• Safety Benchmark Scores May Reflect Policy Definitions More Than Capability. Ten guardrails shared the same architecture and optimization pipeline. Their score ranges were 0.99 on HarmBench-copyright and 0.043 on WildGuard-benign, a 23-fold difference. Soft labels from distilled teachers also transmitted policy.

• CoworkEnv Synthesizes Executable Training Environments for Long-Horizon Office Agents. Supervised fine-tuning with its resources improved Qwen3-VL-8B-Instruct by 31.9 points on Claw-Eval multi-turn tasks. Evidence for reinforcement learning remains preliminary.

• MT Forcing Controls Multi-Turn Editing Drift Across the Entire Trajectory. It supervises image quality with cumulative edit targets generated from the original image. Reference trajectories encourage consistency across turns, but the abstract does not report specific generalization results for long sequences or real user instructions.

Also Notable

• Completed Trajectories Can Train Small External Monitors to Stop Likely Failures Early. StopLoss preserves 99% of successful SWE-bench Verified runs while reducing cost per retained solution by 4.5%.
• FineSplat Reconnects High-Resolution Detail With Two Local Groupings. It combines within-view and 3D spatial grouping, then uses reconstruction-error feedback to correct one-shot feed-forward predictions. The method balances detail and compactness with the same or fewer Gaussian primitives.
• Video Reasoning Chains Need Explicit Spatiotemporal Evidence. TCoVP organizes entities, time ranges, and timestamped spatial evidence into reasoning chains. It converts selected evidence into box prompts and temporal support for video-mask prediction.
• JEPA Interpretability Must Explain How Variables Interact and Affect Predictions. The paper proposes a four-stage agenda: Ground, Structure, Verify, and Act. The final three stages still lack systematic study.
• Generate Stronger Single-Path Continuations From Identified Thought Switches. SteadyThought pairs new responses with originals for thought-level preference training. It cuts token use by up to 63.8% while preserving strong reasoning accuracy.
• Static Rubrics Lose Discrimination as Generation Policies Improve. RARE replaces only sample-level criteria that current outputs already satisfy completely. Across three video-script tasks, defect-targeted updates beat matched static rubrics by 3.82 points on average.
• Different Human and Model Errors Do Not Guarantee Complementarity. Across 35,613 judgments from 416 reviewers, tested assistance methods sometimes raised overall accuracy while reducing human corrections of AI errors.
• Observation Uncertainty Can Protect Important Tokens During Weather-Model Merging. On 72-hour ERA5 forecasts at 1.40625° resolution, UtoMe-E reduced overall error from 67.32 to 53.00. It used 50% fewer tokens and the same 0.55× encoder FLOPs.
• Similar Headline Scores Can Hide Different Medical Risk Calibration. Auto-MD-bench spans 15 tasks across 10 clinical datasets. It finds large model differences under shifts across populations, hospitals, and time.
• Personalized Memory Reduced Response-Distribution Entropy by 34.7% Versus Full Interaction History. Across several models and memory mechanisms, this average result comes from 1,000 benchmark cases built from preferences extracted from real user-assistant conversations.

Read the full edition →