ResearchRubric-Supervised Critic Plus 15.9 SWE-bench 83 Percent Fewer AttemptsarXiv·high signal24 behavioral rubrics from human-agent traces. Semi-supervised critic enables efficient coding agent training with early stopping.SourceSource pagearXiv↳ Follow the threadPolicy dependency / ContrastMake context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 pointsarXiv 2607.05378Stack layer'When Is Routing Meaningful?' Fewer Than 10 Agents Recover Most Diversity in Model SocietiesarXivPolicy dependency / Stack layerReason Less, Verify More: Deterministic Gates Catch a Silent Policy-Violation Failure in Tool-Using AgentsarXivThreat patterngspDAG-FL Secures Decentralized Federated Learning With Gossip and Virtual VotingarXivStack layer / Threat patternMitigating Taint-Style MCP Server Vulnerabilities Without Touching Server Code (SPELLSMITH)arXivStack layer / Update threadSLBench Finds Up to 70% Unsafe Execution When Codex and Claude Code Follow Logical Relations in Skill FilesarXivFollow-up threadLLM Personas Generate Executable Smart-Home Schedules to Replace Privacy-Risky Field DataarXivSame source domain / Semantic neighbor'Hindsight Gating' Cuts Communication in Bandwidth-Constrained Cooperative AgentsarXiv