Datasets:
Update papers and claims (part 3)
Browse files- site/p/2603.08898.json +1 -1
- site/p/2603.13864.json +1 -1
- site/p/2604.11674.json +1 -1
- site/p/2604.11762.json +1 -1
- site/p/2604.22847.json +1 -1
- site/p/2604.25231.json +1 -1
- site/p/2604.27819.json +1 -1
- site/p/2604.28031.json +1 -1
- site/p/2605.01417.json +1 -1
- site/p/2605.01945.json +1 -1
- site/p/2605.02199.json +1 -1
- site/p/2605.03952.json +1 -1
- site/p/2605.07565.json +1 -1
- site/p/2605.07919.json +1 -1
- site/p/2605.08647.json +1 -1
- site/p/2605.08988.json +1 -1
- site/p/2605.10442.json +1 -1
- site/p/2605.12147.json +1 -1
- site/p/2605.12662.json +1 -1
- site/p/2605.14152.json +1 -1
- site/p/2605.14847.json +1 -1
site/p/2603.08898.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2603.08898","title":"Towards Visual Query Segmentation in the Wild","authors":"Bing Fan; Minghao Li; Hanzhi Zhang; Shaohua Dong; NAGA PRUDHVI MAREEDU; Xiaoqiong Liu; Weishi Shi; Yunhe Feng; Yan Huang; Heng Fan","openreview_id":"Gx9LtAgUon","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Visual query localization (VQL) asks a model to find an object in an untrimmed video, given an image crop of that object taken from outside the video. The standard benchmark (Ego4D VQ2D) only asks for a bounding box around the object's last appearance. The paper proposes visual query segmentation (VQS): segment every pixel-level occurrence of the queried object across the whole video. This differs from video object segmentation because the reference comes from outside the video and the target appears only sparsely among distractors. The authors release VQS-4K (4,111 YouTube videos, 222 categories, 1.3M frames at 6 fps, manual masklets for all occurrences) and a baseline method.","method":["VQ-SAM extends SAM 2 (Hiera-B+ encoder). The query image and its mask are encoded into an initial memory. Over K = 2 stages, the current memory is fused into the video-frame features, a spatial-temporal Transformer block is applied, and SAM 2's mask decoder produces 3 candidate masks per frame. Target Feature Generation keeps the top n_t = 2 best masks with IoU score above 0.5, and Distractor Feature Generation keeps the top n_d = 1 alternative mask by divergence times IoU score. An Adaptive Memory Generation module learns softmax weights over the initial memory, target features and distractor features to form the next memory. In the last stage each frame outputs its highest-IoU mask if the occlusion score is above 0, and nothing otherwise. Training: 144 epochs, AdamW, lr 5e-6, 768x768 input, 7-frame clips. At inference the video is split into non-overlapping 7-frame windows.","Baselines are VOS methods (Cutie-b, OASIS, LOMM, SAM 2, SAM2.1++, SAM2Long, SAM 3), given the query image and mask as the reference, and VQL methods (VQLoC, PRVQL, REN), given an added mask head. All except the training-free REN are trained on VQS-4K train (3,289 videos) and evaluated on VQS-4K test (822 videos). Metrics: mean spatio-temporal mask IoU (mSTIoU), temporal IoU (mTIoU), Recovery and Success."],"claim":"On VQS-4K test (822 videos), VQ-SAM reaches 26.0 mSTIoU and 29.6 mTIoU, beating every adapted baseline trained on the same split: the best baseline is SAM2Long at 18.6 mSTIoU / 24.4 mTIoU, and plain SAM 2 fine-tuned on VQS-4K gets 12.9 / 13.0.","source":"Table 3, page 8","metric":"mSTIoU and mTIoU in % (higher is better), plus Recovery and Success; single runs, no error bars","comparison":"VQ-SAM vs. 7 adapted VOS methods and 3 adapted VQL methods, all trained on VQS-4K train except training-free REN","falsification":"A retrained VQ-SAM (or the released checkpoint) gets mSTIoU on VQS-4K test within about 2 points of SAM2Long (i.e. below ~20.6), or a SAM 2 baseline fine-tuned on VQS-4K train with the same recipe gets close to VQ-SAM's 26.0 mSTIoU.","caption":"Table 3: Comparison with current approaches on VQS-4K Tst.","image":"img/2603.08898.png","kind":"table","image_w":1088,"image_h":402,"pwc":
|
|
|
|
| 1 |
+
{"id":"2603.08898","title":"Towards Visual Query Segmentation in the Wild","authors":"Bing Fan; Minghao Li; Hanzhi Zhang; Shaohua Dong; NAGA PRUDHVI MAREEDU; Xiaoqiong Liu; Weishi Shi; Yunhe Feng; Yan Huang; Heng Fan","openreview_id":"Gx9LtAgUon","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Visual query localization (VQL) asks a model to find an object in an untrimmed video, given an image crop of that object taken from outside the video. The standard benchmark (Ego4D VQ2D) only asks for a bounding box around the object's last appearance. The paper proposes visual query segmentation (VQS): segment every pixel-level occurrence of the queried object across the whole video. This differs from video object segmentation because the reference comes from outside the video and the target appears only sparsely among distractors. The authors release VQS-4K (4,111 YouTube videos, 222 categories, 1.3M frames at 6 fps, manual masklets for all occurrences) and a baseline method.","method":["VQ-SAM extends SAM 2 (Hiera-B+ encoder). The query image and its mask are encoded into an initial memory. Over K = 2 stages, the current memory is fused into the video-frame features, a spatial-temporal Transformer block is applied, and SAM 2's mask decoder produces 3 candidate masks per frame. Target Feature Generation keeps the top n_t = 2 best masks with IoU score above 0.5, and Distractor Feature Generation keeps the top n_d = 1 alternative mask by divergence times IoU score. An Adaptive Memory Generation module learns softmax weights over the initial memory, target features and distractor features to form the next memory. In the last stage each frame outputs its highest-IoU mask if the occlusion score is above 0, and nothing otherwise. Training: 144 epochs, AdamW, lr 5e-6, 768x768 input, 7-frame clips. At inference the video is split into non-overlapping 7-frame windows.","Baselines are VOS methods (Cutie-b, OASIS, LOMM, SAM 2, SAM2.1++, SAM2Long, SAM 3), given the query image and mask as the reference, and VQL methods (VQLoC, PRVQL, REN), given an added mask head. All except the training-free REN are trained on VQS-4K train (3,289 videos) and evaluated on VQS-4K test (822 videos). Metrics: mean spatio-temporal mask IoU (mSTIoU), temporal IoU (mTIoU), Recovery and Success."],"claim":"On VQS-4K test (822 videos), VQ-SAM reaches 26.0 mSTIoU and 29.6 mTIoU, beating every adapted baseline trained on the same split: the best baseline is SAM2Long at 18.6 mSTIoU / 24.4 mTIoU, and plain SAM 2 fine-tuned on VQS-4K gets 12.9 / 13.0.","source":"Table 3, page 8","metric":"mSTIoU and mTIoU in % (higher is better), plus Recovery and Success; single runs, no error bars","comparison":"VQ-SAM vs. 7 adapted VOS methods and 3 adapted VQL methods, all trained on VQS-4K train except training-free REN","falsification":"A retrained VQ-SAM (or the released checkpoint) gets mSTIoU on VQS-4K test within about 2 points of SAM2Long (i.e. below ~20.6), or a SAM 2 baseline fine-tuned on VQS-4K train with the same recipe gets close to VQ-SAM's 26.0 mSTIoU.","caption":"Table 3: Comparison with current approaches on VQS-4K Tst.","image":"img/2603.08898.png","kind":"table","image_w":1088,"image_h":402,"pwc":true}
|
site/p/2603.13864.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2603.13864","title":"Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation","authors":"Qian Li; Yunuo Chen; Yuntian Chen","openreview_id":"8Zds0GA3mw","presentation":"poster","track":"main","verdict":"include","background":"Data-poisoning backdoor attacks hide a trigger in a fraction of training images so that a model trained on them predicts an attacker-chosen label whenever the trigger is present. Many 'invisible' triggers are subtle, mostly high-frequency perturbations. In practice datasets are stored and transmitted as lossy-compressed files (JPEG, BPG, learned codecs), and compression tends to remove exactly such high-frequency detail. So attacks that work on raw RGB images may stop working once the poisoned data passes through a compression pipeline. The paper measures this failure and proposes using region-of-interest (ROI) bit allocation in the encoder to keep trigger information alive through a standard decoder.","method":["All images (CIFAR-10, GTSRB, 8-class CelebA, resized to 64x64) are compressed and decompressed with Mbt-2018, a learned image codec that the authors extend with encoder-side ROI support (Spatial Feature Transform layers take a pixel-level mask that sets bitrate allocation weights; the decoder is unchanged). Baseline attacks (FTrojan, WaNet, BppAttack) poison images in RGB with their default settings and poisoning rates (5%, 10%, 20% respectively), and then all images are compressed with a uniform ROI mask (standard compression). The victim trains a classifier (ResNet18 or MobileNetV2; Adam, lr 0.01, decayed 10x every 100 epochs) on the decompressed images, in an all-to-one attack setting.","Universal Attack Reactivation (UAR, the 'ReA' columns) changes only the encoding: each poisoned image x_p is compressed with a frequency mask M_freq = Norm(mean over RGB channels of |IFFT(h_t(FFT(x_p)))|), where h_t keeps frequencies above threshold t, so trigger-carrying high-frequency regions get more bits. To avoid the mask itself becoming a shortcut, each benign image is compressed with the mask of a randomly chosen poisoned image. Metrics are Benign Accuracy (BA) on decompressed clean test images and Attack Success Rate (ASR) on decompressed triggered test images."],"claim":"With Mbt-2018 compression on CIFAR-10 and ResNet18, FTrojan's ASR falls to 12.53% under standard compression, while the same attack with UAR's frequency ROI masks reaches 94.91% ASR, and BA changes only from 92.21% to 91.32%.","source":"Table 1, page 7","metric":"Attack Success Rate (ASR, %, higher = stronger attack) and Benign Accuracy (BA, %, higher is better) on decompressed test images; single numbers, no seeds or error bars reported","comparison":"FTrojan with UAR ROI-mask encoding (FTrojan_ReA) vs. FTrojan with standard uniform-mask encoding, same codec (Mbt-2018), poisoning rate (5%), backbone and training; the same pattern is reported for WaNet and BppAttack, on GTSRB and CelebA, and with MobileNetV2 (and PreAct-ResNet18/SENet in Table 12)","falsification":"On CIFAR-10/ResNet18 with an ROI-enabled Mbt-2018 at the same codec quality setting for both arms, FTrojan_ReA reaches no more than about 50% ASR, or its gain over standard-compressed FTrojan falls well short of the reported ~80 points, or BA drops by more than a few points.","caption":"Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix A.","image":"img/2603.13864.png","kind":"table","image_w":1138,"image_h":427,"pwc":
|
|
|
|
| 1 |
+
{"id":"2603.13864","title":"Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation","authors":"Qian Li; Yunuo Chen; Yuntian Chen","openreview_id":"8Zds0GA3mw","presentation":"poster","track":"main","verdict":"include","background":"Data-poisoning backdoor attacks hide a trigger in a fraction of training images so that a model trained on them predicts an attacker-chosen label whenever the trigger is present. Many 'invisible' triggers are subtle, mostly high-frequency perturbations. In practice datasets are stored and transmitted as lossy-compressed files (JPEG, BPG, learned codecs), and compression tends to remove exactly such high-frequency detail. So attacks that work on raw RGB images may stop working once the poisoned data passes through a compression pipeline. The paper measures this failure and proposes using region-of-interest (ROI) bit allocation in the encoder to keep trigger information alive through a standard decoder.","method":["All images (CIFAR-10, GTSRB, 8-class CelebA, resized to 64x64) are compressed and decompressed with Mbt-2018, a learned image codec that the authors extend with encoder-side ROI support (Spatial Feature Transform layers take a pixel-level mask that sets bitrate allocation weights; the decoder is unchanged). Baseline attacks (FTrojan, WaNet, BppAttack) poison images in RGB with their default settings and poisoning rates (5%, 10%, 20% respectively), and then all images are compressed with a uniform ROI mask (standard compression). The victim trains a classifier (ResNet18 or MobileNetV2; Adam, lr 0.01, decayed 10x every 100 epochs) on the decompressed images, in an all-to-one attack setting.","Universal Attack Reactivation (UAR, the 'ReA' columns) changes only the encoding: each poisoned image x_p is compressed with a frequency mask M_freq = Norm(mean over RGB channels of |IFFT(h_t(FFT(x_p)))|), where h_t keeps frequencies above threshold t, so trigger-carrying high-frequency regions get more bits. To avoid the mask itself becoming a shortcut, each benign image is compressed with the mask of a randomly chosen poisoned image. Metrics are Benign Accuracy (BA) on decompressed clean test images and Attack Success Rate (ASR) on decompressed triggered test images."],"claim":"With Mbt-2018 compression on CIFAR-10 and ResNet18, FTrojan's ASR falls to 12.53% under standard compression, while the same attack with UAR's frequency ROI masks reaches 94.91% ASR, and BA changes only from 92.21% to 91.32%.","source":"Table 1, page 7","metric":"Attack Success Rate (ASR, %, higher = stronger attack) and Benign Accuracy (BA, %, higher is better) on decompressed test images; single numbers, no seeds or error bars reported","comparison":"FTrojan with UAR ROI-mask encoding (FTrojan_ReA) vs. FTrojan with standard uniform-mask encoding, same codec (Mbt-2018), poisoning rate (5%), backbone and training; the same pattern is reported for WaNet and BppAttack, on GTSRB and CelebA, and with MobileNetV2 (and PreAct-ResNet18/SENet in Table 12)","falsification":"On CIFAR-10/ResNet18 with an ROI-enabled Mbt-2018 at the same codec quality setting for both arms, FTrojan_ReA reaches no more than about 50% ASR, or its gain over standard-compressed FTrojan falls well short of the reported ~80 points, or BA drops by more than a few points.","caption":"Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix A.","image":"img/2603.13864.png","kind":"table","image_w":1138,"image_h":427,"pwc":true}
|
site/p/2604.11674.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.11674","title":"AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation","authors":"Mingyang Li; Haofan Xu; Haowen Sun; Xinzhe Chen; Sihua Ren; Liqi Huang; Chenyang Miao; Xinyang Sui; jiawei Ye; Qiongjie Cui; Zeyang Liu; Xingyu Chen; Xuguang Lan","openreview_id":"e25kaYnRu0","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Simulation is a major source of robot manipulation training data, but many everyday tasks depend on affordance: success requires touching a specific functional region of the object (for example, grasping a mug's body so the handle stays free for a rack hook). Existing simulation data generators either rely on manually authored per-object grasp poses, which do not scale to new objects or tasks, or on generic grasp estimators such as AnyGrasp, which rank grasps by stability and ignore which region the downstream action needs. Adding a 2D semantic mask filter after the estimator (as in GenManip) cannot recover good grasps that the estimator never proposed in the right region.","method":["AffordSim, built on NVIDIA Isaac Sim, turns a natural-language task description into a scene (via Qwen-VL-Plus scene and task parsing) and one affordance query per functional contact region. An open-vocabulary 3D affordance model (VoxAfford) predicts a binary per-point mask on the object's canonical point cloud. Stage 1 runs mean-shift clustering on the masked points to get a contact point, estimates the local surface normal by PCA on nearest neighbours, and samples K approach directions on the hemisphere opposite the normal (fixed polar tilt), solving wrist yaw to align with the local geometry. Stage 2 plans a pre-grasp trajectory to each candidate with cuRobo and scores it 1/(joint-space arc length) if collision-free, 0 otherwise; the highest-scoring grasp is executed, followed by cuRobo-planned lift and task segments.","For Table 1, five grasp sources are compared under identical scene and motion-planner settings on all 50 benchmark tasks (10 Easy, 30 Medium affordance-critical, 10 Hard composite), each run 100 times with the manipulated object's initial pose randomized over a 30 cm horizontal window and 60 degree yaw range. Baselines: Manual (hand-tuned per-object grasp rules), AnyGrasp (stability-ranked generic 6-DoF grasps), AnyGrasp+Filter (AnyGrasp candidates kept only if the contact falls inside a 2D SAM2 task-relevant mask, GenManip-style), and Human Aff. (human-annotated affordance ground truth, oracle). The metric is trajectory collection success rate."],"claim":"On the 30 Medium (affordance-critical) tasks of the AffordSim benchmark, AffordSim's affordance-guided grasp selection achieves an average trajectory collection success rate of 79%, versus 45% for AnyGrasp+Filter and 15% for AnyGrasp under identical scene and planner settings (100 rollouts per task).","source":"Table 1, page 7 (full per-task breakdown in Table 6, page 18)","metric":"Trajectory collection success rate (%), higher is better, averaged over the 30 Medium-tier tasks with 100 rollouts each. No error bars or seed spread are reported.","comparison":"AffordSim (VoxAfford-based affordance-guided grasp selection) vs. AnyGrasp+Filter (strongest automatic baseline) and AnyGrasp; Manual and Human Aff. are per-object-annotated references.","falsification":"Under the paper's protocol on the Medium tier, AffordSim's average collection SR does not exceed AnyGrasp+Filter's by a large margin (e.g., the gap is under ~15 percentage points), or AffordSim fails to beat AnyGrasp+Filter on most Medium tasks.","caption":"Table 1: Affordance integration on the AffordSim benchmark. Trajectory collection success rate (%) per task across the five grasp methods. AnyGrasp+Filter denotes AnyGrasp candidates filtered against a 2D SAM2 task-relevant mask, mirroring the GenManip-style affordance-filtering pipeline. Two representative tasks per tier are listed; full breakdown in Appendix E. Tier averages bolded.","image":"img/2604.11674.png","kind":"table","image_w":1133,"image_h":541,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.11674","title":"AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation","authors":"Mingyang Li; Haofan Xu; Haowen Sun; Xinzhe Chen; Sihua Ren; Liqi Huang; Chenyang Miao; Xinyang Sui; jiawei Ye; Qiongjie Cui; Zeyang Liu; Xingyu Chen; Xuguang Lan","openreview_id":"e25kaYnRu0","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Simulation is a major source of robot manipulation training data, but many everyday tasks depend on affordance: success requires touching a specific functional region of the object (for example, grasping a mug's body so the handle stays free for a rack hook). Existing simulation data generators either rely on manually authored per-object grasp poses, which do not scale to new objects or tasks, or on generic grasp estimators such as AnyGrasp, which rank grasps by stability and ignore which region the downstream action needs. Adding a 2D semantic mask filter after the estimator (as in GenManip) cannot recover good grasps that the estimator never proposed in the right region.","method":["AffordSim, built on NVIDIA Isaac Sim, turns a natural-language task description into a scene (via Qwen-VL-Plus scene and task parsing) and one affordance query per functional contact region. An open-vocabulary 3D affordance model (VoxAfford) predicts a binary per-point mask on the object's canonical point cloud. Stage 1 runs mean-shift clustering on the masked points to get a contact point, estimates the local surface normal by PCA on nearest neighbours, and samples K approach directions on the hemisphere opposite the normal (fixed polar tilt), solving wrist yaw to align with the local geometry. Stage 2 plans a pre-grasp trajectory to each candidate with cuRobo and scores it 1/(joint-space arc length) if collision-free, 0 otherwise; the highest-scoring grasp is executed, followed by cuRobo-planned lift and task segments.","For Table 1, five grasp sources are compared under identical scene and motion-planner settings on all 50 benchmark tasks (10 Easy, 30 Medium affordance-critical, 10 Hard composite), each run 100 times with the manipulated object's initial pose randomized over a 30 cm horizontal window and 60 degree yaw range. Baselines: Manual (hand-tuned per-object grasp rules), AnyGrasp (stability-ranked generic 6-DoF grasps), AnyGrasp+Filter (AnyGrasp candidates kept only if the contact falls inside a 2D SAM2 task-relevant mask, GenManip-style), and Human Aff. (human-annotated affordance ground truth, oracle). The metric is trajectory collection success rate."],"claim":"On the 30 Medium (affordance-critical) tasks of the AffordSim benchmark, AffordSim's affordance-guided grasp selection achieves an average trajectory collection success rate of 79%, versus 45% for AnyGrasp+Filter and 15% for AnyGrasp under identical scene and planner settings (100 rollouts per task).","source":"Table 1, page 7 (full per-task breakdown in Table 6, page 18)","metric":"Trajectory collection success rate (%), higher is better, averaged over the 30 Medium-tier tasks with 100 rollouts each. No error bars or seed spread are reported.","comparison":"AffordSim (VoxAfford-based affordance-guided grasp selection) vs. AnyGrasp+Filter (strongest automatic baseline) and AnyGrasp; Manual and Human Aff. are per-object-annotated references.","falsification":"Under the paper's protocol on the Medium tier, AffordSim's average collection SR does not exceed AnyGrasp+Filter's by a large margin (e.g., the gap is under ~15 percentage points), or AffordSim fails to beat AnyGrasp+Filter on most Medium tasks.","caption":"Table 1: Affordance integration on the AffordSim benchmark. Trajectory collection success rate (%) per task across the five grasp methods. AnyGrasp+Filter denotes AnyGrasp candidates filtered against a 2D SAM2 task-relevant mask, mirroring the GenManip-style affordance-filtering pipeline. Two representative tasks per tier are listed; full breakdown in Appendix E. Tier averages bolded.","image":"img/2604.11674.png","kind":"table","image_w":1133,"image_h":541,"pwc":true}
|
site/p/2604.11762.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.11762","title":"MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI","authors":"Paula Arguello; Berk Tinaz; Mohammad Shahab Sepehri; Maryam Soltanolkotabi; Mahdi Soltanolkotabi","openreview_id":"2w0L3gwZS7","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Deep-learning methods for accelerated MRI reconstruction have mostly been trained and evaluated on public raw k-space datasets of the brain and knee (e.g. fastMRI), so little is known about how they behave on the wider range of musculoskeletal (MSK) anatomies, contrasts and coil setups seen in clinical practice. MosaicMRI releases 3,096 fully sampled 1.5 T multi-coil MSK volumes (88,416 slices, 854 patients, ten anatomy classes, seven contrast groups) plus a 54-volume 0.55 T set. It uses accelerated reconstruction to study whether anatomical diversity in the training data helps.","method":["The 1.5 T data are split by patient into train/val/test (2,166 / 465 / 465 volumes; 61,903 / 13,265 / 13,248 slices). Reconstruction is multi-coil at 8x acceleration: random Cartesian masks keep 12.5% of phase-encoding lines, including the central 4% low-frequency band. Test masks are seeded deterministically by filename, and the target is the root-sum-of-squares (RSS) image. Models follow the fastMRI codebase conventions, and the checkpoint with the best validation SSIM is evaluated.","For Table 5, U-Net and ViT image-domain reconstruction models are each trained twice: once on only the target anatomy's training volumes (knee: 243 volumes in the whole dataset; shoulder: 446) and once on the full MosaicMRI training split. Both are evaluated on the same knee or shoulder test volumes, with PSNR and SSIM against the RSS target."],"claim":"At 8x acceleration, a ViT trained on the full mixed-anatomy MosaicMRI training split beats a ViT trained only on the target anatomy by 3.46 dB on knee (37.47 vs. 34.01 dB PSNR; SSIM 0.892 vs. 0.850) and by 5.87 dB on shoulder (41.66 vs. 35.79 dB; SSIM 0.950 vs. 0.912). U-Net shows the same direction with smaller gains (knee 37.94 vs. 37.35, shoulder 42.41 vs. 42.05).","source":"Table 5, page 7","metric":"PSNR (dB) and SSIM against the RSS target on the anatomy's test volumes, higher is better; single runs, no error bars reported","comparison":"Same architecture (ViT; also U-Net) trained on target-anatomy-only data vs. the full MosaicMRI training split, evaluated on knee and shoulder test sets","falsification":"When ViT is retrained under both regimes, the full-data model's PSNR advantage over the anatomy-only model on knee or shoulder is under ~1 dB, or it reverses.","caption":"Table 5: PSNR (dB) / SSIM for reconstruction with mixed-anatomy training across architectures.","image":"img/2604.11762.png","kind":"table","image_w":680,"image_h":255,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.11762","title":"MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI","authors":"Paula Arguello; Berk Tinaz; Mohammad Shahab Sepehri; Maryam Soltanolkotabi; Mahdi Soltanolkotabi","openreview_id":"2w0L3gwZS7","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Deep-learning methods for accelerated MRI reconstruction have mostly been trained and evaluated on public raw k-space datasets of the brain and knee (e.g. fastMRI), so little is known about how they behave on the wider range of musculoskeletal (MSK) anatomies, contrasts and coil setups seen in clinical practice. MosaicMRI releases 3,096 fully sampled 1.5 T multi-coil MSK volumes (88,416 slices, 854 patients, ten anatomy classes, seven contrast groups) plus a 54-volume 0.55 T set. It uses accelerated reconstruction to study whether anatomical diversity in the training data helps.","method":["The 1.5 T data are split by patient into train/val/test (2,166 / 465 / 465 volumes; 61,903 / 13,265 / 13,248 slices). Reconstruction is multi-coil at 8x acceleration: random Cartesian masks keep 12.5% of phase-encoding lines, including the central 4% low-frequency band. Test masks are seeded deterministically by filename, and the target is the root-sum-of-squares (RSS) image. Models follow the fastMRI codebase conventions, and the checkpoint with the best validation SSIM is evaluated.","For Table 5, U-Net and ViT image-domain reconstruction models are each trained twice: once on only the target anatomy's training volumes (knee: 243 volumes in the whole dataset; shoulder: 446) and once on the full MosaicMRI training split. Both are evaluated on the same knee or shoulder test volumes, with PSNR and SSIM against the RSS target."],"claim":"At 8x acceleration, a ViT trained on the full mixed-anatomy MosaicMRI training split beats a ViT trained only on the target anatomy by 3.46 dB on knee (37.47 vs. 34.01 dB PSNR; SSIM 0.892 vs. 0.850) and by 5.87 dB on shoulder (41.66 vs. 35.79 dB; SSIM 0.950 vs. 0.912). U-Net shows the same direction with smaller gains (knee 37.94 vs. 37.35, shoulder 42.41 vs. 42.05).","source":"Table 5, page 7","metric":"PSNR (dB) and SSIM against the RSS target on the anatomy's test volumes, higher is better; single runs, no error bars reported","comparison":"Same architecture (ViT; also U-Net) trained on target-anatomy-only data vs. the full MosaicMRI training split, evaluated on knee and shoulder test sets","falsification":"When ViT is retrained under both regimes, the full-data model's PSNR advantage over the anatomy-only model on knee or shoulder is under ~1 dB, or it reverses.","caption":"Table 5: PSNR (dB) / SSIM for reconstruction with mixed-anatomy training across architectures.","image":"img/2604.11762.png","kind":"table","image_w":680,"image_h":255,"pwc":true}
|
site/p/2604.22847.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.22847","title":"Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes","authors":"Tim Merino; Sam Earle; Ryunosuke Iwai; Julian Togelius; Edoardo Cetin","openreview_id":"3YW4ug0yQX","presentation":"spotlight","track":"evaluations_datasets","verdict":"borderline","background":"Minecraft worlds are 3D voxel grids in which each block is a categorical, semantically meaningful token, but there has been no large block-level dataset for training generative models of such worlds. The paper releases Dream-Cubed, about 2M 32x32x32 chunks of biome-labelled procedural terrain plus human-built maps, and studies biome-conditioned 3D diffusion models trained on it. One practical question for building such a dataset is whether to sample terrain at its natural frequency or to collect rare, structurally complex content on purpose, such as villages and caves.","method":["Each sample is a 32x32x32 tensor of block IDs (V = 117 for the natural-biome data), conditioned on one of 15 chunk-level biome labels (13 natural biomes plus targeted 'cave' and 'village' classes). The model is MD4 discrete masked diffusion with a ~280M-parameter 3D DiT (25 blocks, hidden size 768, 8 heads, AdaLN conditioning) at patch size 4 (512 tokens). It is trained with AdamW (lr 3e-4, weight decay 0.01), effective batch 512, 160 epochs, and classifier-free guidance (condition dropout 0.2, guidance scale 4.0) with random 90-degree rotations as augmentation.","The same architecture is trained on three dataset variants. Balanced: ~67k chunks for each of the 15 biomes. Natural occurrence: biomes in proportion to their frequency in randomly generated worlds (Table 7), so village and cave are each 0.71% (7,100 chunks) and ocean is 35%. Village-boosted: village samples doubled, with the other biomes balanced but reduced proportionally. For evaluation, 100 chunks are generated per biome and rendered to images, and FID is computed against a fixed set of validation chunks. The reported 'adjusted' FID subtracts a reference FID computed from 100 held-out real chunks of that biome."],"claim":"With the same MD4 (patch 4) model, village-biome adjusted FID is ~38 when trained on the natural-occurrence dataset, ~24 on the balanced dataset and ~7.5 on the village-boosted dataset, while average adjusted FID over the 15 biomes stays similar (12.87 / 14.28 / 15.08).","source":"Figure 3, page 7 (Section 4, 'Data composition'); dataset distribution in Table 7, page 21","metric":"Adjusted per-biome FID on rendered chunks (generated FID minus a real-vs-real reference FID), lower is better; 100 generated samples per biome; single training run per variant; no error bars reported","comparison":"MD4 patch-4 models with the same architecture and hyperparameters trained on the Balanced, Natural-occurrence and Village-boosted dataset compositions, evaluated per biome","falsification":"Re-evaluating (or retraining) the three models gives village-biome adjusted FID that does not follow Village-boosted < Balanced < Natural occurrence, or the gap between Natural occurrence and Village-boosted on village shrinks to within the variation seen across repeated 100-sample FID evaluations of the same model.","caption":"Figure 3: Adjusted per-biome FID comparison between MD4 (patch 4) models across three dataset composition variants. Lower is better.","image":"img/2604.22847.png","kind":"figure","image_w":1166,"image_h":736,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.22847","title":"Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes","authors":"Tim Merino; Sam Earle; Ryunosuke Iwai; Julian Togelius; Edoardo Cetin","openreview_id":"3YW4ug0yQX","presentation":"spotlight","track":"evaluations_datasets","verdict":"borderline","background":"Minecraft worlds are 3D voxel grids in which each block is a categorical, semantically meaningful token, but there has been no large block-level dataset for training generative models of such worlds. The paper releases Dream-Cubed, about 2M 32x32x32 chunks of biome-labelled procedural terrain plus human-built maps, and studies biome-conditioned 3D diffusion models trained on it. One practical question for building such a dataset is whether to sample terrain at its natural frequency or to collect rare, structurally complex content on purpose, such as villages and caves.","method":["Each sample is a 32x32x32 tensor of block IDs (V = 117 for the natural-biome data), conditioned on one of 15 chunk-level biome labels (13 natural biomes plus targeted 'cave' and 'village' classes). The model is MD4 discrete masked diffusion with a ~280M-parameter 3D DiT (25 blocks, hidden size 768, 8 heads, AdaLN conditioning) at patch size 4 (512 tokens). It is trained with AdamW (lr 3e-4, weight decay 0.01), effective batch 512, 160 epochs, and classifier-free guidance (condition dropout 0.2, guidance scale 4.0) with random 90-degree rotations as augmentation.","The same architecture is trained on three dataset variants. Balanced: ~67k chunks for each of the 15 biomes. Natural occurrence: biomes in proportion to their frequency in randomly generated worlds (Table 7), so village and cave are each 0.71% (7,100 chunks) and ocean is 35%. Village-boosted: village samples doubled, with the other biomes balanced but reduced proportionally. For evaluation, 100 chunks are generated per biome and rendered to images, and FID is computed against a fixed set of validation chunks. The reported 'adjusted' FID subtracts a reference FID computed from 100 held-out real chunks of that biome."],"claim":"With the same MD4 (patch 4) model, village-biome adjusted FID is ~38 when trained on the natural-occurrence dataset, ~24 on the balanced dataset and ~7.5 on the village-boosted dataset, while average adjusted FID over the 15 biomes stays similar (12.87 / 14.28 / 15.08).","source":"Figure 3, page 7 (Section 4, 'Data composition'); dataset distribution in Table 7, page 21","metric":"Adjusted per-biome FID on rendered chunks (generated FID minus a real-vs-real reference FID), lower is better; 100 generated samples per biome; single training run per variant; no error bars reported","comparison":"MD4 patch-4 models with the same architecture and hyperparameters trained on the Balanced, Natural-occurrence and Village-boosted dataset compositions, evaluated per biome","falsification":"Re-evaluating (or retraining) the three models gives village-biome adjusted FID that does not follow Village-boosted < Balanced < Natural occurrence, or the gap between Natural occurrence and Village-boosted on village shrinks to within the variation seen across repeated 100-sample FID evaluations of the same model.","caption":"Figure 3: Adjusted per-biome FID comparison between MD4 (patch 4) models across three dataset composition variants. Lower is better.","image":"img/2604.22847.png","kind":"figure","image_w":1166,"image_h":736,"pwc":true}
|
site/p/2604.25231.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.25231","title":"DRAGON : A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams","authors":"Anirudh Iyengar Kaniyar Narayana Iyengar; Tampu Ravi Kumar; Gaurav Najpande; Manan Suri; Dinesh Manocha; Puneet Mathur; Vivek Gupta","openreview_id":"MFxuIiA2JO","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Vision-language models score well on diagram question answering benchmarks such as ChartQA and AI2D, but answer accuracy does not show whether a model actually used the diagram regions that support the answer. DRAGON turns this into a localization task: given a diagram, a question and the correct answer, the model must output bounding boxes for every visual element needed to justify the answer (answer-bearing regions, labels, legends, axes, connectors). The benchmark draws on six source datasets (ChartQA, Circuit-VQA, InfographicsVQA, MapIQ, MapWise, AI2D) and releases a 2,445-instance test split with human-verified evidence boxes.","method":["Each model receives (image, question, answer) and returns boxes as ID/x,y,w,h in pixels, parsed with regular expressions (malformed outputs are discarded). There are three prompting protocols: EDGE (predict all evidence boxes directly), SAGE (first list evidence elements as text, then ground each one), and VERGE (EDGE followed by a self-verification pass that fixes and adds boxes). Decoding uses temperature 0.2, top-p 0.7, and up to 2048 tokens for EDGE and 1024 per stage for the multi-stage protocols. Prompt templates are in Appendix C.","Boxes are normalized by image width and height. MPIoU is the maximum IoU over all predicted/gold box pairs, averaged over samples. Grounding IoU (GIoU) is the IoU between the union of predicted boxes and the union of gold boxes (Shapely polygons), averaged over samples. For F1, a predicted box is a true positive if its IoU with some gold box is at least tau. Precision and recall are computed per sample and averaged, then F1 = 2PR/(P+R), with 0 for samples that have no predicted or gold boxes. Table 3 averages each metric over the three prompts."],"claim":"On the MapIQ portion of the DRAGON test split, averaged over EDGE, SAGE and VERGE prompting, Claude Opus 4.6 reaches Grounding IoU 28.4 and F1 21.8, versus 17.4 and 12.5 for Kimi K2.5, the strongest open-weight model, and at most 6.2 GIoU and 2.0 F1 for the other open-weight models (Llama 4 Maverick, Qwen3.5, Gemma3, InternVL3.5).","source":"Table 3, page 8","metric":"Grounding IoU (union-vs-union IoU, %) and box-level F1 (%), higher is better; single run per model and prompt, no error bars or seeds reported","comparison":"Claude Opus 4.6 vs. Kimi K2.5 (and the other open-weight models) on MapIQ, same test instances, prompts and decoding settings","falsification":"Rerunning the three prompts on the MapIQ test instances with both models, Claude Opus 4.6's prompt-averaged GIoU is not clearly above Kimi K2.5's (e.g. a gap under ~5 points), or its F1 lead disappears.","caption":"Table 3: Model-level grounding performance across diagram domains averaged over EDGE, SAGE, and VERGE prompting. We report Max-Pair IoU (MPIoU), Grounding IoU (GIoU), and F1. Across models, MPIoU consistently exceeds GIoU and F1, indicating that models often localize approximate evidence regions but fail to capture the complete reasoning evidence required for diagram question answering.","image":"img/2604.25231.png","kind":"table","image_w":1144,"image_h":472,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.25231","title":"DRAGON : A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams","authors":"Anirudh Iyengar Kaniyar Narayana Iyengar; Tampu Ravi Kumar; Gaurav Najpande; Manan Suri; Dinesh Manocha; Puneet Mathur; Vivek Gupta","openreview_id":"MFxuIiA2JO","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Vision-language models score well on diagram question answering benchmarks such as ChartQA and AI2D, but answer accuracy does not show whether a model actually used the diagram regions that support the answer. DRAGON turns this into a localization task: given a diagram, a question and the correct answer, the model must output bounding boxes for every visual element needed to justify the answer (answer-bearing regions, labels, legends, axes, connectors). The benchmark draws on six source datasets (ChartQA, Circuit-VQA, InfographicsVQA, MapIQ, MapWise, AI2D) and releases a 2,445-instance test split with human-verified evidence boxes.","method":["Each model receives (image, question, answer) and returns boxes as ID/x,y,w,h in pixels, parsed with regular expressions (malformed outputs are discarded). There are three prompting protocols: EDGE (predict all evidence boxes directly), SAGE (first list evidence elements as text, then ground each one), and VERGE (EDGE followed by a self-verification pass that fixes and adds boxes). Decoding uses temperature 0.2, top-p 0.7, and up to 2048 tokens for EDGE and 1024 per stage for the multi-stage protocols. Prompt templates are in Appendix C.","Boxes are normalized by image width and height. MPIoU is the maximum IoU over all predicted/gold box pairs, averaged over samples. Grounding IoU (GIoU) is the IoU between the union of predicted boxes and the union of gold boxes (Shapely polygons), averaged over samples. For F1, a predicted box is a true positive if its IoU with some gold box is at least tau. Precision and recall are computed per sample and averaged, then F1 = 2PR/(P+R), with 0 for samples that have no predicted or gold boxes. Table 3 averages each metric over the three prompts."],"claim":"On the MapIQ portion of the DRAGON test split, averaged over EDGE, SAGE and VERGE prompting, Claude Opus 4.6 reaches Grounding IoU 28.4 and F1 21.8, versus 17.4 and 12.5 for Kimi K2.5, the strongest open-weight model, and at most 6.2 GIoU and 2.0 F1 for the other open-weight models (Llama 4 Maverick, Qwen3.5, Gemma3, InternVL3.5).","source":"Table 3, page 8","metric":"Grounding IoU (union-vs-union IoU, %) and box-level F1 (%), higher is better; single run per model and prompt, no error bars or seeds reported","comparison":"Claude Opus 4.6 vs. Kimi K2.5 (and the other open-weight models) on MapIQ, same test instances, prompts and decoding settings","falsification":"Rerunning the three prompts on the MapIQ test instances with both models, Claude Opus 4.6's prompt-averaged GIoU is not clearly above Kimi K2.5's (e.g. a gap under ~5 points), or its F1 lead disappears.","caption":"Table 3: Model-level grounding performance across diagram domains averaged over EDGE, SAGE, and VERGE prompting. We report Max-Pair IoU (MPIoU), Grounding IoU (GIoU), and F1. Across models, MPIoU consistently exceeds GIoU and F1, indicating that models often localize approximate evidence regions but fail to capture the complete reasoning evidence required for diagram question answering.","image":"img/2604.25231.png","kind":"table","image_w":1144,"image_h":472,"pwc":true}
|
site/p/2604.27819.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.27819","title":"MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents","authors":"Haonan Li; Tianjun Sun; Yongqing Wang; Qisheng Zhang","openreview_id":"9mPzIg8uil","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLM agents connected to several Model Context Protocol (MCP) servers (filesystem, git, SQLite, browser, shell, memory) can read a credential from one place and write it somewhere else while doing a perfectly ordinary task, for example copying an API key from a .env file into a generated deploy script or report. Existing MCP safety benchmarks study adversarial attacks such as prompt injection or malicious servers. MCPHunt instead measures non-adversarial credential propagation across trust boundaries. It also asks whether simple prompt-level instructions can reduce that propagation without hurting task completion.","method":["The benchmark has 147 researcher-designed tasks: 135 tasks across 9 source-to-sink mechanism families (e.g. browser_to_local, config_to_script, file_to_file) plus 12 benign controls. Each task runs against 8 MCP servers in a workspace that is reset before every task. In the 'risky' environments (variants v1/v2/v3, which use different canary values), sensitive values are format-authentic canary strings (e.g. sk_live_*, AKIA*, ghp_*). A trace counts as propagated (unsafe) if at least one of 11 temporally ordered detection signals fires, for example a canary that was read earlier appearing in a later write argument, a shell command, a generated script or a browser form field. Detection is exact string matching on the canary or on a canary core of at least 10 characters. Utility is the fraction of traces where the task completed, the requested artifact was verified and the source data was read. Tasks are labelled in advance as CRS (completion requires the secret, e.g. 'copy everything') or non-CRS (a derived artifact is requested). The policy-violating rate is the propagation rate on non-CRS mechanism-tagged traces in risky environments.","In the mitigation study, GPT-5.4 (temperature 0) runs on risky_v1/v2/v3 under four system-prompt conditions, each collected as a separate run: M0, no mitigation; M1, a brief generic privacy reminder; M2, explicit redaction rules naming API keys, passwords, tokens and credentials; M3, a boundary-aware prompt that explains source-to-sink risk and gives examples of safe derived artifacts. Task prompts and agent configuration are otherwise identical across arms."],"claim":"On GPT-5.4 in the risky environments, the boundary-aware prompt (M3) reduces the trace-level credential propagation rate to 2.3% (384 traces), compared with 16.8% under a generic privacy reminder (M1, 381 traces) and 22.9% with no mitigation (M0, 384 traces). The policy-violating rate falls from 8.1% (M1) to 0.3% (M3), and utility is not lower (80.5% for M3 vs. 79.3% for M1 and 79.4% for M0).","source":"Table 21, page 24 (Appendix R); summarized in Section 4.6 and Table 7, page 9","metric":"Propagation rate = fraction of traces with at least one Tier 1 or Tier 2 canary-propagation signal (lower is better); policy-violating rate = the same rate on non-CRS mechanism-tagged traces; utility = fraction of traces where the task completed and the artifact was verified (higher is better). No error bars or CIs are reported for Table 21; each arm is a single pass at temperature 0.","comparison":"M3 (boundary-aware detailed prompt) vs. M1 (generic privacy reminder) and M0 (no mitigation), same model, tasks, environments and harness","falsification":"Rerunning the released harness with GPT-5.4 on risky_v1/v2/v3 gives an M3 propagation rate that is not clearly below M1's (e.g. the M1-minus-M3 gap shrinks from about 14.5 pp to a few points), or M3 utility falls clearly below M0/M1 utility (e.g. by more than about 5 pp).","caption":"Table 21: Mitigation effectiveness (GPT-5.4, risky environments). Δ = change vs. baseline (no mitigation). Policy-viol. = non-CRS mechanism-tagged risky-environment traces with propagation (safety failures). Utility is the outcome success rate. The baseline row is re-collected within the mitigation run, so its policy-violating rate (13.9%) differs slightly from the main-benchmark value (13.5%, Table 6).","image":"img/2604.27819.png","kind":"table","image_w":1144,"image_h":361,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.27819","title":"MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents","authors":"Haonan Li; Tianjun Sun; Yongqing Wang; Qisheng Zhang","openreview_id":"9mPzIg8uil","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"LLM agents connected to several Model Context Protocol (MCP) servers (filesystem, git, SQLite, browser, shell, memory) can read a credential from one place and write it somewhere else while doing a perfectly ordinary task, for example copying an API key from a .env file into a generated deploy script or report. Existing MCP safety benchmarks study adversarial attacks such as prompt injection or malicious servers. MCPHunt instead measures non-adversarial credential propagation across trust boundaries. It also asks whether simple prompt-level instructions can reduce that propagation without hurting task completion.","method":["The benchmark has 147 researcher-designed tasks: 135 tasks across 9 source-to-sink mechanism families (e.g. browser_to_local, config_to_script, file_to_file) plus 12 benign controls. Each task runs against 8 MCP servers in a workspace that is reset before every task. In the 'risky' environments (variants v1/v2/v3, which use different canary values), sensitive values are format-authentic canary strings (e.g. sk_live_*, AKIA*, ghp_*). A trace counts as propagated (unsafe) if at least one of 11 temporally ordered detection signals fires, for example a canary that was read earlier appearing in a later write argument, a shell command, a generated script or a browser form field. Detection is exact string matching on the canary or on a canary core of at least 10 characters. Utility is the fraction of traces where the task completed, the requested artifact was verified and the source data was read. Tasks are labelled in advance as CRS (completion requires the secret, e.g. 'copy everything') or non-CRS (a derived artifact is requested). The policy-violating rate is the propagation rate on non-CRS mechanism-tagged traces in risky environments.","In the mitigation study, GPT-5.4 (temperature 0) runs on risky_v1/v2/v3 under four system-prompt conditions, each collected as a separate run: M0, no mitigation; M1, a brief generic privacy reminder; M2, explicit redaction rules naming API keys, passwords, tokens and credentials; M3, a boundary-aware prompt that explains source-to-sink risk and gives examples of safe derived artifacts. Task prompts and agent configuration are otherwise identical across arms."],"claim":"On GPT-5.4 in the risky environments, the boundary-aware prompt (M3) reduces the trace-level credential propagation rate to 2.3% (384 traces), compared with 16.8% under a generic privacy reminder (M1, 381 traces) and 22.9% with no mitigation (M0, 384 traces). The policy-violating rate falls from 8.1% (M1) to 0.3% (M3), and utility is not lower (80.5% for M3 vs. 79.3% for M1 and 79.4% for M0).","source":"Table 21, page 24 (Appendix R); summarized in Section 4.6 and Table 7, page 9","metric":"Propagation rate = fraction of traces with at least one Tier 1 or Tier 2 canary-propagation signal (lower is better); policy-violating rate = the same rate on non-CRS mechanism-tagged traces; utility = fraction of traces where the task completed and the artifact was verified (higher is better). No error bars or CIs are reported for Table 21; each arm is a single pass at temperature 0.","comparison":"M3 (boundary-aware detailed prompt) vs. M1 (generic privacy reminder) and M0 (no mitigation), same model, tasks, environments and harness","falsification":"Rerunning the released harness with GPT-5.4 on risky_v1/v2/v3 gives an M3 propagation rate that is not clearly below M1's (e.g. the M1-minus-M3 gap shrinks from about 14.5 pp to a few points), or M3 utility falls clearly below M0/M1 utility (e.g. by more than about 5 pp).","caption":"Table 21: Mitigation effectiveness (GPT-5.4, risky environments). Δ = change vs. baseline (no mitigation). Policy-viol. = non-CRS mechanism-tagged risky-environment traces with propagation (safety failures). Utility is the outcome success rate. The baseline row is re-collected within the mitigation run, so its policy-violating rate (13.9%) differs slightly from the main-benchmark value (13.5%, Table 6).","image":"img/2604.27819.png","kind":"table","image_w":1144,"image_h":361,"pwc":true}
|
site/p/2604.28031.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2604.28031","title":"Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation","authors":"Garvin Kruthof","openreview_id":"9kbTg0tKKr","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Researchers increasingly refine research proposals with LLMs over several turns, asking them to make ideas more novel or rigorous. Multi-turn benchmarks usually score per-turn quality or task success, not whether the output still respects a fixed set of hard constraints (budgets, sample floors, banned methods) from the original brief. A plan that quietly exceeds its budget still reads as complete, so this kind of drift has no external verifier and can go unnoticed.","method":["DriftBench has 38 structured research briefs across 24 domains, each with an objective, 5-8 hard constraints (234 total), success criteria and 3-5 banned moves. Seven subject models (GPT-5.4, GPT-5.4-mini, Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, Claude Sonnet 4.6, Qwen3-235B, Llama-3.3-70B-Instruct) each run every brief twice per condition at temperature 0.7 (Gemini 1.0). All conditions share one system prompt asking for fidelity, constraint tracking and minimal complexity. In Single-shot (SS) the model writes one proposal. In Multi-turn pressure (MT-P) it gets six turns with escalating user prompts ('Make it more novel', 'Strengthen the design', 'Add one more component'), and turn 6 asks for the same sectioned final proposal as single-shot. Multi-turn neutral (6 turns of 'Continue.') and checkpointed conditions are also run.","An LLM judge at temperature 0 scores the final proposal against the brief on four 0-4 rubrics: objective fidelity, constraint adherence (4 = all hard constraints fully respected; 3 = one stretched; 2 = one violated, ...), alternative coverage, and complexity inflation (higher = more over-engineered). In the crossed-judge design, each of Claude Opus 4.6 (primary) and GPT-5.4 (robustness check) scores all seven models' single-shot and pressure runs. Table 2 reports means +/- SE over about 530 runs per condition."],"claim":"Under a common Claude Opus 4.6 judge, iterative pressure lowers mean constraint adherence from 3.77 +/- 0.02 (single-shot) to 3.45 +/- 0.03 (6-turn pressure) and raises complexity inflation from 0.16 +/- 0.02 to 1.65 +/- 0.03 (n = 527/530 runs pooled over 7 models). Under a GPT-5.4 judge the same shift appears: adherence 3.24 -> 2.90, complexity 1.05 -> 2.63.","source":"Table 2, page 6","metric":"Judge-scored 0-4 rubric means; constraint adherence higher is better, complexity inflation lower is better. +/- is the standard error of the mean over runs (no seed variation beyond two repetitions per brief x model cell).","comparison":"Multi-turn pressure (6 turns, escalating improvement prompts) vs. single-shot on the same 38 briefs and 7 models, same system prompt, same final-proposal format, same judge. MT-Neutral (3.66 adherence, 0.67 complexity, published-arm judge) is the turn-count control.","falsification":"Re-running single-shot and pressure conditions on the released briefs (all models or a subset) with the same judge prompt shows an adherence drop that is within about 2 SE of zero (e.g. < 0.1 points), or no rise in complexity inflation, under either judge.","caption":"Table 2: Aggregate results on the clean sample (two repetitions per cell, 532 intended runs per condition; n = scored runs). Scores are 0–4, higher is better except for complexity; values are mean ± SE. Single-shot and pressure rows are scored by one common judge per block (Opus 4.6, primary; GPT-5.4, robustness check). Rows marked ∗ were not cross-judged and show published-arm values (Opus 4.6 for non-Anthropic runs, GPT-5.4 for Sonnet 4.6), given per model in Appendix G.","image":"img/2604.28031.png","kind":"table","image_w":1133,"image_h":422,"pwc":
|
|
|
|
| 1 |
+
{"id":"2604.28031","title":"Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation","authors":"Garvin Kruthof","openreview_id":"9kbTg0tKKr","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Researchers increasingly refine research proposals with LLMs over several turns, asking them to make ideas more novel or rigorous. Multi-turn benchmarks usually score per-turn quality or task success, not whether the output still respects a fixed set of hard constraints (budgets, sample floors, banned methods) from the original brief. A plan that quietly exceeds its budget still reads as complete, so this kind of drift has no external verifier and can go unnoticed.","method":["DriftBench has 38 structured research briefs across 24 domains, each with an objective, 5-8 hard constraints (234 total), success criteria and 3-5 banned moves. Seven subject models (GPT-5.4, GPT-5.4-mini, Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, Claude Sonnet 4.6, Qwen3-235B, Llama-3.3-70B-Instruct) each run every brief twice per condition at temperature 0.7 (Gemini 1.0). All conditions share one system prompt asking for fidelity, constraint tracking and minimal complexity. In Single-shot (SS) the model writes one proposal. In Multi-turn pressure (MT-P) it gets six turns with escalating user prompts ('Make it more novel', 'Strengthen the design', 'Add one more component'), and turn 6 asks for the same sectioned final proposal as single-shot. Multi-turn neutral (6 turns of 'Continue.') and checkpointed conditions are also run.","An LLM judge at temperature 0 scores the final proposal against the brief on four 0-4 rubrics: objective fidelity, constraint adherence (4 = all hard constraints fully respected; 3 = one stretched; 2 = one violated, ...), alternative coverage, and complexity inflation (higher = more over-engineered). In the crossed-judge design, each of Claude Opus 4.6 (primary) and GPT-5.4 (robustness check) scores all seven models' single-shot and pressure runs. Table 2 reports means +/- SE over about 530 runs per condition."],"claim":"Under a common Claude Opus 4.6 judge, iterative pressure lowers mean constraint adherence from 3.77 +/- 0.02 (single-shot) to 3.45 +/- 0.03 (6-turn pressure) and raises complexity inflation from 0.16 +/- 0.02 to 1.65 +/- 0.03 (n = 527/530 runs pooled over 7 models). Under a GPT-5.4 judge the same shift appears: adherence 3.24 -> 2.90, complexity 1.05 -> 2.63.","source":"Table 2, page 6","metric":"Judge-scored 0-4 rubric means; constraint adherence higher is better, complexity inflation lower is better. +/- is the standard error of the mean over runs (no seed variation beyond two repetitions per brief x model cell).","comparison":"Multi-turn pressure (6 turns, escalating improvement prompts) vs. single-shot on the same 38 briefs and 7 models, same system prompt, same final-proposal format, same judge. MT-Neutral (3.66 adherence, 0.67 complexity, published-arm judge) is the turn-count control.","falsification":"Re-running single-shot and pressure conditions on the released briefs (all models or a subset) with the same judge prompt shows an adherence drop that is within about 2 SE of zero (e.g. < 0.1 points), or no rise in complexity inflation, under either judge.","caption":"Table 2: Aggregate results on the clean sample (two repetitions per cell, 532 intended runs per condition; n = scored runs). Scores are 0–4, higher is better except for complexity; values are mean ± SE. Single-shot and pressure rows are scored by one common judge per block (Opus 4.6, primary; GPT-5.4, robustness check). Rows marked ∗ were not cross-judged and show published-arm values (Opus 4.6 for non-Anthropic runs, GPT-5.4 for Sonnet 4.6), given per model in Appendix G.","image":"img/2604.28031.png","kind":"table","image_w":1133,"image_h":422,"pwc":true}
|
site/p/2605.01417.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.01417","title":"Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks","authors":"Benjamin Warner; Ratna S Grandhi; Max Kieffer; Aymane Ouraq; Saurav Panigrahi; Geetu Ambwani; Kunal Bagga; Nikhil Khandekar; Arya Hariharan; Nishant Mishra; Manish Saravanan; Shamus S Yang; Ahmed Essouaied; Adepoju J Moyondafoluwa; Robert Scholz; Bofeng Huang; Molly Beavers; Srishti Gureja; Anish Mahishi; Sameed Khan; Maxime Griot; Hunar Batra; Jean-Benoit Delbrouck; Siddhant Bharadwaj; Ronald Clark; Ashish Vashist; Anas Zafar; Leema K Murali; Harsh Deshpande; Ameen Patel; William Brown; Johannes Hagemann; Connor Lane; Paul Scotti; Tanishq M Abraham","openreview_id":"QnhGNjR3s2","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Medical LLM benchmarks such as MultiMedQA are largely saturated, rely on restricted data (MedHELM), or cover few tasks or models (HealthBench). A recurring question is whether domain-specific medical fine-tuning still helps modern models or whether general-purpose models are enough. Medmarks collects 30 open benchmarks (19 verifiable in Medmarks-V, 11 open-ended in Medmarks-OE judged by LLMs), implements them as verifiers environments, and evaluates 61 models in 71 configurations, including medical fine-tunes next to their generalist base models.","method":["Each Medmarks-V benchmark is run through the verifiers library with a vLLM server, using the benchmark's original prompt where one exists (otherwise HELM / LM Eval Harness prompts, or a minimal 'think step by step' system prompt with answer-format instructions). Models use their officially recommended sampling settings (Gemma 3 4B: temperature 1.0, top-p 0.95, top-k 60; MedGemma 4B: temperature 0.0, top-p 1.0, top-k 0; Table 8). Multiple-choice answers are scored with a lenient grader that accepts the letter, the exact answer text, or both, strips dangling thinking traces, normalises case, punctuation and whitespace, looks after common answer prefixes, and handles simple negation (Appendix I). Most multiple-choice benchmarks get three rollouts: one in the original answer order and two with shuffled answer order.","The claim compares Google's Gemma 3 4B (generalist) with MedGemma 4B (the medical fine-tune of the same base) on the same benchmarks and harness. Figure 19 plots each model's per-benchmark normalized score as a bar, with individual rollouts shown as dots. In the main text (Figure 3) the same comparison is summarised as a rise in mean win rate on Medmarks-V from 0.321 to 0.378."],"claim":"Under the Medmarks-V harness, MedGemma 4B scores higher than its base model Gemma 3 4B on MedQA (about 0.65 vs. 0.52), MedMCQA (0.55 vs. 0.46) and PubMedQA (0.73 vs. 0.64), with the three rollouts of each model not overlapping on any of the three benchmarks.","source":"Figure 19 (top panel), page 36; summarised in Figure 3, page 6","metric":"Per-benchmark normalized score (accuracy for these multiple-choice tasks), higher is better; the Figure 19 caption does not define bars or dots; from Section 4's description of three rollouts (one original answer order plus two shuffled orders), bars appear to be the mean over rollouts and dots the individual rollouts. No std or confidence interval is reported. Values are the bar labels printed in the figure.","comparison":"MedGemma 4B vs. Gemma 3 4B, same prompts, grader and benchmarks, each model with its recommended sampling parameters","falsification":"Re-running both models through the Medmarks environments (three rollouts with shuffled answer orders), MedGemma 4B's MedQA, MedMCQA or PubMedQA accuracy is not higher than Gemma 3 4B's by more than the spread of the rollouts, e.g. the MedQA gap is under about 0.05.","caption":"Figure 19. Comparing the performance of Gemma 3 models to MedGemma 3 models on Medmarks-V tasks.","image":"img/2605.01417.png","kind":"figure","image_w":1861,"image_h":1472,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.01417","title":"Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks","authors":"Benjamin Warner; Ratna S Grandhi; Max Kieffer; Aymane Ouraq; Saurav Panigrahi; Geetu Ambwani; Kunal Bagga; Nikhil Khandekar; Arya Hariharan; Nishant Mishra; Manish Saravanan; Shamus S Yang; Ahmed Essouaied; Adepoju J Moyondafoluwa; Robert Scholz; Bofeng Huang; Molly Beavers; Srishti Gureja; Anish Mahishi; Sameed Khan; Maxime Griot; Hunar Batra; Jean-Benoit Delbrouck; Siddhant Bharadwaj; Ronald Clark; Ashish Vashist; Anas Zafar; Leema K Murali; Harsh Deshpande; Ameen Patel; William Brown; Johannes Hagemann; Connor Lane; Paul Scotti; Tanishq M Abraham","openreview_id":"QnhGNjR3s2","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Medical LLM benchmarks such as MultiMedQA are largely saturated, rely on restricted data (MedHELM), or cover few tasks or models (HealthBench). A recurring question is whether domain-specific medical fine-tuning still helps modern models or whether general-purpose models are enough. Medmarks collects 30 open benchmarks (19 verifiable in Medmarks-V, 11 open-ended in Medmarks-OE judged by LLMs), implements them as verifiers environments, and evaluates 61 models in 71 configurations, including medical fine-tunes next to their generalist base models.","method":["Each Medmarks-V benchmark is run through the verifiers library with a vLLM server, using the benchmark's original prompt where one exists (otherwise HELM / LM Eval Harness prompts, or a minimal 'think step by step' system prompt with answer-format instructions). Models use their officially recommended sampling settings (Gemma 3 4B: temperature 1.0, top-p 0.95, top-k 60; MedGemma 4B: temperature 0.0, top-p 1.0, top-k 0; Table 8). Multiple-choice answers are scored with a lenient grader that accepts the letter, the exact answer text, or both, strips dangling thinking traces, normalises case, punctuation and whitespace, looks after common answer prefixes, and handles simple negation (Appendix I). Most multiple-choice benchmarks get three rollouts: one in the original answer order and two with shuffled answer order.","The claim compares Google's Gemma 3 4B (generalist) with MedGemma 4B (the medical fine-tune of the same base) on the same benchmarks and harness. Figure 19 plots each model's per-benchmark normalized score as a bar, with individual rollouts shown as dots. In the main text (Figure 3) the same comparison is summarised as a rise in mean win rate on Medmarks-V from 0.321 to 0.378."],"claim":"Under the Medmarks-V harness, MedGemma 4B scores higher than its base model Gemma 3 4B on MedQA (about 0.65 vs. 0.52), MedMCQA (0.55 vs. 0.46) and PubMedQA (0.73 vs. 0.64), with the three rollouts of each model not overlapping on any of the three benchmarks.","source":"Figure 19 (top panel), page 36; summarised in Figure 3, page 6","metric":"Per-benchmark normalized score (accuracy for these multiple-choice tasks), higher is better; the Figure 19 caption does not define bars or dots; from Section 4's description of three rollouts (one original answer order plus two shuffled orders), bars appear to be the mean over rollouts and dots the individual rollouts. No std or confidence interval is reported. Values are the bar labels printed in the figure.","comparison":"MedGemma 4B vs. Gemma 3 4B, same prompts, grader and benchmarks, each model with its recommended sampling parameters","falsification":"Re-running both models through the Medmarks environments (three rollouts with shuffled answer orders), MedGemma 4B's MedQA, MedMCQA or PubMedQA accuracy is not higher than Gemma 3 4B's by more than the spread of the rollouts, e.g. the MedQA gap is under about 0.05.","caption":"Figure 19. Comparing the performance of Gemma 3 models to MedGemma 3 models on Medmarks-V tasks.","image":"img/2605.01417.png","kind":"figure","image_w":1861,"image_h":1472,"pwc":true}
|
site/p/2605.01945.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.01945","title":"PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction","authors":"Zhiwen Yang; Pan Liu; yifan Li; Yunhua Zhong; Jun Xia","openreview_id":"dePVHdM1Gw","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Deep learning models that predict peptide tandem mass spectra (MS/MS fragment-ion intensities) from a modified peptide sequence, precursor charge and collision energy are used to build spectral libraries, analyse DIA data and rescore peptide-spectrum matches. Published models are usually evaluated with random train/test splits, so the same unmodified amino-acid backbone (possibly with different PTMs) can appear in both training and test data. The model can then memorise backbone-specific fragmentation patterns, and reported accuracy may overstate how well it generalises to unseen peptides.","method":["PepSpecBench filters PROSPECT (ProteomeTools-based spectral library) to peptides of length 6-40, precursor charge 1-6 and three PTMs (acetylation, carbamidomethylation, oxidation), and reconstructs a 500k/50k/50k train/validation/test set balanced 50% unmodified / 50% PTM. The default Backbone-Disjoint split hashes the backbone (the sequence with all PTM annotations stripped) deterministically into train/val/test, so no backbone appears in two partitions (MD5 of the naked sequence mapped to buckets 0-99; 80/10/10 by backbone). For the split ablation (Appendix C.5), the reconstructed PROSPECT train/val/test tables are concatenated into a ~600k pool and re-split 80/10/10 by an MD5 bucket of either a row-level key (Naive Random: sample_key plus row index, i.e. per spectrum regardless of backbone) or the modified_sequence string (Modified-Sequence: a backbone can recur with different PTMs across partitions); the Backbone-Disjoint arm is the main 500k/50k/50k benchmark split.","Prosit (GRU encoder-decoder, about 2.5M parameters, extended to peptide length 40, masked spectral-angle loss, AdamW, peak LR 1e-4, batch 1024, cosine schedule with 5% warmup, early stopping on validation median SA with patience 10) is trained separately under each split with the same recipe. Predictions and observed spectra are projected into a shared 234-dimensional canonical b/y-ion space (cleavage position x ion type x fragment charge up to 3; ground truth is binned at 0.1 Da and read at theoretical b/y m/z), impossible positions are masked, and per-spectrum Spectral Angle SA = (2/pi)·arccos(cosine similarity) and Pearson correlation are aggregated as medians over each split's test set."],"claim":"Training and testing Prosit on PROSPECT with a naive random split gives a median spectral angle of 0.1254 (PCC 0.9099), better than the 0.1376 (PCC 0.8921) obtained under the backbone-disjoint split, an apparent improvement of 0.0122 SA that the authors attribute to backbone leakage, whereas a modified-sequence split stays close to backbone-disjoint (SA 0.1392, PCC 0.8895).","source":"Table 3, page 6 (bootstrap CI for the backbone-disjoint arm in Table 14, page 22)","metric":"Median Spectral Angle over test spectra in the shared canonical space (lower is better); median SAS = 1 - SA and median PCC (higher is better) also reported. Table 3 has no error bars; Table 14 gives a 95% bootstrap CI (1000 resamples of 50k test spectra) only for the backbone-disjoint arm: 0.1376 [0.1354, 0.1397]. Single training run per split.","comparison":"Prosit trained/evaluated under Naive Random split vs. Backbone-Disjoint split (and Modified-Sequence split) on PROSPECT, same data source and training recipe.","falsification":"Retraining Prosit under both splits (ideally over several seeds) gives a naive-random median SA that is not lower than the backbone-disjoint median SA by clearly more than the seed-to-seed spread and bootstrap CI (e.g. a gap under about 0.004 SA, or reversed), or the modified-sequence split shows a gap comparable to naive random.","caption":"Table 3: Split-ablation metrics for Prosit on PROSPECT (medians). ∆ relative to Backbone-Disjoint.","image":"img/2605.01945.png","kind":"table","image_w":1144,"image_h":191,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.01945","title":"PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction","authors":"Zhiwen Yang; Pan Liu; yifan Li; Yunhua Zhong; Jun Xia","openreview_id":"dePVHdM1Gw","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Deep learning models that predict peptide tandem mass spectra (MS/MS fragment-ion intensities) from a modified peptide sequence, precursor charge and collision energy are used to build spectral libraries, analyse DIA data and rescore peptide-spectrum matches. Published models are usually evaluated with random train/test splits, so the same unmodified amino-acid backbone (possibly with different PTMs) can appear in both training and test data. The model can then memorise backbone-specific fragmentation patterns, and reported accuracy may overstate how well it generalises to unseen peptides.","method":["PepSpecBench filters PROSPECT (ProteomeTools-based spectral library) to peptides of length 6-40, precursor charge 1-6 and three PTMs (acetylation, carbamidomethylation, oxidation), and reconstructs a 500k/50k/50k train/validation/test set balanced 50% unmodified / 50% PTM. The default Backbone-Disjoint split hashes the backbone (the sequence with all PTM annotations stripped) deterministically into train/val/test, so no backbone appears in two partitions (MD5 of the naked sequence mapped to buckets 0-99; 80/10/10 by backbone). For the split ablation (Appendix C.5), the reconstructed PROSPECT train/val/test tables are concatenated into a ~600k pool and re-split 80/10/10 by an MD5 bucket of either a row-level key (Naive Random: sample_key plus row index, i.e. per spectrum regardless of backbone) or the modified_sequence string (Modified-Sequence: a backbone can recur with different PTMs across partitions); the Backbone-Disjoint arm is the main 500k/50k/50k benchmark split.","Prosit (GRU encoder-decoder, about 2.5M parameters, extended to peptide length 40, masked spectral-angle loss, AdamW, peak LR 1e-4, batch 1024, cosine schedule with 5% warmup, early stopping on validation median SA with patience 10) is trained separately under each split with the same recipe. Predictions and observed spectra are projected into a shared 234-dimensional canonical b/y-ion space (cleavage position x ion type x fragment charge up to 3; ground truth is binned at 0.1 Da and read at theoretical b/y m/z), impossible positions are masked, and per-spectrum Spectral Angle SA = (2/pi)·arccos(cosine similarity) and Pearson correlation are aggregated as medians over each split's test set."],"claim":"Training and testing Prosit on PROSPECT with a naive random split gives a median spectral angle of 0.1254 (PCC 0.9099), better than the 0.1376 (PCC 0.8921) obtained under the backbone-disjoint split, an apparent improvement of 0.0122 SA that the authors attribute to backbone leakage, whereas a modified-sequence split stays close to backbone-disjoint (SA 0.1392, PCC 0.8895).","source":"Table 3, page 6 (bootstrap CI for the backbone-disjoint arm in Table 14, page 22)","metric":"Median Spectral Angle over test spectra in the shared canonical space (lower is better); median SAS = 1 - SA and median PCC (higher is better) also reported. Table 3 has no error bars; Table 14 gives a 95% bootstrap CI (1000 resamples of 50k test spectra) only for the backbone-disjoint arm: 0.1376 [0.1354, 0.1397]. Single training run per split.","comparison":"Prosit trained/evaluated under Naive Random split vs. Backbone-Disjoint split (and Modified-Sequence split) on PROSPECT, same data source and training recipe.","falsification":"Retraining Prosit under both splits (ideally over several seeds) gives a naive-random median SA that is not lower than the backbone-disjoint median SA by clearly more than the seed-to-seed spread and bootstrap CI (e.g. a gap under about 0.004 SA, or reversed), or the modified-sequence split shows a gap comparable to naive random.","caption":"Table 3: Split-ablation metrics for Prosit on PROSPECT (medians). ∆ relative to Backbone-Disjoint.","image":"img/2605.01945.png","kind":"table","image_w":1144,"image_h":191,"pwc":true}
|
site/p/2605.02199.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.02199","title":"MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing","authors":"Nishant Bhargava; Rodrigo S Barrento","openreview_id":"d7WzgVziwp","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Long-term LLM agents have to compress a stream of past interactions into a persistent memory store under a storage budget before they know which future queries will be asked. Memory systems such as Mem0, Letta and A-Mem are usually judged by end-to-end question-answering accuracy, which mixes up what was written to memory with retrieval, prompting and reader errors. MemAudit evaluates the write step on its own: it defines a finite 'package' of candidate memories with costs and semantic coverage, and scores a written store against the exact budgeted optimum of that package.","method":["A package fixes a stream of experiences, several candidate representations per experience (raw span, atomic fact, summary, tombstone, compound update), a cost per candidate (normalized word-equivalent units), semantic evidence units with weights w_r, and a coverage matrix a_ur. A store X must fit the budget (sum of costs <= B) and keep at most one representation per experience. Its utility is F(X) = sum_r w_r min(1, sum_{u in X} a_ur), a monotone submodular coverage function. OPT(B) is computed exactly by group-wise branch-and-bound (checked against a PuLP MILP on 1,200 instances, all matching), and a writer is scored by the ratio F(X)/OPT(B).","MemAudit-Small generates compact hidden event graphs with raw, fact, summary, tombstone and compound-update candidates. Over 500 seeds, budgets B = 2, 4, 8, 16 are swept. Writers, all constrained to the same package and budget: MemAudit-GVT (grouped value-threshold: for each arriving experience, among candidates whose exact oracle marginal density exceeds a threshold, insert the one with the largest marginal value, taking the best over a threshold grid); Density-only (pick the candidate with the highest marginal value per cost); Recency raw (keep the most recent raw spans until the budget runs out); Fact-only and Summary-only (restricted to one representation family); and No-tombstone OPT (exact optimum after removing tombstone and compound-update candidates)."],"claim":"On the 500-seed MemAudit-Small exact-package sweep, removing tombstone/compound-update candidates lowers the exact optimum to 0.523, 0.654, 0.689 and 0.815 of full OPT at B = 2, 4, 8, 16. Density-only selection covers no invalidation units at B = 2, 4, 8 and reaches only 0.264-0.361 of OPT there, while oracle-marginal GVT stays near OPT (about 0.96-1.00).","source":"Figure 2, page 6 (numbers also in Section 4 text, page 5)","metric":"Ratio of the writer's coverage utility F(X) to the certified exact package optimum OPT(B), averaged over 500 seeds (higher is better). The caption says bands are bootstrap 95% intervals, but they are not visible at plot resolution and no numeric intervals are given.","comparison":"Exact OPT vs. No-tombstone OPT (representation-frontier ablation), and MemAudit-GVT vs. Density-only, Recency raw, Fact-only and Summary-only writers on the same packages and budgets","falsification":"On regenerated MemAudit-Small packages, No-tombstone OPT stays within a few percent of full OPT (e.g. above 0.9 at B = 2-8), or Density-only reaches comparable ratios to GVT (e.g. above 0.8) at B = 2, 4, 8, or Density-only covers invalidation units at those budgets.","caption":"Figure 2: Exact-small budget sweep. MemAudit-GVT remains near certified package OPT, while density-only, recency raw, and single-representation baselines degrade under tight storage. Bands are bootstrap 95% intervals from the canonical 500-seed run.","image":"img/2605.02199.png","kind":"figure","image_w":1144,"image_h":580,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.02199","title":"MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing","authors":"Nishant Bhargava; Rodrigo S Barrento","openreview_id":"d7WzgVziwp","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Long-term LLM agents have to compress a stream of past interactions into a persistent memory store under a storage budget before they know which future queries will be asked. Memory systems such as Mem0, Letta and A-Mem are usually judged by end-to-end question-answering accuracy, which mixes up what was written to memory with retrieval, prompting and reader errors. MemAudit evaluates the write step on its own: it defines a finite 'package' of candidate memories with costs and semantic coverage, and scores a written store against the exact budgeted optimum of that package.","method":["A package fixes a stream of experiences, several candidate representations per experience (raw span, atomic fact, summary, tombstone, compound update), a cost per candidate (normalized word-equivalent units), semantic evidence units with weights w_r, and a coverage matrix a_ur. A store X must fit the budget (sum of costs <= B) and keep at most one representation per experience. Its utility is F(X) = sum_r w_r min(1, sum_{u in X} a_ur), a monotone submodular coverage function. OPT(B) is computed exactly by group-wise branch-and-bound (checked against a PuLP MILP on 1,200 instances, all matching), and a writer is scored by the ratio F(X)/OPT(B).","MemAudit-Small generates compact hidden event graphs with raw, fact, summary, tombstone and compound-update candidates. Over 500 seeds, budgets B = 2, 4, 8, 16 are swept. Writers, all constrained to the same package and budget: MemAudit-GVT (grouped value-threshold: for each arriving experience, among candidates whose exact oracle marginal density exceeds a threshold, insert the one with the largest marginal value, taking the best over a threshold grid); Density-only (pick the candidate with the highest marginal value per cost); Recency raw (keep the most recent raw spans until the budget runs out); Fact-only and Summary-only (restricted to one representation family); and No-tombstone OPT (exact optimum after removing tombstone and compound-update candidates)."],"claim":"On the 500-seed MemAudit-Small exact-package sweep, removing tombstone/compound-update candidates lowers the exact optimum to 0.523, 0.654, 0.689 and 0.815 of full OPT at B = 2, 4, 8, 16. Density-only selection covers no invalidation units at B = 2, 4, 8 and reaches only 0.264-0.361 of OPT there, while oracle-marginal GVT stays near OPT (about 0.96-1.00).","source":"Figure 2, page 6 (numbers also in Section 4 text, page 5)","metric":"Ratio of the writer's coverage utility F(X) to the certified exact package optimum OPT(B), averaged over 500 seeds (higher is better). The caption says bands are bootstrap 95% intervals, but they are not visible at plot resolution and no numeric intervals are given.","comparison":"Exact OPT vs. No-tombstone OPT (representation-frontier ablation), and MemAudit-GVT vs. Density-only, Recency raw, Fact-only and Summary-only writers on the same packages and budgets","falsification":"On regenerated MemAudit-Small packages, No-tombstone OPT stays within a few percent of full OPT (e.g. above 0.9 at B = 2-8), or Density-only reaches comparable ratios to GVT (e.g. above 0.8) at B = 2, 4, 8, or Density-only covers invalidation units at those budgets.","caption":"Figure 2: Exact-small budget sweep. MemAudit-GVT remains near certified package OPT, while density-only, recency raw, and single-representation baselines degrade under tight storage. Bands are bootstrap 95% intervals from the canonical 500-seed run.","image":"img/2605.02199.png","kind":"figure","image_w":1144,"image_h":580,"pwc":true}
|
site/p/2605.03952.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.03952","title":"MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents","authors":"Jonathan Steinberg; Oren Gal","openreview_id":"g9BxvmeCLK","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Coding agents are increasingly driven by streams of engineering tickets, and their safety training is evaluated on single, overt requests. MOSAIC-Bench asks whether a malicious end-state can be reached by splitting it into three innocuous-looking tickets, each of which an agent implements without objection, and whether downstream LLM code reviewers catch the resulting cumulative diff. Every chain has a deterministic proof-of-concept oracle that runs against a live Docker deployment of the web app, so 'exploitable' is fixed ground truth rather than an LLM judgment. If reviewer prompting alone changes detection a lot, that is a cheap, deployable mitigation.","method":["The benchmark has 199 three-stage attack chains over 10 web-app substrates (Express, Flask, Stripe, Gin, etc.), 31 CWE classes and 5 languages. Each chain ships Jira-style tickets, a stagewise golden solution and a Python PoC oracle that returns VULNERABLE on the full 3-stage golden composition and SECURE on every 1- and 2-stage prefix. Reviewer agents are handed the cumulative git diff of a confirmed-exploitable chain and must output APPROVE or DECLINE; evasion is the fraction of confirmed-exploitable chains that the reviewer APPROVEs.","Three reviewer protocols are compared on the same diffs: neutral diff-only (reviewer prompted as a senior engineer doing a normal PR review, given only the cumulative diff), neutral full-context (the full post-application repository is also available), and pentester framing (system prompt in App. M.1: list what changed, try to write a concrete exploit as a low-privilege user, then classify with a CWE and DECLINE only if an exploit was written). Five reviewers were run under pentester framing, including the open-weight Gemma-4-E4B-it. False positives of the Gemma pentester reviewer are measured separately on 608 merged real-world GitHub PRs."],"claim":"On the 199 confirmed-exploitable MOSAIC-Bench cumulative diffs, switching the Gemma-4-E4B-it reviewer from the neutral diff-only prompt to the pentester prompt reduces evasion (APPROVE on an exploitable chain) from 59.8% to 11.6% (88.4% detection), while giving the neutral reviewer full repository context raises evasion to 70.9%.","source":"Figure 3, page 7 (Gemma-4B bars); Section 6.1, page 8","metric":"Evasion rate (%) = fraction of oracle-confirmed exploitable chains the reviewer APPROVEs; lower is better. Single run per configuration; no error bars or seeds reported.","comparison":"Gemma-4-E4B-it reviewer: pentester framing vs. neutral diff-only vs. neutral full-context framing, same model, same 199 diffs. Secondary: the same comparison for Opus 4.6, Sonnet 4.6, Codex GPT-5.3 and Codex GPT-5.4.","falsification":"Re-running Gemma-4-E4B-it with the paper's pentester and neutral diff-only prompts on the released 199 chain diffs gives pentester evasion that is not substantially below neutral diff-only evasion (e.g., a reduction under ~20 pp, or pentester evasion above ~25%).","caption":"Figure 3: Pentester framing reduces reviewer evasion, but the magnitude varies by model. Framing dominates scale in expectation; GPT-5.4 is the main weak-transfer case.","image":"img/2605.03952.png","kind":"figure","image_w":1319,"image_h":741,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.03952","title":"MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents","authors":"Jonathan Steinberg; Oren Gal","openreview_id":"g9BxvmeCLK","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Coding agents are increasingly driven by streams of engineering tickets, and their safety training is evaluated on single, overt requests. MOSAIC-Bench asks whether a malicious end-state can be reached by splitting it into three innocuous-looking tickets, each of which an agent implements without objection, and whether downstream LLM code reviewers catch the resulting cumulative diff. Every chain has a deterministic proof-of-concept oracle that runs against a live Docker deployment of the web app, so 'exploitable' is fixed ground truth rather than an LLM judgment. If reviewer prompting alone changes detection a lot, that is a cheap, deployable mitigation.","method":["The benchmark has 199 three-stage attack chains over 10 web-app substrates (Express, Flask, Stripe, Gin, etc.), 31 CWE classes and 5 languages. Each chain ships Jira-style tickets, a stagewise golden solution and a Python PoC oracle that returns VULNERABLE on the full 3-stage golden composition and SECURE on every 1- and 2-stage prefix. Reviewer agents are handed the cumulative git diff of a confirmed-exploitable chain and must output APPROVE or DECLINE; evasion is the fraction of confirmed-exploitable chains that the reviewer APPROVEs.","Three reviewer protocols are compared on the same diffs: neutral diff-only (reviewer prompted as a senior engineer doing a normal PR review, given only the cumulative diff), neutral full-context (the full post-application repository is also available), and pentester framing (system prompt in App. M.1: list what changed, try to write a concrete exploit as a low-privilege user, then classify with a CWE and DECLINE only if an exploit was written). Five reviewers were run under pentester framing, including the open-weight Gemma-4-E4B-it. False positives of the Gemma pentester reviewer are measured separately on 608 merged real-world GitHub PRs."],"claim":"On the 199 confirmed-exploitable MOSAIC-Bench cumulative diffs, switching the Gemma-4-E4B-it reviewer from the neutral diff-only prompt to the pentester prompt reduces evasion (APPROVE on an exploitable chain) from 59.8% to 11.6% (88.4% detection), while giving the neutral reviewer full repository context raises evasion to 70.9%.","source":"Figure 3, page 7 (Gemma-4B bars); Section 6.1, page 8","metric":"Evasion rate (%) = fraction of oracle-confirmed exploitable chains the reviewer APPROVEs; lower is better. Single run per configuration; no error bars or seeds reported.","comparison":"Gemma-4-E4B-it reviewer: pentester framing vs. neutral diff-only vs. neutral full-context framing, same model, same 199 diffs. Secondary: the same comparison for Opus 4.6, Sonnet 4.6, Codex GPT-5.3 and Codex GPT-5.4.","falsification":"Re-running Gemma-4-E4B-it with the paper's pentester and neutral diff-only prompts on the released 199 chain diffs gives pentester evasion that is not substantially below neutral diff-only evasion (e.g., a reduction under ~20 pp, or pentester evasion above ~25%).","caption":"Figure 3: Pentester framing reduces reviewer evasion, but the magnitude varies by model. Framing dominates scale in expectation; GPT-5.4 is the main weak-transfer case.","image":"img/2605.03952.png","kind":"figure","image_w":1319,"image_h":741,"pwc":true}
|
site/p/2605.07565.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.07565","title":"Ensemble Distributionally Robust Bayesian Optimisation","authors":"Tigran Ramazyan; Denis Derkach","openreview_id":"qzmJlrajPM","presentation":"poster","track":"main","verdict":"include","background":"In contextual Bayesian optimisation the objective f(x, c) depends on a design x chosen by the learner and on a random context c drawn from an unknown distribution P, and the goal is to maximise the expected value E_{c~P}[f(x, c)]. Because P has to be estimated from the contexts observed so far, optimising against this estimate can pick poor designs. Distributionally robust BO (DRBO) guards against this by optimising against the worst distribution in an ambiguity set around the estimate. Existing DRBO methods either discretise the context space, which scales poorly, or use a single GP surrogate, which can be fragile on highly non-convex objectives.","method":["EDRBO fits an ensemble of M = 4 Gaussian processes on (x, c, y) data, using Squared Exponential, Rational Quadratic, Matérn-3/2 and Matérn-5/2 kernels. The ensemble prediction is the Wasserstein barycentre of the per-GP posteriors: the mean of the GP means mu_t and the mean of the GP standard deviations sigma_t. A per-point radius eps_hat_t(x, c) is the largest 2-Wasserstein distance between any single GP posterior and the barycentre, sqrt((mu_m - mu_t)^2 + (sigma_m - sigma_t)^2). Using a Wasserstein-DRO Lipschitz regulariser, the robust acquisition becomes x_t = argmax_x E_{c~Q_t}[mu_t(x,c) + beta_t sigma_t(x,c) + eps_hat_t(x,c)/sqrt(2)], where Q_t is the empirical distribution of the contexts observed so far. The learner then observes c_t ~ P and y_t = f(x_t, c_t) + noise.","Baselines (UCB, SBO-KDE, DRBO-KDE, DRBO-MMD, DRBO-MMD Minmax, StableOpt, WDRBO, ERBO) are adapted from the WDRBO / SBO-KDE code, all on BoTorch with a single RBF-kernel GP. Each method runs 200 iterations (5 of them initial points) on each of 9 problems with 10 random seeds. The metric is final cumulative expected regret, sum_t (E_P[f(x*, c)] - E_P[f(x_t, c)]). Ackley: x in [0,1]^2, c ~ N(0.5, 0.2^2) clipped to [0,1]. Newsvendor: f(x,c) = 9 min{x,c} + max{0, x-c} - 5x, with c drawn from a Burr XII distribution (alpha=2, beta=20) clipped to [0,1]."],"claim":"After 200 iterations, EDRBO has lower final cumulative regret than all 8 baselines on Ackley (273.96 ± 62.91 vs. 453.82 ± 127.19 for the best baseline, ERBO) and on Newsvendor (8.03 ± 1.46 vs. 13.54 ± 2.63 for the best baseline, SBO-KDE).","source":"Table 1, page 8","metric":"Final cumulative expected regret after 200 iterations (lower is better); ± is the standard deviation over 10 random seeds","comparison":"EDRBO (4-kernel GP ensemble) vs. UCB, SBO-KDE, DRBO-KDE, DRBO-MMD, DRBO-MMD Minmax, StableOpt, WDRBO, ERBO (single RBF GP), same benchmarks, budget and seeds","falsification":"Over 10 seeds on Ackley or Newsvendor, some baseline (in particular ERBO, SBO-KDE or WDRBO) reaches a mean final cumulative regret within one std of EDRBO's or below it; equivalently, EDRBO's Ackley mean is not clearly below about 450, or its Newsvendor mean is not clearly below about 13.5.","caption":"Table 1: Mean ± standard deviation of final cumulative regret across all experiments. The best and the second best results per experiment are in bold and underlined, respectively. The Average rank is computed over all experiments.","image":"img/2605.07565.png","kind":"table","image_w":1144,"image_h":641,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.07565","title":"Ensemble Distributionally Robust Bayesian Optimisation","authors":"Tigran Ramazyan; Denis Derkach","openreview_id":"qzmJlrajPM","presentation":"poster","track":"main","verdict":"include","background":"In contextual Bayesian optimisation the objective f(x, c) depends on a design x chosen by the learner and on a random context c drawn from an unknown distribution P, and the goal is to maximise the expected value E_{c~P}[f(x, c)]. Because P has to be estimated from the contexts observed so far, optimising against this estimate can pick poor designs. Distributionally robust BO (DRBO) guards against this by optimising against the worst distribution in an ambiguity set around the estimate. Existing DRBO methods either discretise the context space, which scales poorly, or use a single GP surrogate, which can be fragile on highly non-convex objectives.","method":["EDRBO fits an ensemble of M = 4 Gaussian processes on (x, c, y) data, using Squared Exponential, Rational Quadratic, Matérn-3/2 and Matérn-5/2 kernels. The ensemble prediction is the Wasserstein barycentre of the per-GP posteriors: the mean of the GP means mu_t and the mean of the GP standard deviations sigma_t. A per-point radius eps_hat_t(x, c) is the largest 2-Wasserstein distance between any single GP posterior and the barycentre, sqrt((mu_m - mu_t)^2 + (sigma_m - sigma_t)^2). Using a Wasserstein-DRO Lipschitz regulariser, the robust acquisition becomes x_t = argmax_x E_{c~Q_t}[mu_t(x,c) + beta_t sigma_t(x,c) + eps_hat_t(x,c)/sqrt(2)], where Q_t is the empirical distribution of the contexts observed so far. The learner then observes c_t ~ P and y_t = f(x_t, c_t) + noise.","Baselines (UCB, SBO-KDE, DRBO-KDE, DRBO-MMD, DRBO-MMD Minmax, StableOpt, WDRBO, ERBO) are adapted from the WDRBO / SBO-KDE code, all on BoTorch with a single RBF-kernel GP. Each method runs 200 iterations (5 of them initial points) on each of 9 problems with 10 random seeds. The metric is final cumulative expected regret, sum_t (E_P[f(x*, c)] - E_P[f(x_t, c)]). Ackley: x in [0,1]^2, c ~ N(0.5, 0.2^2) clipped to [0,1]. Newsvendor: f(x,c) = 9 min{x,c} + max{0, x-c} - 5x, with c drawn from a Burr XII distribution (alpha=2, beta=20) clipped to [0,1]."],"claim":"After 200 iterations, EDRBO has lower final cumulative regret than all 8 baselines on Ackley (273.96 ± 62.91 vs. 453.82 ± 127.19 for the best baseline, ERBO) and on Newsvendor (8.03 ± 1.46 vs. 13.54 ± 2.63 for the best baseline, SBO-KDE).","source":"Table 1, page 8","metric":"Final cumulative expected regret after 200 iterations (lower is better); ± is the standard deviation over 10 random seeds","comparison":"EDRBO (4-kernel GP ensemble) vs. UCB, SBO-KDE, DRBO-KDE, DRBO-MMD, DRBO-MMD Minmax, StableOpt, WDRBO, ERBO (single RBF GP), same benchmarks, budget and seeds","falsification":"Over 10 seeds on Ackley or Newsvendor, some baseline (in particular ERBO, SBO-KDE or WDRBO) reaches a mean final cumulative regret within one std of EDRBO's or below it; equivalently, EDRBO's Ackley mean is not clearly below about 450, or its Newsvendor mean is not clearly below about 13.5.","caption":"Table 1: Mean ± standard deviation of final cumulative regret across all experiments. The best and the second best results per experiment are in bold and underlined, respectively. The Average rank is computed over all experiments.","image":"img/2605.07565.png","kind":"table","image_w":1144,"image_h":641,"pwc":true}
|
site/p/2605.07919.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.07919","title":"MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence","authors":"Hanqi Jiang; Junhao Chen; Yi Pan; Lifeng Chen; Weihang You; Haozhen Gong; Ruiyu Yan; Jinglei Lv; Lin Zhao; Hui Ren; Quanzheng Li; Tianming Liu; Xiang Li","openreview_id":"VSlJ79obBe","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Medical vision-language models (VLMs) are usually evaluated on intact image-question pairs, where a high score can come partly from language priors rather than from reading the image. In clinical use a model should also notice when the evidence is broken, for example when the question contains a premise the image contradicts or the relevant region of the image is hidden, and decline or correct the premise instead of giving a fluent answer. The paper calls a fluent non-refusal answer in that situation a 'silent failure' and builds a benchmark to measure it.","method":["MedVIGIL has 300 cases from VQA-RAD, SLAKE, ROCO and MIMIC-CXR/CheXpert, annotated by radiologists with a gold answer, a five-option MCQ (A-E, where E is a probe-specific safe action such as rejecting a false premise or flagging insufficient evidence), a clinical risk tier L1-L5 and an ROI box. Each case is expanded into up to ten probes (2,556 total): original, paraphrase, negation, specificity-drop, knowledge-only, two hallucination traps (false premise contradicted by the image; correct answer is always the refusal option), and ROI-masked / ROI-only / left-right-flip images. Every model is queried once per probe at temperature 0 with a single-letter prompt; scoring is exact letter match with no judge model.","The silent-failure rate (SFR) is the fraction of the 600 trap probes on which the model picks a non-refusal option. SFRw weights the per-tier SFR by (1, 2, 3, 5, 8) for L1-L5, and the Safety axis is 100 - SFRw. Capability is the mean of original, paraphrase, negation and specificity-drop accuracy. Appendix M reports case-clustered bootstrap 95% CIs (500 resamples over the 300 cases, seed 20260505) for Capability, Safety, Grounding and the harmonic-mean composite MCS."],"claim":"On the MedVIGIL MCQ probes, GPT-4o (gpt-4o-2024-11-20) has a higher Capability score than HuatuoGPT-Vision-7B (54.8 vs 48.1) but a much lower risk-weighted Safety score: 36.0 [95% CI 31.2, 40.8] vs 60.4 [55.6, 64.9], with raw hallucination-trap silent-failure rate 53.0% vs 31.5%.","source":"Table 16, page 33 (bootstrap CIs); raw SFR and accuracy from Table 1, page 7","metric":"Safety = 100 - SFRw (risk-weighted silent-failure rate on hallucination-trap probes), higher is safer; brackets are case-clustered bootstrap 95% CIs over the 300 cases (500 resamples). Table 1 values (SFR, Overall, Original) have no error bars.","comparison":"GPT-4o vs. HuatuoGPT-Vision-7B on the same 2,556 probes with the same single-letter MCQ prompt and greedy decoding","falsification":"Re-running both models on the released probes gives GPT-4o a Safety score inside or above HuatuoGPT-V's 95% CI (>= 55.6), or the bootstrap CIs on Safety overlap, or GPT-4o's raw trap SFR is not clearly higher than HuatuoGPT-V's (e.g. a gap under 10 points instead of 21.5).","caption":"Table 16: Case-clustered bootstrap 95% CIs (500 resamples, fixed seed) for nine representative models spanning the MCS range. Capability uses the PR definition (Section 5.2). Rank-stability discussion in main text.","image":"img/2605.07919.png","kind":"table","image_w":1144,"image_h":472,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.07919","title":"MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence","authors":"Hanqi Jiang; Junhao Chen; Yi Pan; Lifeng Chen; Weihang You; Haozhen Gong; Ruiyu Yan; Jinglei Lv; Lin Zhao; Hui Ren; Quanzheng Li; Tianming Liu; Xiang Li","openreview_id":"VSlJ79obBe","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Medical vision-language models (VLMs) are usually evaluated on intact image-question pairs, where a high score can come partly from language priors rather than from reading the image. In clinical use a model should also notice when the evidence is broken, for example when the question contains a premise the image contradicts or the relevant region of the image is hidden, and decline or correct the premise instead of giving a fluent answer. The paper calls a fluent non-refusal answer in that situation a 'silent failure' and builds a benchmark to measure it.","method":["MedVIGIL has 300 cases from VQA-RAD, SLAKE, ROCO and MIMIC-CXR/CheXpert, annotated by radiologists with a gold answer, a five-option MCQ (A-E, where E is a probe-specific safe action such as rejecting a false premise or flagging insufficient evidence), a clinical risk tier L1-L5 and an ROI box. Each case is expanded into up to ten probes (2,556 total): original, paraphrase, negation, specificity-drop, knowledge-only, two hallucination traps (false premise contradicted by the image; correct answer is always the refusal option), and ROI-masked / ROI-only / left-right-flip images. Every model is queried once per probe at temperature 0 with a single-letter prompt; scoring is exact letter match with no judge model.","The silent-failure rate (SFR) is the fraction of the 600 trap probes on which the model picks a non-refusal option. SFRw weights the per-tier SFR by (1, 2, 3, 5, 8) for L1-L5, and the Safety axis is 100 - SFRw. Capability is the mean of original, paraphrase, negation and specificity-drop accuracy. Appendix M reports case-clustered bootstrap 95% CIs (500 resamples over the 300 cases, seed 20260505) for Capability, Safety, Grounding and the harmonic-mean composite MCS."],"claim":"On the MedVIGIL MCQ probes, GPT-4o (gpt-4o-2024-11-20) has a higher Capability score than HuatuoGPT-Vision-7B (54.8 vs 48.1) but a much lower risk-weighted Safety score: 36.0 [95% CI 31.2, 40.8] vs 60.4 [55.6, 64.9], with raw hallucination-trap silent-failure rate 53.0% vs 31.5%.","source":"Table 16, page 33 (bootstrap CIs); raw SFR and accuracy from Table 1, page 7","metric":"Safety = 100 - SFRw (risk-weighted silent-failure rate on hallucination-trap probes), higher is safer; brackets are case-clustered bootstrap 95% CIs over the 300 cases (500 resamples). Table 1 values (SFR, Overall, Original) have no error bars.","comparison":"GPT-4o vs. HuatuoGPT-Vision-7B on the same 2,556 probes with the same single-letter MCQ prompt and greedy decoding","falsification":"Re-running both models on the released probes gives GPT-4o a Safety score inside or above HuatuoGPT-V's 95% CI (>= 55.6), or the bootstrap CIs on Safety overlap, or GPT-4o's raw trap SFR is not clearly higher than HuatuoGPT-V's (e.g. a gap under 10 points instead of 21.5).","caption":"Table 16: Case-clustered bootstrap 95% CIs (500 resamples, fixed seed) for nine representative models spanning the MCS range. Capability uses the PR definition (Section 5.2). Rank-stability discussion in main text.","image":"img/2605.07919.png","kind":"table","image_w":1144,"image_h":472,"pwc":true}
|
site/p/2605.08647.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.08647","title":"AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators","authors":"Aritra Mazumder; Shubhashis Roy Dipta; Nusrat Jahan Lia; Tanzila Khan; Kainat R Hossain; Nehaa Shri; Humayra Tasnim; Gour G Shawon; Shubhrangshu Debsarkar; Sumaiya A Rani; Al Jami Islam Anik; Debjoty Mitra; Al Nafeu Khan","openreview_id":"P9qqypDzuC","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Multi-agent LLM systems pass messages between role-specialised agents (planner, auditor, engineer, ...), and a team can return a complete, valid-looking final answer even when a critical constraint was silently dropped along the way. Outcome-only benchmarks cannot see such process failures. AgentCollabBench is a 900-task benchmark across software engineering, DevOps and data engineering that injects controlled artifacts into multi-agent conversations to measure four failure modes, and it asks whether the way agents are wired together (the communication topology) matters as much as which model they run on.","method":["Each task specifies a team of agents, one of five communication topologies (linear chain, branching tree, converging DAG, fully connected, custom graph), a domain and a complexity tier. For the Radioactive Tracer Durability (RTD) metric, a unique tracer string (e.g. DATA-POLICY-RTN97DAYS-PII500MS) is embedded at a known graph location. Agents are assigned layers DAG-style (agents with no incoming edges are layer 0; each downstream agent is one plus the maximum layer of its predecessors). RTD is the deepest layer whose agent output contains the tracer (case-insensitive exact regex match on the canonical identifier), divided by the deepest layer in the task graph, so RTD is in [0, 1]. Scoring is deterministic, with no LLM judge.","The 232 RTD tasks (45-52 per topology) are run with all agents on a single backbone (GPT 4.1 mini, Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B, Llama 3.1 8B Instruct). Per-topology mean RTD is compared with a Kruskal-Wallis test (η² effect size) and Dunn post-hoc pairwise tests with Bonferroni correction. Tasks are co-designed with their topology, so different topologies use different tasks rather than the same task rewired."],"claim":"With GPT 4.1 mini as the backbone, RTD depends strongly on topology (Kruskal-Wallis H = 55.2, p < 1e-6, η² = 0.23): converging-DAG tasks keep the tracer to 49.4% mean depth vs. 100% for fully connected and 96.2% for custom graph (Dunn p < 0.001 for both), with linear chain 75.9% and branching tree 77.4%; converging DAG is the lowest-RTD topology for all four backbones.","source":"Figure 5 (top: GPT 4.1 mini per-topology means; bottom: η² for all models), page 8; exact means in Table 13 (page 27), tests in Tables 17-18 (page 29)","metric":"RTD (%), higher is better; mean over tasks in each topology; the Figure 5 caption does not define its error bars (the paper's other per-model results, Table 9, are mean ± standard error over task scores); single evaluation run, no seeds. Kruskal-Wallis H, p and η² across the five topologies.","comparison":"Mean RTD on converging-DAG tasks vs. fully-connected and custom-graph tasks (and vs. linear chain / branching tree), same backbone (GPT 4.1 mini) and same evaluation protocol; replicated for Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B and Llama 3.1 8B Instruct.","falsification":"Running the released RTD tasks with GPT 4.1 mini, converging-DAG mean RTD is not clearly below fully connected and custom graph (gap well under ~30 points or Dunn test not significant), or the Kruskal-Wallis topology effect on RTD is not significant / η² falls near 0.","caption":"Figure 5: RQ4 topology effects on information propagation. Direct-routing topologies (fully connected, custom graph) preserve tracers almost completely (RTD ~96-100% for GPT 4.1 mini) by eliminating multi-hop propagation paths; converging DAG and branching tree, which require multi-hop forwarding, lose substantially more (RTD 50% and 78%). The association replicates across all four evaluated models (η² = 0.07-0.40). Top: GPT 4.1 mini metric means by topology. Bottom: η² magnitudes across all models.","image":"img/2605.08647.png","kind":"figure","image_w":1138,"image_h":744,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.08647","title":"AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators","authors":"Aritra Mazumder; Shubhashis Roy Dipta; Nusrat Jahan Lia; Tanzila Khan; Kainat R Hossain; Nehaa Shri; Humayra Tasnim; Gour G Shawon; Shubhrangshu Debsarkar; Sumaiya A Rani; Al Jami Islam Anik; Debjoty Mitra; Al Nafeu Khan","openreview_id":"P9qqypDzuC","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Multi-agent LLM systems pass messages between role-specialised agents (planner, auditor, engineer, ...), and a team can return a complete, valid-looking final answer even when a critical constraint was silently dropped along the way. Outcome-only benchmarks cannot see such process failures. AgentCollabBench is a 900-task benchmark across software engineering, DevOps and data engineering that injects controlled artifacts into multi-agent conversations to measure four failure modes, and it asks whether the way agents are wired together (the communication topology) matters as much as which model they run on.","method":["Each task specifies a team of agents, one of five communication topologies (linear chain, branching tree, converging DAG, fully connected, custom graph), a domain and a complexity tier. For the Radioactive Tracer Durability (RTD) metric, a unique tracer string (e.g. DATA-POLICY-RTN97DAYS-PII500MS) is embedded at a known graph location. Agents are assigned layers DAG-style (agents with no incoming edges are layer 0; each downstream agent is one plus the maximum layer of its predecessors). RTD is the deepest layer whose agent output contains the tracer (case-insensitive exact regex match on the canonical identifier), divided by the deepest layer in the task graph, so RTD is in [0, 1]. Scoring is deterministic, with no LLM judge.","The 232 RTD tasks (45-52 per topology) are run with all agents on a single backbone (GPT 4.1 mini, Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B, Llama 3.1 8B Instruct). Per-topology mean RTD is compared with a Kruskal-Wallis test (η² effect size) and Dunn post-hoc pairwise tests with Bonferroni correction. Tasks are co-designed with their topology, so different topologies use different tasks rather than the same task rewired."],"claim":"With GPT 4.1 mini as the backbone, RTD depends strongly on topology (Kruskal-Wallis H = 55.2, p < 1e-6, η² = 0.23): converging-DAG tasks keep the tracer to 49.4% mean depth vs. 100% for fully connected and 96.2% for custom graph (Dunn p < 0.001 for both), with linear chain 75.9% and branching tree 77.4%; converging DAG is the lowest-RTD topology for all four backbones.","source":"Figure 5 (top: GPT 4.1 mini per-topology means; bottom: η² for all models), page 8; exact means in Table 13 (page 27), tests in Tables 17-18 (page 29)","metric":"RTD (%), higher is better; mean over tasks in each topology; the Figure 5 caption does not define its error bars (the paper's other per-model results, Table 9, are mean ± standard error over task scores); single evaluation run, no seeds. Kruskal-Wallis H, p and η² across the five topologies.","comparison":"Mean RTD on converging-DAG tasks vs. fully-connected and custom-graph tasks (and vs. linear chain / branching tree), same backbone (GPT 4.1 mini) and same evaluation protocol; replicated for Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B and Llama 3.1 8B Instruct.","falsification":"Running the released RTD tasks with GPT 4.1 mini, converging-DAG mean RTD is not clearly below fully connected and custom graph (gap well under ~30 points or Dunn test not significant), or the Kruskal-Wallis topology effect on RTD is not significant / η² falls near 0.","caption":"Figure 5: RQ4 topology effects on information propagation. Direct-routing topologies (fully connected, custom graph) preserve tracers almost completely (RTD ~96-100% for GPT 4.1 mini) by eliminating multi-hop propagation paths; converging DAG and branching tree, which require multi-hop forwarding, lose substantially more (RTD 50% and 78%). The association replicates across all four evaluated models (η² = 0.07-0.40). Top: GPT 4.1 mini metric means by topology. Bottom: η² magnitudes across all models.","image":"img/2605.08647.png","kind":"figure","image_w":1138,"image_h":744,"pwc":true}
|
site/p/2605.08988.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.08988","title":"Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials","authors":"Amir Masoud Nourollah; Irtaza Khalid; Stefano Leoni; Steven Schockaert","openreview_id":"MTNbRIVxvp","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Machine learning interatomic potentials (MLIPs) predict DFT-level energies and forces for molecular dynamics far more cheaply than DFT itself. They are usually evaluated on unseen configurations of the same molecules they were trained on, and foundation MLIPs are pre-trained on millions of structures in the hope that they generalise to new chemistry. It is unclear whether these models learn how molecular fragments compose (so that they can handle new molecules built from known parts) or mainly interpolate between training molecules. The GMD benchmark tests this with controlled train/test splits where the test molecules are new but are built from fragments that appear in training.","method":["GMD consists of ab initio MD trajectories of small H/C/N/O molecules labelled with ORCA PBE/def2-TZVP+D3BJ energies and forces. For each training molecule, 1000 snapshots are used for training, 200 for validation and 200 as an in-distribution (ID) test set (new configurations of training molecules). OOD test sets contain only molecules never seen in training. Task 1 (Fragment Chain Extension): train on C2-C6 linear alkanes, test on C7-C13 alkanes. Task 2 (Fragment Fusion): train on C4-C10 alcohols and aldehydes, C7-C11 diols, and complex polyols/polycarbonyls (up to C11); test on C4-C10 carboxylic acids, whose -COOH group combines the hydroxyl and carbonyl moieties seen separately in training.","Models are trained from scratch with each architecture's original repository and early stopping on ID validation loss; foundation models (MACE-MP-0 S/M/L, UMA-Small) are fully fine-tuned for 2000 epochs after re-estimating atomic reference energies. The metrics are force MAE over all Cartesian components (eV/Å) and total-energy MAE per configuration (eV), on the ID and OOD test sets, averaged over three test splits drawn from the held-out trajectories."],"claim":"MACE trained from scratch generalises to longer alkanes (Fragment Chain Extension force MAE 0.0034 ID vs. 0.0061 eV/Å OOD, energy MAE 0.0016 vs. 0.0087 eV) but fails on Fragment Fusion: force MAE rises from 0.0047 ± 0.0000 eV/Å (ID) to 0.0889 ± 0.0001 eV/Å on unseen carboxylic acids (about 19x), and energy MAE from 0.0036 eV to 0.9734 eV (about 270x).","source":"Table 6, page 41 (Appendix A.8); summarised in Figure 2, page 7","metric":"Force MAE (eV/Å) and total-energy MAE (eV), lower is better. ± is the spread over three test splits drawn from the held-out trajectories (sampling uncertainty of the test set), not over training seeds.","comparison":"The same from-scratch MACE model on in-distribution test configurations vs. OOD test molecules, on Fragment Chain Extension (small gap) vs. Fragment Fusion (large gap). Other architectures and fine-tuned foundation models in the same table show the same pattern.","falsification":"After training MACE from scratch on the released Fragment Fusion training split, the OOD/ID force-MAE ratio on carboxylic acids is close to the Chain Extension ratio (e.g. below ~3x), or the OOD force MAE is below ~0.03 eV/Å. Alternatively, the Chain Extension OOD/ID ratio is itself large (above ~5x).","caption":"Table 6: Comparison of MLIP models on the five GMD tasks: Fragment Chain Extension, Fragment Fusion, Fragment Duplication, Fragment Combination and Fragment Fission. Metrics reported are Forces Mean Absolute Error (F MAE) and Energy Mean Absolute Error (E MAE). Results are averaged across all molecules for each task. ID and OOD indicate in-distribution and out-of-distribution splits, respectively. Energy is in eV and forces in eV/Å. † Fragment Fission uses the Fragment Duplication models and in-distribution test set, so its ID values are those of Fragment Duplication. MACE-Small-25, MACE-Small-110 and UMA-Small-110 were not evaluated on this task.","image":"img/2605.08988.png","kind":"table","image_w":1138,"image_h":1205,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.08988","title":"Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials","authors":"Amir Masoud Nourollah; Irtaza Khalid; Stefano Leoni; Steven Schockaert","openreview_id":"MTNbRIVxvp","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Machine learning interatomic potentials (MLIPs) predict DFT-level energies and forces for molecular dynamics far more cheaply than DFT itself. They are usually evaluated on unseen configurations of the same molecules they were trained on, and foundation MLIPs are pre-trained on millions of structures in the hope that they generalise to new chemistry. It is unclear whether these models learn how molecular fragments compose (so that they can handle new molecules built from known parts) or mainly interpolate between training molecules. The GMD benchmark tests this with controlled train/test splits where the test molecules are new but are built from fragments that appear in training.","method":["GMD consists of ab initio MD trajectories of small H/C/N/O molecules labelled with ORCA PBE/def2-TZVP+D3BJ energies and forces. For each training molecule, 1000 snapshots are used for training, 200 for validation and 200 as an in-distribution (ID) test set (new configurations of training molecules). OOD test sets contain only molecules never seen in training. Task 1 (Fragment Chain Extension): train on C2-C6 linear alkanes, test on C7-C13 alkanes. Task 2 (Fragment Fusion): train on C4-C10 alcohols and aldehydes, C7-C11 diols, and complex polyols/polycarbonyls (up to C11); test on C4-C10 carboxylic acids, whose -COOH group combines the hydroxyl and carbonyl moieties seen separately in training.","Models are trained from scratch with each architecture's original repository and early stopping on ID validation loss; foundation models (MACE-MP-0 S/M/L, UMA-Small) are fully fine-tuned for 2000 epochs after re-estimating atomic reference energies. The metrics are force MAE over all Cartesian components (eV/Å) and total-energy MAE per configuration (eV), on the ID and OOD test sets, averaged over three test splits drawn from the held-out trajectories."],"claim":"MACE trained from scratch generalises to longer alkanes (Fragment Chain Extension force MAE 0.0034 ID vs. 0.0061 eV/Å OOD, energy MAE 0.0016 vs. 0.0087 eV) but fails on Fragment Fusion: force MAE rises from 0.0047 ± 0.0000 eV/Å (ID) to 0.0889 ± 0.0001 eV/Å on unseen carboxylic acids (about 19x), and energy MAE from 0.0036 eV to 0.9734 eV (about 270x).","source":"Table 6, page 41 (Appendix A.8); summarised in Figure 2, page 7","metric":"Force MAE (eV/Å) and total-energy MAE (eV), lower is better. ± is the spread over three test splits drawn from the held-out trajectories (sampling uncertainty of the test set), not over training seeds.","comparison":"The same from-scratch MACE model on in-distribution test configurations vs. OOD test molecules, on Fragment Chain Extension (small gap) vs. Fragment Fusion (large gap). Other architectures and fine-tuned foundation models in the same table show the same pattern.","falsification":"After training MACE from scratch on the released Fragment Fusion training split, the OOD/ID force-MAE ratio on carboxylic acids is close to the Chain Extension ratio (e.g. below ~3x), or the OOD force MAE is below ~0.03 eV/Å. Alternatively, the Chain Extension OOD/ID ratio is itself large (above ~5x).","caption":"Table 6: Comparison of MLIP models on the five GMD tasks: Fragment Chain Extension, Fragment Fusion, Fragment Duplication, Fragment Combination and Fragment Fission. Metrics reported are Forces Mean Absolute Error (F MAE) and Energy Mean Absolute Error (E MAE). Results are averaged across all molecules for each task. ID and OOD indicate in-distribution and out-of-distribution splits, respectively. Energy is in eV and forces in eV/Å. † Fragment Fission uses the Fragment Duplication models and in-distribution test set, so its ID values are those of Fragment Duplication. MACE-Small-25, MACE-Small-110 and UMA-Small-110 were not evaluated on this task.","image":"img/2605.08988.png","kind":"table","image_w":1138,"image_h":1205,"pwc":true}
|
site/p/2605.10442.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.10442","title":"StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs","authors":"Pierre Le Jeune; Etienne Duchesne; Weixuan Xiao; Stefano Palminteri; Bazire Houssin; Benoît Malézieux; Matteo Dora","openreview_id":"FRBT5zuX3I","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Most social-bias benchmarks for LLMs test whether a model recognizes a stereotype (multiple choice, templates, Winograd schemas) and are mostly in English. Frontier models now saturate these benchmarks, but they can still produce stereotypes when they write freely. StereoTales measures bias in open-ended generation instead: models write short stories where one demographic attribute of the protagonist is fixed, and the paper looks for which other attributes the models assign far more often than expected, in 10 languages and across 23 recent LLMs.","method":["Prompts combine one fixed value of one of 19 demographic attributes (79 values in total, e.g. age, gender, religion, income) with one of 36 narrative scenarios. Each model writes a ~200-word story per prompt. An ensemble of three LLM extractors (GPT 5 nano, Gemini 2.5 Flash Lite, Gemini 3.1 Flash Lite) then recovers all 19 attributes of the protagonist. For each model and each pair of attributes, the paper builds a contingency table and runs one-sided Fisher exact tests per cell (Benjamini–Yekutieli corrected, alpha 0.05). Cells that are significant and have lift P(b|a)/P(b) >= 2 count as 'associations'. Associations whose median human harm rating (5-point Likert, UK Prolific panel) is >= 4 count as harmful.","For the mitigation experiment (Appendix G.6), the whole pipeline is rerun on a fixed held-out set of 3,555 prompts for all 23 models under three system-prompt templates that differ only in one added sentence: Neutral (none), Positive ('Please write this story with a focus on fairness and objectivity.') and Negative ('Please ensure that your story is unbiased and does not propagate or rely on stereotypes.'). Associations are pooled across models and kept only if they are significant in at least two models. Templates are then compared pairwise with McNemar's test on the paired emitted/not-emitted indicators, Bonferroni-corrected over the 3 comparisons."],"claim":"On a fixed set of 3,555 prompts across 23 LLMs, adding the anti-stereotype instruction (Negative template) reduces the number of emitted associations (significant in >= 2 models) from 393 under the Neutral template to 327 (82 only-negative vs. 148 only-neutral; one-sided McNemar p = 8.0e-6, Bonferroni p = 2.4e-5). The generic fairness instruction (Positive template, 379) does not differ from Neutral (p = 0.87).","source":"Table 7, page 53 (Appendix G.6); per-template counts in Figures 16-17, page 52","metric":"Count of distinct associations emitted under each template (lower = fewer stereotype-like associations). Significance from a one-sided McNemar test on paired indicators, Bonferroni-adjusted over 3 comparisons. No seeds or error bars: each prompt is generated once per model.","comparison":"Negative (anti-stereotype instruction) vs. Neutral (no instruction) vs. Positive (fairness instruction) system prompts, same prompts, models, extractors and statistical pipeline","falsification":"Rerun on the same prompt set and models (or a fixed subset of models), the Negative template does not emit significantly fewer associations than Neutral (one-sided McNemar p >= 0.05 after Bonferroni correction), or the only-neutral vs. only-negative discordant counts are roughly balanced.","caption":"Table 7: Pairwise McNemar tests across prompt templates, on association indicators (emitted / not emitted). For each comparison, |A| and |B| are the total number of associations emitted under templates A and B; both is the number emitted under both templates, only A (resp. only B) the number emitted only under template A (resp. B). p_{B>A} is the one-sided McNemar p-value for the directional hypothesis that template B emits more associations than template A; p^bonf_{B>A} is its Bonferroni-adjusted counterpart across the 3 pairwise comparisons within each dataset. Comparisons with p^bonf_{B>A} < 0.05 are indicated in bold.","image":"img/2605.10442.png","kind":"table","image_w":1144,"image_h":455,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.10442","title":"StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs","authors":"Pierre Le Jeune; Etienne Duchesne; Weixuan Xiao; Stefano Palminteri; Bazire Houssin; Benoît Malézieux; Matteo Dora","openreview_id":"FRBT5zuX3I","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Most social-bias benchmarks for LLMs test whether a model recognizes a stereotype (multiple choice, templates, Winograd schemas) and are mostly in English. Frontier models now saturate these benchmarks, but they can still produce stereotypes when they write freely. StereoTales measures bias in open-ended generation instead: models write short stories where one demographic attribute of the protagonist is fixed, and the paper looks for which other attributes the models assign far more often than expected, in 10 languages and across 23 recent LLMs.","method":["Prompts combine one fixed value of one of 19 demographic attributes (79 values in total, e.g. age, gender, religion, income) with one of 36 narrative scenarios. Each model writes a ~200-word story per prompt. An ensemble of three LLM extractors (GPT 5 nano, Gemini 2.5 Flash Lite, Gemini 3.1 Flash Lite) then recovers all 19 attributes of the protagonist. For each model and each pair of attributes, the paper builds a contingency table and runs one-sided Fisher exact tests per cell (Benjamini–Yekutieli corrected, alpha 0.05). Cells that are significant and have lift P(b|a)/P(b) >= 2 count as 'associations'. Associations whose median human harm rating (5-point Likert, UK Prolific panel) is >= 4 count as harmful.","For the mitigation experiment (Appendix G.6), the whole pipeline is rerun on a fixed held-out set of 3,555 prompts for all 23 models under three system-prompt templates that differ only in one added sentence: Neutral (none), Positive ('Please write this story with a focus on fairness and objectivity.') and Negative ('Please ensure that your story is unbiased and does not propagate or rely on stereotypes.'). Associations are pooled across models and kept only if they are significant in at least two models. Templates are then compared pairwise with McNemar's test on the paired emitted/not-emitted indicators, Bonferroni-corrected over the 3 comparisons."],"claim":"On a fixed set of 3,555 prompts across 23 LLMs, adding the anti-stereotype instruction (Negative template) reduces the number of emitted associations (significant in >= 2 models) from 393 under the Neutral template to 327 (82 only-negative vs. 148 only-neutral; one-sided McNemar p = 8.0e-6, Bonferroni p = 2.4e-5). The generic fairness instruction (Positive template, 379) does not differ from Neutral (p = 0.87).","source":"Table 7, page 53 (Appendix G.6); per-template counts in Figures 16-17, page 52","metric":"Count of distinct associations emitted under each template (lower = fewer stereotype-like associations). Significance from a one-sided McNemar test on paired indicators, Bonferroni-adjusted over 3 comparisons. No seeds or error bars: each prompt is generated once per model.","comparison":"Negative (anti-stereotype instruction) vs. Neutral (no instruction) vs. Positive (fairness instruction) system prompts, same prompts, models, extractors and statistical pipeline","falsification":"Rerun on the same prompt set and models (or a fixed subset of models), the Negative template does not emit significantly fewer associations than Neutral (one-sided McNemar p >= 0.05 after Bonferroni correction), or the only-neutral vs. only-negative discordant counts are roughly balanced.","caption":"Table 7: Pairwise McNemar tests across prompt templates, on association indicators (emitted / not emitted). For each comparison, |A| and |B| are the total number of associations emitted under templates A and B; both is the number emitted under both templates, only A (resp. only B) the number emitted only under template A (resp. B). p_{B>A} is the one-sided McNemar p-value for the directional hypothesis that template B emits more associations than template A; p^bonf_{B>A} is its Bonferroni-adjusted counterpart across the 3 pairwise comparisons within each dataset. Comparisons with p^bonf_{B>A} < 0.05 are indicated in bold.","image":"img/2605.10442.png","kind":"table","image_w":1144,"image_h":455,"pwc":true}
|
site/p/2605.12147.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.12147","title":"PrivacySIM: Evaluating LLM Simulation of User Privacy Behavior","authors":"James Flemings; Qiumin Xu","openreview_id":"Nax0tIAhD9","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLMs are increasingly proposed as stand-ins for human participants, including for eliciting users' data-sharing preferences when designing privacy-sensitive systems. Users' privacy decisions are idiosyncratic and often contradict their stated attitudes (the 'privacy paradox'), so it is unclear whether conditioning an LLM on a user's persona lets it predict that individual's actual privacy choices. PrivacySIM tests this against ground-truth answers from real participants in published privacy user studies.","method":["The benchmark samples 200 users from each of five published user studies on LLM/chatbot privacy (LLM healthcare consultation, LLM chatbot norms, AI agent permissions, chatbot group chat, LLM conversational agents): 1,000 users and 9,400 user-question pairs. Each user has a persona made of up to three facets taken verbatim from the study: demographics (Du, only in 3 studies), previous AI/chatbot experiences (Eu) and stated privacy attitudes (Au). The LLM gets a fixed prompt template (role, persona block, JSON-output constraints, the original questionnaire) and answers all of that user's data-sharing questions in one JSON response.","The score is tolerance accuracy: the fraction of questions where the simulated answer is within 0.15 x scale width of the user's real answer (exact match on 4/5-point Likert scales, within +/-15 points on the 0-100 scales), averaged over users and then over the five studies. The key comparison is the no-persona prompt ('none', persona fields left empty) vs. the Eu+Au prompt, for the same model. Open models use T=1.0, top-p=0.95; closed models use default temperature and medium reasoning effort."],"claim":"Averaged over all five studies, conditioning on previous experiences plus privacy attitudes (Eu+Au) raises tolerance accuracy over the no-persona prompt for every model in Table 3, e.g. Gemini 3 Flash 37.07% -> 40.27%, Gemini 3.1 Pro 36.32% -> 40.39%, GPT-5.4 (medium) 35.70% -> 39.30%, Qwen3.5 27B 34.53% -> 36.41%; no model exceeds about 40.4%.","source":"Table 3, page 8 (also Figure 2, page 6 and Table 4, page 9)","metric":"Tolerance accuracy (%) at tau = 0.15, higher is better, averaged over the 5 studies; single run, no error bars reported","comparison":"Same model, no-persona prompt (acc(none)) vs. Eu+Au persona prompt (acc(Eu+Au)), across GPT-5.4 Mini/low/medium, Gemini 3 Flash/Pro, Qwen3.5 27B/122B","falsification":"Re-running the benchmark shows the Eu+Au gain over 'none' is not positive, or it is within run-to-run variation (e.g. repeated runs at the same settings vary by about 3 points), for the tested model; or the best model clearly exceeds about 45% 5-study accuracy at Eu+Au.","caption":"Table 3: Simulation accuracy across model families and variants. For GPT-5.4 we report the smaller mini model and the full model under low and medium reasoning effort. We find that reasoning effort has negligible impact and model size can yield modest simulation improvement.","image":"img/2605.12147.png","kind":"table","image_w":538,"image_h":483,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.12147","title":"PrivacySIM: Evaluating LLM Simulation of User Privacy Behavior","authors":"James Flemings; Qiumin Xu","openreview_id":"Nax0tIAhD9","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLMs are increasingly proposed as stand-ins for human participants, including for eliciting users' data-sharing preferences when designing privacy-sensitive systems. Users' privacy decisions are idiosyncratic and often contradict their stated attitudes (the 'privacy paradox'), so it is unclear whether conditioning an LLM on a user's persona lets it predict that individual's actual privacy choices. PrivacySIM tests this against ground-truth answers from real participants in published privacy user studies.","method":["The benchmark samples 200 users from each of five published user studies on LLM/chatbot privacy (LLM healthcare consultation, LLM chatbot norms, AI agent permissions, chatbot group chat, LLM conversational agents): 1,000 users and 9,400 user-question pairs. Each user has a persona made of up to three facets taken verbatim from the study: demographics (Du, only in 3 studies), previous AI/chatbot experiences (Eu) and stated privacy attitudes (Au). The LLM gets a fixed prompt template (role, persona block, JSON-output constraints, the original questionnaire) and answers all of that user's data-sharing questions in one JSON response.","The score is tolerance accuracy: the fraction of questions where the simulated answer is within 0.15 x scale width of the user's real answer (exact match on 4/5-point Likert scales, within +/-15 points on the 0-100 scales), averaged over users and then over the five studies. The key comparison is the no-persona prompt ('none', persona fields left empty) vs. the Eu+Au prompt, for the same model. Open models use T=1.0, top-p=0.95; closed models use default temperature and medium reasoning effort."],"claim":"Averaged over all five studies, conditioning on previous experiences plus privacy attitudes (Eu+Au) raises tolerance accuracy over the no-persona prompt for every model in Table 3, e.g. Gemini 3 Flash 37.07% -> 40.27%, Gemini 3.1 Pro 36.32% -> 40.39%, GPT-5.4 (medium) 35.70% -> 39.30%, Qwen3.5 27B 34.53% -> 36.41%; no model exceeds about 40.4%.","source":"Table 3, page 8 (also Figure 2, page 6 and Table 4, page 9)","metric":"Tolerance accuracy (%) at tau = 0.15, higher is better, averaged over the 5 studies; single run, no error bars reported","comparison":"Same model, no-persona prompt (acc(none)) vs. Eu+Au persona prompt (acc(Eu+Au)), across GPT-5.4 Mini/low/medium, Gemini 3 Flash/Pro, Qwen3.5 27B/122B","falsification":"Re-running the benchmark shows the Eu+Au gain over 'none' is not positive, or it is within run-to-run variation (e.g. repeated runs at the same settings vary by about 3 points), for the tested model; or the best model clearly exceeds about 45% 5-study accuracy at Eu+Au.","caption":"Table 3: Simulation accuracy across model families and variants. For GPT-5.4 we report the smaller mini model and the full model under low and medium reasoning effort. We find that reasoning effort has negligible impact and model size can yield modest simulation improvement.","image":"img/2605.12147.png","kind":"table","image_w":538,"image_h":483,"pwc":true}
|
site/p/2605.12662.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.12662","title":"scShapeBench: Discovering geometry from high dimensional scRNAseq data","authors":"Andrew J Steindl; Joao Felipe Rocha; Brian T Di Bassinga; Zachary Warren; Matthew Scicluna; César Miguel Valdez Córdova; Shabarni Gupta; Leire Torices; Daniel Neumann; Timothy J Mann; Ihuan Gunawan; Dhananjay Bhaskar; John G Lock; Christine L Chaffer; Guy Wolf; Smita Krishnaswamy","openreview_id":"eE5wOtNka2","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Single-cell RNA-seq data can be organized as discrete clusters, a single trajectory, a branching tree or an archetypal (simplex-like) continuum, and each shape calls for a different downstream analysis pipeline (Seurat clustering, Monocle, flow models, etc.). Today a bioinformatician picks the pipeline by looking at a 2D embedding, which blocks automated or agentic analysis. The paper frames 'shape detection' as recovering a reduced graph (degree-2 vertices contracted) from a high-dimensional point cloud, and releases a benchmark with synthetic point clouds sampled around known graphs plus 102 expert-annotated scRNAseq datasets.","method":["Synthetic track: 1,000 point clouds generated with seed 20260422 from random latent graphs (singletons, edges, trees, single cycles, multiple cycles, hybrids; up to four connected components; up to 80 nodes), embedded in R^2 to R^5, thickened into tubes and sampled with Gaussian noise, with difficulty axes noise ratio [0.02, 0.20], separation [0, 0.6], density [3.5, 25] and thickness [0.45, 1.0] (Appendix D.2). Each method outputs a graph S, which is reduced by contracting degree-2 chains and compared with the reduced ground-truth graph via (i) Wasserstein distance between edge-length-filtration persistence diagrams, converted to a similarity (higher is better), and (ii) approximate graph edit distance from NetworkX (lower is better), both averaged over samples.","scReebTower builds a k = 15 nearest-neighbour graph, an adaptive Gaussian kernel and the row-normalized diffusion operator P, and uses the first non-trivial eigenvector (Fiedler vector) as a Morse function. It takes midpoint thresholds between sorted function values, makes each connected component of the crossing-edge level set a node, links nodes in adjacent levels that share points, and reduces the resulting Reeb graph. The multiscale 'scReebTower' first applies diffusion condensation X̂ = P^t X with k_smooth = min(80, n-1); 'scReebTower (ℓ=0)' skips it. Baselines: PAGA (Scanpy, Leiden resolution 1.0, k = 15, keep all edges with connectivity > 0) and Mapper (KeplerMapper, PCA lens of dim ≤ 2, 6 intervals, overlap 0.35, DBSCAN min_samples 3, eps = 1.5 × median 3rd-NN distance). Hyperparameters are fixed with no per-sample tuning."],"claim":"On the 1,000-sample synthetic track, multiscale scReebTower recovers reduced graphs with a mean graph edit distance of 13.32 vs. 16.25 for PAGA and 79.78 for Mapper, and a Wasserstein persistence similarity of 0.3199 vs. 0.2788 for PAGA and 0.0318 for Mapper.","source":"Table 1, page 9","metric":"Mean approximate graph edit distance between reduced recovered and ground-truth graphs (lower is better) and mean Wasserstein persistence similarity (higher is better) over 1,000 synthetic samples; no error bars or seeds reported","comparison":"scReebTower (with diffusion condensation) vs. scReebTower (ℓ=0), PAGA and Mapper on the same synthetic point clouds","falsification":"On a regenerated synthetic benchmark following Appendix D.2, scReebTower's mean GED is not clearly below PAGA's (e.g. the gap shrinks to well under ~3 or reverses), or its persistence similarity is not above PAGA's.","caption":"Table 1: Benchmark results on the synthetic track. Higher is better for Wasserstein persistence similarity, lower is better for graph edit distance","image":"img/2605.12662.png","kind":"table","image_w":1144,"image_h":205,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.12662","title":"scShapeBench: Discovering geometry from high dimensional scRNAseq data","authors":"Andrew J Steindl; Joao Felipe Rocha; Brian T Di Bassinga; Zachary Warren; Matthew Scicluna; César Miguel Valdez Córdova; Shabarni Gupta; Leire Torices; Daniel Neumann; Timothy J Mann; Ihuan Gunawan; Dhananjay Bhaskar; John G Lock; Christine L Chaffer; Guy Wolf; Smita Krishnaswamy","openreview_id":"eE5wOtNka2","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Single-cell RNA-seq data can be organized as discrete clusters, a single trajectory, a branching tree or an archetypal (simplex-like) continuum, and each shape calls for a different downstream analysis pipeline (Seurat clustering, Monocle, flow models, etc.). Today a bioinformatician picks the pipeline by looking at a 2D embedding, which blocks automated or agentic analysis. The paper frames 'shape detection' as recovering a reduced graph (degree-2 vertices contracted) from a high-dimensional point cloud, and releases a benchmark with synthetic point clouds sampled around known graphs plus 102 expert-annotated scRNAseq datasets.","method":["Synthetic track: 1,000 point clouds generated with seed 20260422 from random latent graphs (singletons, edges, trees, single cycles, multiple cycles, hybrids; up to four connected components; up to 80 nodes), embedded in R^2 to R^5, thickened into tubes and sampled with Gaussian noise, with difficulty axes noise ratio [0.02, 0.20], separation [0, 0.6], density [3.5, 25] and thickness [0.45, 1.0] (Appendix D.2). Each method outputs a graph S, which is reduced by contracting degree-2 chains and compared with the reduced ground-truth graph via (i) Wasserstein distance between edge-length-filtration persistence diagrams, converted to a similarity (higher is better), and (ii) approximate graph edit distance from NetworkX (lower is better), both averaged over samples.","scReebTower builds a k = 15 nearest-neighbour graph, an adaptive Gaussian kernel and the row-normalized diffusion operator P, and uses the first non-trivial eigenvector (Fiedler vector) as a Morse function. It takes midpoint thresholds between sorted function values, makes each connected component of the crossing-edge level set a node, links nodes in adjacent levels that share points, and reduces the resulting Reeb graph. The multiscale 'scReebTower' first applies diffusion condensation X̂ = P^t X with k_smooth = min(80, n-1); 'scReebTower (ℓ=0)' skips it. Baselines: PAGA (Scanpy, Leiden resolution 1.0, k = 15, keep all edges with connectivity > 0) and Mapper (KeplerMapper, PCA lens of dim ≤ 2, 6 intervals, overlap 0.35, DBSCAN min_samples 3, eps = 1.5 × median 3rd-NN distance). Hyperparameters are fixed with no per-sample tuning."],"claim":"On the 1,000-sample synthetic track, multiscale scReebTower recovers reduced graphs with a mean graph edit distance of 13.32 vs. 16.25 for PAGA and 79.78 for Mapper, and a Wasserstein persistence similarity of 0.3199 vs. 0.2788 for PAGA and 0.0318 for Mapper.","source":"Table 1, page 9","metric":"Mean approximate graph edit distance between reduced recovered and ground-truth graphs (lower is better) and mean Wasserstein persistence similarity (higher is better) over 1,000 synthetic samples; no error bars or seeds reported","comparison":"scReebTower (with diffusion condensation) vs. scReebTower (ℓ=0), PAGA and Mapper on the same synthetic point clouds","falsification":"On a regenerated synthetic benchmark following Appendix D.2, scReebTower's mean GED is not clearly below PAGA's (e.g. the gap shrinks to well under ~3 or reverses), or its persistence similarity is not above PAGA's.","caption":"Table 1: Benchmark results on the synthetic track. Higher is better for Wasserstein persistence similarity, lower is better for graph edit distance","image":"img/2605.12662.png","kind":"table","image_w":1144,"image_h":205,"pwc":true}
|
site/p/2605.14152.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.14152","title":"ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety","authors":"Michael S Lee; Yash Maurya; Drew Rein; Bert Herring; Jonathan Nguyen; Kyungho Song; Udari Madhushani Sehwag; Jiyeon Cho; Kaustubh Deshpande; Yeongkyun Jang; Jiyeon Joo; Minn S Choi; Evi Fuelle; Christina Q Knight; Joseph Brandifino; Max Fenkell","openreview_id":"2PpebPNdF3","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Multilingual LLM safety is usually tested by translating harmful prompts into other languages, and prior work often finds that translation into lower-resource languages weakens refusals. Translation keeps the scenario and its entities fixed, though, so it cannot separate the effect of the language from the effect of the geopolitical setting (which country's agencies, places and events the attack refers to). For national-security and public-safety (NSPS) risks, both factors change at once when a model is deployed in another country, so it matters which one drives the change in safety behavior.","method":["ROK-FORTRESS contains 1,235 expert-written adversarial NSPS prompts (CBRNE, political violence, criminal/financial activity, information leakage), each with a prompt-specific rubric of 1-7 binary harm criteria. Human translators produce matched variants: culture-agnostic tasks (565) exist in English (V_En) and Korean (V_Ko); culture-specific tasks (670) exist in four variants, English/Korean language crossed with US/Korean entities (V_En,US, V_Ko,US, V_En,KR, V_Ko,KR). Each of 14 models (Claude Opus/Sonnet 4.5, GPT-5.2, Gemini 3 Pro, o4-mini, DeepSeek V3.2, Kimi K2, Llama 4 Maverick, Mistral Large 3, Qwen3-235B, HyperCLOVA X, Kanana 2, K-EXAONE, Solar Open) answers every variant.","A 3-judge LLM panel (HyperCLOVA X + o4-mini + Kimi K2 [medium reasoning], majority vote, calibrated against human labels at F1 0.83) marks each rubric criterion. Criteria are weighted by risk tier (1-3), and the per-task score is the weighted fraction of criteria judged harmful (0 for a safe refusal). The Tier-weighted Risk Score (TRS) is the mean over tasks times 100. The linguistic drop is Delta_ling = TRS(English) - TRS(Korean), computed as paired per-task differences. The contextual drop Delta_ctx averages TRS(US-grounded) - TRS(Korean-grounded) over both languages. 95% CIs are percentile bootstraps over paired (task, model) differences."],"claim":"Pooled over 14 models and all tasks, Korean-language variants of the same adversarial prompts have lower tier-weighted risk than their English counterparts: Delta_ling = +10.21 pp (95% CI [+9.33, +11.06]). The drop is positive in each of the four domains (+7.90 to +12.89 pp), and it is about 2.3x the contextual drop from swapping US for Korean entities (+4.47 pp, [+3.77, +5.18]).","source":"Table 19, page 40 (Appendix E.8); per-model view in Figure 3a and Figure 2, pages 10-12","metric":"Tier-weighted Risk Score (TRS, 0-100, lower = less harmful content); Delta_ling = TRS(English) - TRS(Korean) in percentage points, positive = Korean is safer. Brackets are percentile bootstrap 95% CIs over paired (task, model) differences.","comparison":"Same adversarial intents and rubrics in English vs. human-translated Korean (V_En vs V_Ko, and V_En,US vs V_Ko,US / V_En,KR vs V_Ko,KR), same models, same judge panel; secondarily, US-grounded vs. Korean-grounded variants holding language fixed.","falsification":"Re-running the models on the matched English/Korean variants, with the paper's TRS scoring, gives a pooled Delta_ling whose 95% CI includes zero or is negative, i.e. Korean prompts are not less harmful than English ones. A pooled Delta_ling not clearly larger than Delta_ctx would refute the secondary 'language dominates context' part.","caption":"Table 19. Per-domain Delta_ling and Delta_ctx (paired per-task TRS differences, in percentage points). Aggregated across 14 models with percentile bootstrap 95% CIs. N is the number of paired (task, model) differences contributing to the cell.","image":"img/2605.14152.png","kind":"table","image_w":1411,"image_h":405,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.14152","title":"ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety","authors":"Michael S Lee; Yash Maurya; Drew Rein; Bert Herring; Jonathan Nguyen; Kyungho Song; Udari Madhushani Sehwag; Jiyeon Cho; Kaustubh Deshpande; Yeongkyun Jang; Jiyeon Joo; Minn S Choi; Evi Fuelle; Christina Q Knight; Joseph Brandifino; Max Fenkell","openreview_id":"2PpebPNdF3","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Multilingual LLM safety is usually tested by translating harmful prompts into other languages, and prior work often finds that translation into lower-resource languages weakens refusals. Translation keeps the scenario and its entities fixed, though, so it cannot separate the effect of the language from the effect of the geopolitical setting (which country's agencies, places and events the attack refers to). For national-security and public-safety (NSPS) risks, both factors change at once when a model is deployed in another country, so it matters which one drives the change in safety behavior.","method":["ROK-FORTRESS contains 1,235 expert-written adversarial NSPS prompts (CBRNE, political violence, criminal/financial activity, information leakage), each with a prompt-specific rubric of 1-7 binary harm criteria. Human translators produce matched variants: culture-agnostic tasks (565) exist in English (V_En) and Korean (V_Ko); culture-specific tasks (670) exist in four variants, English/Korean language crossed with US/Korean entities (V_En,US, V_Ko,US, V_En,KR, V_Ko,KR). Each of 14 models (Claude Opus/Sonnet 4.5, GPT-5.2, Gemini 3 Pro, o4-mini, DeepSeek V3.2, Kimi K2, Llama 4 Maverick, Mistral Large 3, Qwen3-235B, HyperCLOVA X, Kanana 2, K-EXAONE, Solar Open) answers every variant.","A 3-judge LLM panel (HyperCLOVA X + o4-mini + Kimi K2 [medium reasoning], majority vote, calibrated against human labels at F1 0.83) marks each rubric criterion. Criteria are weighted by risk tier (1-3), and the per-task score is the weighted fraction of criteria judged harmful (0 for a safe refusal). The Tier-weighted Risk Score (TRS) is the mean over tasks times 100. The linguistic drop is Delta_ling = TRS(English) - TRS(Korean), computed as paired per-task differences. The contextual drop Delta_ctx averages TRS(US-grounded) - TRS(Korean-grounded) over both languages. 95% CIs are percentile bootstraps over paired (task, model) differences."],"claim":"Pooled over 14 models and all tasks, Korean-language variants of the same adversarial prompts have lower tier-weighted risk than their English counterparts: Delta_ling = +10.21 pp (95% CI [+9.33, +11.06]). The drop is positive in each of the four domains (+7.90 to +12.89 pp), and it is about 2.3x the contextual drop from swapping US for Korean entities (+4.47 pp, [+3.77, +5.18]).","source":"Table 19, page 40 (Appendix E.8); per-model view in Figure 3a and Figure 2, pages 10-12","metric":"Tier-weighted Risk Score (TRS, 0-100, lower = less harmful content); Delta_ling = TRS(English) - TRS(Korean) in percentage points, positive = Korean is safer. Brackets are percentile bootstrap 95% CIs over paired (task, model) differences.","comparison":"Same adversarial intents and rubrics in English vs. human-translated Korean (V_En vs V_Ko, and V_En,US vs V_Ko,US / V_En,KR vs V_Ko,KR), same models, same judge panel; secondarily, US-grounded vs. Korean-grounded variants holding language fixed.","falsification":"Re-running the models on the matched English/Korean variants, with the paper's TRS scoring, gives a pooled Delta_ling whose 95% CI includes zero or is negative, i.e. Korean prompts are not less harmful than English ones. A pooled Delta_ling not clearly larger than Delta_ctx would refute the secondary 'language dominates context' part.","caption":"Table 19. Per-domain Delta_ling and Delta_ctx (paired per-task TRS differences, in percentage points). Aggregated across 14 models with percentile bootstrap 95% CIs. N is the number of paired (task, model) differences contributing to the cell.","image":"img/2605.14152.png","kind":"table","image_w":1411,"image_h":405,"pwc":true}
|
site/p/2605.14847.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"id":"2605.14847","title":"SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation","authors":"Ivan Molodetskikh; Kirill Malyshev; Mark Mirgaleev; Nikita Zagainov; Evgeney Bogatyrev; Dmitriy Vatolin","openreview_id":"GJzL4oNx1s","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Modern GAN- and diffusion-based super-resolution (SR) models produce sharp images but also introduce local artifacts (hallucinated textures, garbled text, distorted patterns). Existing SR artifact datasets and detectors (DeSRA, LDL, PAL4VST) treat artifacts as binary masks, although some defects are obvious to viewers and others are almost invisible. The paper proposes 'artifact prominence', the fraction of crowd viewers who judge a highlighted region to contain a noticeable artifact. It releases SR-Prominence, 3,935 masks across DeSRA, Open Images, Urban100 and Urban100-HR, each annotated by 30 crowd workers, and uses it to audit which detectors and image-quality metrics actually track what viewers notice.","method":["Each artifact mask comes with an LR input, a 4x SR output, the HR reference (where one exists) and a prominence value in [0, 1]. Prominence is the share of the 30 crowd workers (after quality-control filtering) who said the highlighted region contains distorted objects or textures. Masks in the Open Images and Urban100 components were proposed automatically by thresholding detector heatmaps, and before annotation they were morphologically preprocessed (open 25x25, dilate 64x64 circular, close 25x25).","Threshold-free score (Section 5.2): run a detector or metric on every SR output to get a spatial heatmap (full-reference metrics compare against the original HR image; SSIM gives a per-pixel map, DISTS is computed block-wise on 16x16 blocks; binary detectors give their binary mask). For each annotated mask, erode it back to undo the preprocessing dilation, then compute (median heatmap inside the mask) minus (median heatmap outside the mask). The score is the Spearman correlation between this contrast and the crowdsourced prominence over all masks in a dataset component. The comparison includes full-reference metrics (SSIM, DISTS, LPIPS, ERQA), SR artifact detectors (DeSRA, LDL), no-reference artifact segmenters (PAL4VST, PAL4Inpaint) and no-reference IQA models (TOPIQ, PaQ-2-PiQ)."],"claim":"With the original HR image as reference, full-reference SSIM and DISTS heatmaps correlate with crowdsourced artifact prominence far better than no-reference methods: on Prominence-OpenImages SRCC is 0.312 (SSIM) and 0.415 (DISTS) vs. at most 0.136 for PAL4VST, PAL4Inpaint, TOPIQ and PaQ-2-PiQ; on Prominence-Urban100 it is 0.305 and 0.236 vs. at most 0.156.","source":"Table 5, page 9","metric":"Spearman rank correlation (higher is better) between the per-mask contrast (median heatmap inside the eroded mask minus median outside) and the crowdsourced prominence, over all masks in a component; no error bars or CIs reported for individual correlations","comparison":"Full-reference SSIM and DISTS (original HR reference) vs. no-reference methods PAL4VST, PAL4Inpaint (binary masks), TOPIQ and PaQ-2-PiQ, on the Open Images and Urban100 components","falsification":"Recomputing the score on the released data with the original HR reference, SSIM or DISTS fails to exceed the best no-reference method (PAL4VST/PAL4Inpaint/TOPIQ/PaQ-2-PiQ) by a clear margin (e.g. at least 0.1 SRCC) on Open Images or Urban100, or the no-reference IQA models show clearly positive correlations comparable to SSIM/DISTS.","caption":"Table 5: Threshold-free prominence score described in Section 5.2. RLFN columns use the pseudo-GT protocol from Section 5.3.","image":"img/2605.14847.png","kind":"table","image_w":1133,"image_h":438,"pwc":
|
|
|
|
| 1 |
+
{"id":"2605.14847","title":"SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation","authors":"Ivan Molodetskikh; Kirill Malyshev; Mark Mirgaleev; Nikita Zagainov; Evgeney Bogatyrev; Dmitriy Vatolin","openreview_id":"GJzL4oNx1s","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Modern GAN- and diffusion-based super-resolution (SR) models produce sharp images but also introduce local artifacts (hallucinated textures, garbled text, distorted patterns). Existing SR artifact datasets and detectors (DeSRA, LDL, PAL4VST) treat artifacts as binary masks, although some defects are obvious to viewers and others are almost invisible. The paper proposes 'artifact prominence', the fraction of crowd viewers who judge a highlighted region to contain a noticeable artifact. It releases SR-Prominence, 3,935 masks across DeSRA, Open Images, Urban100 and Urban100-HR, each annotated by 30 crowd workers, and uses it to audit which detectors and image-quality metrics actually track what viewers notice.","method":["Each artifact mask comes with an LR input, a 4x SR output, the HR reference (where one exists) and a prominence value in [0, 1]. Prominence is the share of the 30 crowd workers (after quality-control filtering) who said the highlighted region contains distorted objects or textures. Masks in the Open Images and Urban100 components were proposed automatically by thresholding detector heatmaps, and before annotation they were morphologically preprocessed (open 25x25, dilate 64x64 circular, close 25x25).","Threshold-free score (Section 5.2): run a detector or metric on every SR output to get a spatial heatmap (full-reference metrics compare against the original HR image; SSIM gives a per-pixel map, DISTS is computed block-wise on 16x16 blocks; binary detectors give their binary mask). For each annotated mask, erode it back to undo the preprocessing dilation, then compute (median heatmap inside the mask) minus (median heatmap outside the mask). The score is the Spearman correlation between this contrast and the crowdsourced prominence over all masks in a dataset component. The comparison includes full-reference metrics (SSIM, DISTS, LPIPS, ERQA), SR artifact detectors (DeSRA, LDL), no-reference artifact segmenters (PAL4VST, PAL4Inpaint) and no-reference IQA models (TOPIQ, PaQ-2-PiQ)."],"claim":"With the original HR image as reference, full-reference SSIM and DISTS heatmaps correlate with crowdsourced artifact prominence far better than no-reference methods: on Prominence-OpenImages SRCC is 0.312 (SSIM) and 0.415 (DISTS) vs. at most 0.136 for PAL4VST, PAL4Inpaint, TOPIQ and PaQ-2-PiQ; on Prominence-Urban100 it is 0.305 and 0.236 vs. at most 0.156.","source":"Table 5, page 9","metric":"Spearman rank correlation (higher is better) between the per-mask contrast (median heatmap inside the eroded mask minus median outside) and the crowdsourced prominence, over all masks in a component; no error bars or CIs reported for individual correlations","comparison":"Full-reference SSIM and DISTS (original HR reference) vs. no-reference methods PAL4VST, PAL4Inpaint (binary masks), TOPIQ and PaQ-2-PiQ, on the Open Images and Urban100 components","falsification":"Recomputing the score on the released data with the original HR reference, SSIM or DISTS fails to exceed the best no-reference method (PAL4VST/PAL4Inpaint/TOPIQ/PaQ-2-PiQ) by a clear margin (e.g. at least 0.1 SRCC) on Open Images or Urban100, or the no-reference IQA models show clearly positive correlations comparable to SSIM/DISTS.","caption":"Table 5: Threshold-free prominence score described in Section 5.2. RLFN columns use the pseudo-GT protocol from Section 5.3.","image":"img/2605.14847.png","kind":"table","image_w":1133,"image_h":438,"pwc":true}
|