ResearchCUDA Agent Agentic RL for GPU Kernel Generation Outperforms Opus 4.5arXiv·high signalXBlueskyLinkedInCopy linkAgentic RL system generates optimized CUDA kernels outperforming torch.compile by 100% and Claude Opus 4.5 by 40% on hardest tasksSourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerAgentic Workflows Reveal Their Dependency Graph Only at Runtime — ASGE-RR Reserves Capacity for Calls That Haven't Happened YetarXiv 2608.06033Stack layer / Threat patternAn automated red-teamer plants instruction backdoors in customized coding LLMs at 94.5% success while evading detection entirelyarXiv 2608.05659Stack layer / ContrastGenerative Reward Models Underperform in RL Because They Rank, Not Score — RRC Fixes the MismatcharXiv 2608.06310Policy dependency / Stack layerAgentOPSD: Critic-Free Turn-Level Credit Assignment via Bayesian Belief Updates Hits 89.1% on ALFWorld With a 7B ModelarXiv / HuggingFace Daily Papers (57 upvotes)Stack layer / Threat patternTamarin-to-ProVerif Translation Across 121 Models: The Two Verifiers Agree on 246 of 247 Tasks, but ProVerif Is 6.7x FasterarXiv 2608.06315Stack layer / Threat pattern"Agentic Posture Vulnerability": A Vulnerability-Management Record for Agent Exposures That Have No CVEarXiv 2608.05884Policy dependency / Stack layerEnvACE trains tool-use agents by making the policy rehearse the environment instead of calling onearXivStack layer / ContrastFinEvo-Bench measures whether agents actually learn from experience — Letta scores 91.65, Codex gains the most at +19.37arXiv