block's reference panel, and takes the cosine similarity of the two means. The prompt ablation computes, for each (model, block) pair, CosSim(L1) − CosSim(L0), and reports the mean over the 672 pairs with a 95% interval from 5,000 bootstrap resamples of the matched pairs, plus a city-balanced estimate that averages within each city first."],"claim":"Adding the street and neighborhood names to a city-only prompt (L1 vs. L0) raises DINOv2 panel CosSim to the target block's real reference panel by +0.036 on average over 672 matched (model, block) pairs (bootstrap 95% CI [0.029, 0.042]; city-balanced +0.042 [0.025, 0.060], positive in 21 of 25 cities), and the gain is positive for all six models (e.g. SDXL 0.512 → 0.536, SD 3.5 Large 0.432 → 0.515).","source":"Table 6, page 7 (per-model values from Table 5, page 6)","metric":"CosSim: cosine similarity between mean DINOv2 ViT-B/14 embeddings of the four-image generated panel and the block's real reference panel, higher is better. Brackets are 95% bootstrap intervals; the paper states they use 5,000 resamples of the matched (model, block) pairs, and for the city-balanced column says only that estimates average within each city first and then weight cities equally (no separate resampling unit is specified). Table 5 per-model values have no error bars.","comparison":"L1 prompts (city + street + neighborhood names) vs. L0 prompts (city + country only), same models, same blocks, same reference panels, paired by (model, block)","falsification":"Regenerating L0 and L1 images with the stated settings and scoring them against the released reference panels gives a matched mean CosSim gain whose 95% bootstrap interval includes zero, or (for a single-model check) SDXL's L1 − L0 gain is near zero or negative instead of roughly +0.02.","caption":"Table 6: Prompt-ablation uncertainty checks for CosSim. Matched units are (model, block) pairs; city-balanced units give each city equal weight.","image":"img/2606.23669.png","kind":"table","image_w":1144,"image_h":238,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.23669","title":"GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity in Text-to-Image Street-View Generation","authors":"Kaizhen Tan","openreview_id":"LSACZLYAsF","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Text-to-image models can produce believable street scenes for a prompt like 'a street in New York City', but standard metrics (FID, CLIPScore, LPIPS) measure image quality or text alignment and cannot tell whether an image looks like the requested place or just a generic version of the city. GeoFidelity-Bench tests geographic fidelity at the level of named street blocks: each block (a named OpenStreetMap way, or a segment of one) has a panel of real Mapillary street-view images, so generated images can be compared with the real appearance of that street and with other streets in the same city.","method":["The benchmark has 112 named street blocks in 25 cities, each with a filtered panel of real Mapillary images (7,563 block-image assignments, at least 25 and on average 67.5 per block; four images per block are held out as real queries and the rest form the reference panel). Prompts come in three levels sharing a photorealistic, daytime, clear-weather suffix: L0 'taken in {city}, {country}'; L1 'taken on {street} in the {neighborhood} district of {city}, {country}'; L2 = L1 plus 'near GPS coordinates ({lat}, {lon})'. Six open-weight models (SDXL, SD 3.5 Large, FLUX.1-dev, FLUX.1-schnell, PixArt-Σ, HunyuanDiT) each generate four 1024×1024 images per block and prompt level, with fixed per-model settings (e.g. SDXL 30 steps, guidance 5.0, fp16) and a shared negative prompt (not used for FLUX). L0 images for a block are drawn from a pool of city-only generations for its city.","The main metric, CosSim, encodes every image with DINOv2 ViT-B/14, averages the embeddings within the four-image generated panel and within the block's reference panel, and takes the cosine similarity of the two means. The prompt ablation computes, for each (model, block) pair, CosSim(L1) − CosSim(L0), and reports the mean over the 672 pairs with a 95% interval from 5,000 bootstrap resamples of the matched pairs, plus a city-balanced estimate that averages within each city first."],"claim":"Adding the street and neighborhood names to a city-only prompt (L1 vs. L0) raises DINOv2 panel CosSim to the target block's real reference panel by +0.036 on average over 672 matched (model, block) pairs (bootstrap 95% CI [0.029, 0.042]; city-balanced +0.042 [0.025, 0.060], positive in 21 of 25 cities), and the gain is positive for all six models (e.g. SDXL 0.512 → 0.536, SD 3.5 Large 0.432 → 0.515).","source":"Table 6, page 7 (per-model values from Table 5, page 6)","metric":"CosSim: cosine similarity between mean DINOv2 ViT-B/14 embeddings of the four-image generated panel and the block's real reference panel, higher is better. Brackets are 95% bootstrap intervals; the paper states they use 5,000 resamples of the matched (model, block) pairs, and for the city-balanced column says only that estimates average within each city first and then weight cities equally (no separate resampling unit is specified). Table 5 per-model values have no error bars.","comparison":"L1 prompts (city + street + neighborhood names) vs. L0 prompts (city + country only), same models, same blocks, same reference panels, paired by (model, block)","falsification":"Regenerating L0 and L1 images with the stated settings and scoring them against the released reference panels gives a matched mean CosSim gain whose 95% bootstrap interval includes zero, or (for a single-model check) SDXL's L1 − L0 gain is near zero or negative instead of roughly +0.02.","caption":"Table 6: Prompt-ablation uncertainty checks for CosSim. Matched units are (model, block) pairs; city-balanced units give each city equal weight.","image":"img/2606.23669.png","kind":"table","image_w":1144,"image_h":238,"pwc":true}
|
site/p/2606.24819.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2606.24819","title":"HelpBench: Assessing the Ability of LLMs to Provide Privacy, Safety, and Security Advice","authors":"Sarah Meiklejohn; Sunny Consolvo; Patrick G Kelley; Tara Matthews; Sai Teja Peddinti; Renee Shelby; Lenin Simicich; Kurt Thomas","openreview_id":"ySxqajUUaa","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"People increasingly ask LLMs for help with digital privacy, safety and security (PSS) problems such as recovering a locked account, judging whether a message is a scam, or dealing with harassment and stalking. Bad advice in this domain can cause account compromise, financial loss or physical risk, yet there was no help-seeking benchmark for consumer PSS advice comparable to HealthBench or LegalBench. HelpBench fills this gap with realistic, open-ended questions and expert-written per-question rubrics, so that model advice can be scored automatically and repeatedly.","method":["HelpBench has 450 questions, 50 for each of nine topics (accounts, moderation tools, security tools, privacy tools, compromise, platform actions, scams, harassment, data concerns), sampled from Reddit help-seeking posts (2021-2024), stratified by length, whether existing practices are mentioned, and need type, then rephrased with Gemini 2.5 Flash and manually checked. Each question has an expert rubric of positive and negative factual criteria worth 1, 2, 3 or 5 points, plus shared delivery criteria (tone, relevance, clarity). A response's score is (points for met positive criteria + points for avoided negative criteria) / maximum points.","An auto-rater (the prompt in Appendix E, run on Gemini 2.5 Pro at temperature 0) judges each criterion as met or unmet; its scores correlate with expert scores at Pearson r = 0.85 (0.96 factual, 0.78 delivery). 18 models from 7 families (GPT, Claude, Gemini, Grok, DeepSeek, GLM, Qwen) are queried 5 times per question with default parameters through their chat versions (40,500 responses). Scores are averaged over the 5 responses per question, then over the questions in each category."],"claim":"Averaged over 18 LLMs, HelpBench scores are highest on scam questions (88%) and lowest on moderation-tool questions (77%), and every one of the 18 models scores higher on scams than on moderation tools (gaps from 3 points for Gemini 3.1 Pro, 85% vs. 82%, to 16 points for Qwen 3, 78% vs. 62%).","source":"Table 3, page 24 (summary in Table 1, page 6, and Section 5.1.3)","metric":"HelpBench rubric score in percent (higher is better), auto-rated by Gemini 2.5 Pro, averaged over 5 responses per question and then over the 50 questions per topic. The second line under each cell in Table 3 is the average standard deviation across the 5 responses to a question (spread across responses, not a standard error or CI). The paper reports a Kruskal-Wallis test showing topic differences are significant (p < 0.05) for 17 of 18 models.","comparison":"Scores on the 50 scam questions vs. the 50 moderation-tool questions for the same model, same rubric scoring and same auto-rater","falsification":"Generating 5 responses per question from one or more of the listed models on the scams and moderation-tools questions and scoring them with the released rubrics and auto-rater prompt (Gemini 2.5 Pro, temperature 0), the scam score is not clearly higher than the moderation-tool score (e.g. the gap is under ~3 points averaged over the models tested, or reverses for a model whose reported gap is 7+ points, such as GPT 5.0 at 93% vs. 81%).","caption":"Table 3: The average performance of each model, both overall and for questions within a given category. Performance is averaged first per question—where we scored five separate responses from the same model for the same question—and then per category. Each row below this average shows the average standard deviation across the five responses obtained per question.","image":"img/2606.24819.png","kind":"table","image_w":1458,"image_h":2083,"pwc":
|
|
|
|
| 1 |
+
{"id":"2606.24819","title":"HelpBench: Assessing the Ability of LLMs to Provide Privacy, Safety, and Security Advice","authors":"Sarah Meiklejohn; Sunny Consolvo; Patrick G Kelley; Tara Matthews; Sai Teja Peddinti; Renee Shelby; Lenin Simicich; Kurt Thomas","openreview_id":"ySxqajUUaa","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"People increasingly ask LLMs for help with digital privacy, safety and security (PSS) problems such as recovering a locked account, judging whether a message is a scam, or dealing with harassment and stalking. Bad advice in this domain can cause account compromise, financial loss or physical risk, yet there was no help-seeking benchmark for consumer PSS advice comparable to HealthBench or LegalBench. HelpBench fills this gap with realistic, open-ended questions and expert-written per-question rubrics, so that model advice can be scored automatically and repeatedly.","method":["HelpBench has 450 questions, 50 for each of nine topics (accounts, moderation tools, security tools, privacy tools, compromise, platform actions, scams, harassment, data concerns), sampled from Reddit help-seeking posts (2021-2024), stratified by length, whether existing practices are mentioned, and need type, then rephrased with Gemini 2.5 Flash and manually checked. Each question has an expert rubric of positive and negative factual criteria worth 1, 2, 3 or 5 points, plus shared delivery criteria (tone, relevance, clarity). A response's score is (points for met positive criteria + points for avoided negative criteria) / maximum points.","An auto-rater (the prompt in Appendix E, run on Gemini 2.5 Pro at temperature 0) judges each criterion as met or unmet; its scores correlate with expert scores at Pearson r = 0.85 (0.96 factual, 0.78 delivery). 18 models from 7 families (GPT, Claude, Gemini, Grok, DeepSeek, GLM, Qwen) are queried 5 times per question with default parameters through their chat versions (40,500 responses). Scores are averaged over the 5 responses per question, then over the questions in each category."],"claim":"Averaged over 18 LLMs, HelpBench scores are highest on scam questions (88%) and lowest on moderation-tool questions (77%), and every one of the 18 models scores higher on scams than on moderation tools (gaps from 3 points for Gemini 3.1 Pro, 85% vs. 82%, to 16 points for Qwen 3, 78% vs. 62%).","source":"Table 3, page 24 (summary in Table 1, page 6, and Section 5.1.3)","metric":"HelpBench rubric score in percent (higher is better), auto-rated by Gemini 2.5 Pro, averaged over 5 responses per question and then over the 50 questions per topic. The second line under each cell in Table 3 is the average standard deviation across the 5 responses to a question (spread across responses, not a standard error or CI). The paper reports a Kruskal-Wallis test showing topic differences are significant (p < 0.05) for 17 of 18 models.","comparison":"Scores on the 50 scam questions vs. the 50 moderation-tool questions for the same model, same rubric scoring and same auto-rater","falsification":"Generating 5 responses per question from one or more of the listed models on the scams and moderation-tools questions and scoring them with the released rubrics and auto-rater prompt (Gemini 2.5 Pro, temperature 0), the scam score is not clearly higher than the moderation-tool score (e.g. the gap is under ~3 points averaged over the models tested, or reverses for a model whose reported gap is 7+ points, such as GPT 5.0 at 93% vs. 81%).","caption":"Table 3: The average performance of each model, both overall and for questions within a given category. Performance is averaged first per question—where we scored five separate responses from the same model for the same question—and then per category. Each row below this average shows the average standard deviation across the five responses obtained per question.","image":"img/2606.24819.png","kind":"table","image_w":1458,"image_h":2083,"pwc":true}
|
site/p/2607.02680.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2607.02680","title":"K9-Bench: Evaluating Multimodal LLMs on Canine-Centric Videos","authors":"Khush Attarde; Yusuf Ali; Megha Thukral; Divye Bhutani; Thomas Ploetz; Zsolt Kira","openreview_id":"JleiohZb5K","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Video benchmarks for multimodal LLMs are almost entirely about humans, so it is unclear whether these models can follow animal behaviour, such as a dog's posture, ear and tail cues, and its interactions with people over long clips. This matters for pet-welfare monitoring and home robots. K9-Bench is a video question-answering benchmark built from 907 YouTube videos of domestic dogs, with 4,744 five-way multiple-choice questions in five categories (posture, action sequence, context, cause-effect, interaction), generated by Gemini-2.5-Flash from an expert-built canine behaviour ontology and filtered to remove questions that text-only LLMs can answer.","method":["Questions and correct answers are generated by Gemini-2.5-Flash from the video (and audio) plus a canine-behaviour knowledge base; four distractors are generated from a text narration of the video. A 'deaf-blind' filter removes any question that a majority of three text-only frontier LLMs answer correctly (removing 42.58% of 8,263 QAs), and GLM-4.5-Flash rewrites questions to strip speaker identities and audio references.","Open-source models (Qwen3-VL 4B/8B/32B Instruct and Thinking, InternVL-3.5 8B/14B) receive 32 uniformly sampled frames plus the question and are asked to give a free-form answer (Thinking variants get up to 2,048 response tokens). The answer is embedded with Qwen3-Embedding-8B and compared by cosine similarity to the five options; it counts as correct only if similarity to the ground-truth option exceeds 0.5 and no wrong option has a higher similarity. The metric is accuracy over all 4,744 questions (chance is 20%)."],"claim":"On K9-Bench (4,744 five-way questions, 32 frames, embedding-matched free-form answers), Qwen3-VL Thinking variants score below their Instruct counterparts: 4B 22.4% vs. 27.6%, 8B 25.3% vs. 28.8%, 32B 30.7% vs. 31.4%, and no open-source model exceeds 33.0% (InternVL-3.5-14B).","source":"Table 1, page 7","metric":"Multiple-choice accuracy (%) over all 4,744 questions, higher is better; single run, no error bars or seeds reported","comparison":"Qwen3-VL-{4B, 8B, 32B}-Thinking vs. Qwen3-VL-{4B, 8B, 32B}-Instruct under the same 32-frame input and the same embedding-based scorer; InternVL-3.5-8B/14B as the other open-source references","falsification":"Re-evaluating with the released data and the paper's protocol, Qwen3-VL-8B-Thinking matches or beats Qwen3-VL-8B-Instruct (or 4B-Thinking matches or beats 4B-Instruct), or Qwen3-VL-8B-Instruct accuracy differs from 28.8% by more than a few points, or an open-source model of these sizes scores well above 33%.","caption":"Table 1: MCQ evaluation results on K9-Bench: Closed-source models lead overall, while open-source models show varied strengths, particularly in interaction and context reasoning. * indicates models evaluated with 32 frames. † indicates models evaluated using a 1 FPS setting.","image":"img/2607.02680.png","kind":"table","image_w":1133,"image_h":580,"pwc":
|
|
|
|
| 1 |
+
{"id":"2607.02680","title":"K9-Bench: Evaluating Multimodal LLMs on Canine-Centric Videos","authors":"Khush Attarde; Yusuf Ali; Megha Thukral; Divye Bhutani; Thomas Ploetz; Zsolt Kira","openreview_id":"JleiohZb5K","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Video benchmarks for multimodal LLMs are almost entirely about humans, so it is unclear whether these models can follow animal behaviour, such as a dog's posture, ear and tail cues, and its interactions with people over long clips. This matters for pet-welfare monitoring and home robots. K9-Bench is a video question-answering benchmark built from 907 YouTube videos of domestic dogs, with 4,744 five-way multiple-choice questions in five categories (posture, action sequence, context, cause-effect, interaction), generated by Gemini-2.5-Flash from an expert-built canine behaviour ontology and filtered to remove questions that text-only LLMs can answer.","method":["Questions and correct answers are generated by Gemini-2.5-Flash from the video (and audio) plus a canine-behaviour knowledge base; four distractors are generated from a text narration of the video. A 'deaf-blind' filter removes any question that a majority of three text-only frontier LLMs answer correctly (removing 42.58% of 8,263 QAs), and GLM-4.5-Flash rewrites questions to strip speaker identities and audio references.","Open-source models (Qwen3-VL 4B/8B/32B Instruct and Thinking, InternVL-3.5 8B/14B) receive 32 uniformly sampled frames plus the question and are asked to give a free-form answer (Thinking variants get up to 2,048 response tokens). The answer is embedded with Qwen3-Embedding-8B and compared by cosine similarity to the five options; it counts as correct only if similarity to the ground-truth option exceeds 0.5 and no wrong option has a higher similarity. The metric is accuracy over all 4,744 questions (chance is 20%)."],"claim":"On K9-Bench (4,744 five-way questions, 32 frames, embedding-matched free-form answers), Qwen3-VL Thinking variants score below their Instruct counterparts: 4B 22.4% vs. 27.6%, 8B 25.3% vs. 28.8%, 32B 30.7% vs. 31.4%, and no open-source model exceeds 33.0% (InternVL-3.5-14B).","source":"Table 1, page 7","metric":"Multiple-choice accuracy (%) over all 4,744 questions, higher is better; single run, no error bars or seeds reported","comparison":"Qwen3-VL-{4B, 8B, 32B}-Thinking vs. Qwen3-VL-{4B, 8B, 32B}-Instruct under the same 32-frame input and the same embedding-based scorer; InternVL-3.5-8B/14B as the other open-source references","falsification":"Re-evaluating with the released data and the paper's protocol, Qwen3-VL-8B-Thinking matches or beats Qwen3-VL-8B-Instruct (or 4B-Thinking matches or beats 4B-Instruct), or Qwen3-VL-8B-Instruct accuracy differs from 28.8% by more than a few points, or an open-source model of these sizes scores well above 33%.","caption":"Table 1: MCQ evaluation results on K9-Bench: Closed-source models lead overall, while open-source models show varied strengths, particularly in interaction and context reasoning. * indicates models evaluated with 32 frames. † indicates models evaluated using a 1 FPS setting.","image":"img/2607.02680.png","kind":"table","image_w":1133,"image_h":580,"pwc":true}
|