15 articles — 🔴 2 critical , 🟡 7 important , 🟢 6 interesting
Meta Superintelligence Labs has unveiled Muse Image and Muse Video — models that operate as agents, internally invoking code and web-search tools, ranking #2 and #3 on the Arena leaderboard, with mandatory Content Seal watermarking.
Researchers from Tsinghua University and Zhipu AI proposed Direct On-Policy Distillation (Direct-OPD) — a method that transfers reinforcement learning gains from a weaker teacher model to a stronger student without imitating the final policy, achieving a jump from 48.3% to 62.4% on AIME 2024.
Researchers from Stanford and UC Berkeley have identified verification — the ability to assess the correctness of a solution — as a new scaling axis for LLMs, orthogonal to existing paradigms. The framework generates continuous verification scores and achieves SOTA on four benchmarks: Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%).
Google has added background execution, remote MCP integration, custom function calling, and credential refresh to the Gemini API — four capabilities that enable developers to deploy agents beyond prototype environments.
Vivek Trivedy of LangChain argues that systematically improving AI agents is fundamentally a problem of mining execution traces at scale. The recommended sequence: harness engineering, then fine-tuning, then additional harness optimization. Key advice: deploy agents early to start the data collection loop.
Anthropic released Claude Code v2.1.203 on July 7 with a login token expiry warning, automatic recovery of frozen background sessions, and a roughly 7 MB reduction in binary size. Several bugs slowing macOS and disrupting sub-agent behavior have also been fixed.
Researchers have released AgentGym2, an evaluation framework testing 15 LLM agents under realistic conditions with incomplete information and unknown tools. All models, including GPT-5 and Gemini, struggle significantly — revealing a wide gap between the current state of technology and production deployment requirements. The paper was accepted at ACL 2026.
The AMD ROCm team published a technical guide on manually computing GPU occupancy for the CDNA4 architecture. Key finding: matrix-bound kernels on the MI355X achieve peak throughput at just 12% occupancy — a counterintuitive result for practitioners coming from the CUDA ecosystem.
The NVIDIA Vera CPU with Olympus cores delivers 50% higher IPC than the Grace processor and 3.4 TB/s core-to-core bandwidth. Perplexity reports 1.5x faster job completion on coding agent workflows.
The UK AI Security Institute used frontier models to audit its own research platform and found a five-step attack chain that had previously escaped standard tools — all for under £150 in token costs.
Researchers from Seoul National University and the Universities of Indiana and Wisconsin introduce Agent Data Injection (ADI) — a new attack type that embeds malicious data into trusted data structures of agents, achieving arbitrary click attacks and remote code execution on Claude Code, Codex, and Gemini CLI while bypassing existing defenses.
