ResearchControllable Reasoning Models Are Private ThinkersarXiv·high signalTraining reasoning models to follow instructions in traces improves privacy by up to 51.9 percentage pointsSourceSource pagearXiv↳ Follow the threadStack layer / Update threadSLBench Finds Up to 70% Unsafe Execution When Codex and Claude Code Follow Logical Relations in Skill FilesarXivStack layer / Follow-up threadExplanation-Guided Training Steers Models Toward Faithful Partial-Dependence ExplanationsarXivThreat pattern / Follow-up threadMalaika Grounds Agentic Malware Analysis in Domain, Program Semantics, and Threat IntelarXivPolicy dependency / Stack layerReason Less, Verify More: Deterministic Gates Catch a Silent Policy-Violation Failure in Tool-Using AgentsarXivStack layer / Threat patternMitigating Taint-Style MCP Server Vulnerabilities Without Touching Server Code (SPELLSMITH)arXivFollow-up threadS²AE Makes Sparse Autoencoders Learn Modality-Consistent Concepts in VLMsarXivStack layer / Update threadDominoTree Improves Speculative Decoding With Path-Dependent Tree DraftingarXivPolicy dependency / ContrastMake context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 pointsarXiv 2607.05378