• A 7B Model Uses Search to Offset Limited Capacity. ZGCM-1 combines internal reasoning, external tools, and a 256K context. Its 16K pretraining reaches the same loss about 4.2 times faster.

• Safety Oversight Must Cover Real Execution Traces. HazardAuditor normalizes runtime interactions across frameworks. Its abstract reports gains of up to 16.5 percentage points over the strongest existing guard.

• Understanding, Action, and Prediction Can Share One Sequence Objective. PhysBrain 1.5 turns human interaction videos into embodied supervision. It encodes answers, end-effector motion, and future visual states as discrete sequences.

• Executable Reconstruction Separates Perception From Factual Accuracy. BVB’s best model scores 88.6 on perceptual similarity but preserves only 53.7% of the original video’s spatiotemporal facts.

Also Notable

• Research Agents Are Turning Tool Use Into Verifiable Experience Pipelines — Atria Dawn Preview focuses on production training loops instead of stopping at the “superintelligence” label.
• Recursive Self-Improvement May Require Co-Adapting Exploration and the World — Dream-RSI studies high-value solution discovery inside changing worlds and search spaces.
• Multi-Agent Systems Can Build Causal Memory Without Retraining — RSIAgent combines curriculum, execution, and verification agents to adapt to new interfaces after deployment.
• Explicit Controls May Reduce Video Drift and Character Distortion — LynnReal-Omni combines diffusion-based image quality with controls from 3D scenes and game states.
• Discovery Models Must Ask Questions, Invent Representations, and Explain — Discovery Foundation Models offers a conceptual framework for discussing discovery intelligence.
• Short GUI Actions Directly Test Blockwise Diffusion Latency — LLaDA-UI asks whether parallel generation produces real gains in interface interaction.
• Image Swaps Reveal Whether Medical VLMs Read Scans — ModaLens measures image sensitivity more directly than aggregate accuracy.
• Combining Reasoning Strengths May Beat Choosing One Model — Lightning Weave combines model capabilities to improve the accuracy-cost frontier.
• Frozen Models May Already Contain a Native Skill Router — The Router Within avoids filling the context with skill metadata or adding a separate retriever.
• Visual RAG Must Select and Combine Sparse Evidence — Explicit context selection and merging may work better than expanding one-shot input.
• Design Agents Need Different Concepts, Not Higher Temperature — This work adds structured creative exploration to natural-language-to-interface and frontend-code generation.

Read the full edition →