ResearchCUDA Agent Agentic RL for GPU Kernel Generation Outperforms Opus 4.5arXiv·high signalAgentic RL system generates optimized CUDA kernels outperforming torch.compile by 100% and Claude Opus 4.5 by 40% on hardest tasksSourceSource pagearXiv↳ Follow the threadStack layer / Threat patternPiSAs Benchmarks Cross-User Privacy Leaks in Shared Agentic SystemsarXivStack layer / Update threadVisual Action Outcome Reasoning Alignment Improves VLM Physical ReasoningarXivStack layerVera Framework Auto-Generates Agent Safety Tests; 93.9% Attack Success on Production FrameworksarXivStack layerCoding Before Testing Biases an LLM's Own Test Suite Toward Its Buggy CodearXivStack layerSemi-formal reasoning: force the agent to state premises and trace execution before it verifies a patcharXivStack layerCoding Agents Internally Encode a Forward 'Programming Horizon'arXivPolicy dependencyTREK Rescues GRPO on Hard Prompts Outside the Student's On-Policy SupportarXivThreat patternLook-Ahead Bias Reframed as a Verifiable Non-Interference Property for Agentic TradingarXiv