ResearchEsoLang-Bench: LLM Coding Benchmarks Measure Memorization Not ReasoningarXiv·high signalNovel benchmark using esoteric languages proves massive gap between mainstream vs esoteric performance. Current benchmarks largely measure memorization.SourceSource pagearXiv↳ Follow the threadStack layer / ContrastSuper Weights Don't Generalize, and Training Them in Isolation Destroys the ModelarXivThreat pattern / ContrastOut of Sight Defends Web Content by Attacking Agent Context CompressionarXivStack layer / ContrastTool-Making Agents Compile Repeated SOP Steps Into Versioned Tools Before DeploymentarXivStack layer / Update threadCodeTracer Performs Forensic Attribution of Backdoored Code CompletionsarXivStack layerMCP-Grounded Agent Pipeline Converts Legacy Docs to NIST OSCAL for Continuous CompliancearXivStack layer / Threat patternTRACE Watermarks Agent Trajectories Against Adversaries Who Control the LogsarXivStack layer / Threat patternOpen-Source Multi-Agent Firewall Intercepts Both HTTP(S) and WebSocket LLM TrafficarXivPolicy dependency / Stack layerHalluSquatting weaponizes LLM hallucinations to push botnet malware into agentsThe Hacker News