HKU's XLANG Lab and Alibaba's Qwen team have open-sourced FineVLA, a framework that refines language supervision in robot policy learning from "what to do" to "how to do it," reaching 86.8% success in RoboTwin simulation (+15) and 62.7/100 on a real dual-arm robot, with code, models, and benchmark all open-sourced.

OpenAI expands its Daybreak security program with the full release of GPT-5.5-Cyber (CyberGym 85.6%, ahead of Mythos 5), a Codex Security plugin update, and the Patch the Planet open-source remediation effort — pushing vulnerabilities all the way from discovery to landed patches.

Zhongguancun Academy × USTC × SJTU × EIT propose VLA-JEPA, fusing VLA models and world models in latent space: assembly tasks completed with just 13 trajectories, 97.2% on LIBERO, shared by LeCun and Saining Xie.

Fudan × Shanghai Innovation Institute × Oxford publish the ICML 2026 paper AutoControl-Arena, which auto-synthesizes executable test environments to surface latent Agent risks in long-tail scenarios, reproducing Anthropic/OpenAI safety-report behaviors with a 0.87 correlation.

At Automate 2026, Nvidia launched Halos for Robotics, a full-stack robot safety system that opens 18,600 engineering-years of autonomous-driving safety know-how to the embodied AI industry — the core safety framework is already open, with 43 companies on board.

ISCAS open-sources the PerfEvolve framework, converting static tuning docs into executable procedural skills for Agents and delivering up to 58.9% performance gains on PostgreSQL v16 — a direct fix for LLM tuning failures.