• Confidence Needs a Track Record: XConf matched or beat 10-sample self-consistency in 23 of 24 comparisons at one-tenth the generation cost.

• Scientific Code Can Become Training Infrastructure: ScienceIDE turns research software into trainable, testable environments using domain specifications and correctness criteria.

• Smooth Critics May Be Flattening Real Value: Value Flattening prevents PPO critics from capturing reward swings across reasoning trajectories, weakening their ability to reduce variance.

• Multi-Agent Research Needs Reproducible Lineage: Agora records research assets in an append-only Git DAG. Thirteen agents produced 165 independent replications in nearly 12 days.

• Demonstrations Can Work at Deployment Time: GPT-Policy lets VLMs adapt robot behavior from contextual examples while a constrained controller validates execution.

Also Notable

• Zing-0.5 Makes Real-Time Control a New Bar for Generated Worlds — The 5B autoregressive model accepts keyboard actions and live text instructions for playable environments.
• ComPO Uses Comparison Oracles for Preference Alignment — The zero-order approach seeks directional signals without relying on small likelihood differences or conventional gradients.
• PANORAMA Combines Complete Descriptions With Pixel-Level Grounding — Its mask-candidate selection method targets the tradeoff between descriptive coverage and segmentation accuracy.
• Shared Gaze May Reveal Mutual Understanding — The study unifies gaze labels across the MapTask and MUNDEX collaborative-task corpora.
• EvolveTrade Treats System Prompts as Adaptable Trading Policies — It explores whether agents can adjust research, tool use, verification, and risk controls as markets change.
• Long Tasks Need More Than Final Success Rates — Locating failures, recovery behavior, and irreversible damage gives teams clearer targets for reliability improvements.
• Shared Text Can Expose Dataset-Level Secrets — QuanText extends privacy threat models to properties such as gender, diagnosis, or opinion distributions.
• DualSQL Jointly Trains Schema Linking and SQL Generation — Multi-agent reinforcement learning tests whether the two modules can provide useful learning signals to each other.
• MoRE Combines Expert Mixtures With Weight Reuse — It aims to preserve expert capacity without making parameter counts grow linearly with the number of experts.
• Generalist–Expert Mixtures Target Rare Multimodal Pathologies — The study examines whether one model can preserve shared representations alongside modality-specific expertise.

Read the full edition →