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 layerHolding Back Ready Agent Turns Instead of Releasing Them Eagerly Cuts P95 Workflow Latency up to 3.50xarXiv 2609.10964Stack layer / Update threadMarket Concentration Barely Changes Model Collapse: Pushing One Model to 90% Share Moves Five-Generation Endpoints a Few PercentarXiv 2609.11146Policy dependency / Follow-up threadMOSAIC Picks a GraphRAG Traversal Policy per Query and Beats the Best Fixed Policy by 9.96 PointsarXiv 2609.11065Stack layer / Follow-up threadBenchmark Radar Ships a Daily-Updated Catalog of 1,283 AI Benchmarks With 12,916 Numeric Score ObservationsarXiv 2609.11115Stack layer / Threat patternClaude Code 2.1.269 Ships a Plugin Eval Runner and a Knob to Raise the Workflow Tool's Concurrent Agent Cap to 256Anthropic (claude-code CHANGELOG)Stack layerReqEvolve Takes Plain-Language User Requests to Running Code at 89.2% Pass@1 Across 72 Evolution CasesarXiv 2609.10590Stack layerScore retrieval sufficiency structurally before generation, not from the model's own confidencearXiv 2609.11023Stack layerActMap Turns a Generation's Whole Activation Trajectory Into a 96 KiB Tensor You Can Score in Under a MillisecondarXiv 2609.11498