SourcesSAGE — Self-Aware Guided Efficient Reasoning Reduces Inference CostarXiv·high signalXBlueskyLinkedInCopy linkByteDance. Reasoning models know when to stop thinking but sampling obscures it. SAGE selects concise correct paths via confidence.SourceSource pagearXiv↳ Follow the threadPolicy dependency / ContrastCORAL Puts an LLM Agent in a Live Closed Loop on Production Recommenders and Wins Two A/B TestsarXiv 2609.02730Stack layer / ContrastContext Privilege Escalation Attacks Hit 12 Real Agent Harnesses, Including Claude Code and CodexarXiv 2609.01222Stack layer / ContrastCordisBench measures whether a model can reason about the harness it is allowed to rewritearXivStack layer / Update threadControlled Experiment Confirms Representational Entanglement Causes Collateral Damage in UnlearningarXiv 2609.02285Threat pattern / ContrastA web-browsing agent leaks recoverable signal about records it silently loaded, with no jailbreak and no disclosurearXiv 2609.01663Policy dependency / Stack layerTyped Provenance Guardrails Block All 19 Unsafe Releases From a Persistent Agent's Autobiographical MemoryarXiv 2609.02127Stack layer / Update threadPersistent agents get an architecture that survives swapping the model, the harness and the hostarXivStack layer / ContrastReveree Scores Reverse-Engineering Agents Past Solve Rate and Finds Bigger, Newer Models Are Not Reliably StrongerarXiv 2609.01185