• User Edits Cut Agent Solve Rates by 7.7 Percentage Points: SWE-Touch finds that nine coding models often preserve conflicts or overwrite user work. Collaboration tests need change detection, ownership checks, and regression testing.

• GradCuit Gives Hidden-State Updates a Traceable Gradient Path: Across five models and three reasoning benchmarks, it averages 64.5% accuracy. That beats chain-of-thought prompting by 6.6 percentage points, but gradient attribution does not explain reasoning.

• UEmbed Unifies Sparse, Dense, and Multimodal Retrieval in One 9B Model: It scores 71.8 for dense and 71.0 for sparse retrieval on MMEB-v2. Its value still depends on latency, indexing costs, and deployment complexity.

Also Notable

• World Models Must React Plausibly, Not Just Generate Realistic Scenes — WorldExam tests whether models learn worlds whose states can change coherently.
• Flood Segmentation Gets a Large Cross-Sensor Testbed — GEOID-Flood pairs bi-temporal SAR with registered optical imagery to test geographic foundation models across regions and sensors.
• SKT Scales Skill Training With 27,164 Verified Trajectories — Its synthetic tasks train the full process of selecting and executing single or multiple skills.
• Agents May Keep Searching After Their Internal Map Finds the Answer — ScrambleToolBench exposes the gap between discovering an environment and acting on that knowledge.
• Human Motion Capture Shifts From Ideal Accuracy to Field Resilience — Sen-Cap directly tests missing sensors, changing layouts, and noisy inputs across LiDAR and cameras.
• Human Judgments Can Be Aggregated Before Calibration — This method offers theoretical guarantees when reviewers use inconsistent scales and no objective ground truth exists.

Read the full edition →