ResearchCalibrated Speculative Decoding: Frequency-Guided Selection Achieves 2.33x Peak ThroughputarXiv·medium signalXBlueskyLinkedInCopy linkConventional speculative decoding suffers from frequent false rejections when draft models produce semantically correct but lexically divergent tokens. CSD adds Online Correction Memory (aggregates historical rejection patterns as rescue candidates) and Semantic Consistency Gating (probability-ratio verification instead of exact token matching). Achieves 2.33x peak throughput speedup over baselines.SourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerSkillRise Collapses Extract-Retrieve-Execute Into One RL Loop: a Single Policy Alternates Between Solving Tasks and Rewriting Its Own Skill DocumentarXiv / HuggingFace Daily PapersPolicy dependency / Stack layerStop authorizing tool calls off the model's own rationale — convert it to typed claims and check them against server-held factsarXiv 2607.25364Policy dependency / Stack layerYour CLAUDE.md doesn't actually constrain the agent: best model passes only 36.2% of long-horizon policy-compliance tasksarXiv 2607.25398Policy dependency / Stack layerCROSS-CATEGORY: Three Independent July 28-29 Artifacts All Say Agents Cannot Be Governed by Written InstructionsMultiple Sources (arXiv 2607.25398, Enklype Salt, SaaStr)Stack layer / Update threadMetis moves agent memory inside the backbone: gradient-free updates via a single forward pass, with frozen weights at inferencearXiv 2607.26760Stack layer / ContrastStop feeding failed code back to the model — a placebo-controlled test shows blind resampling beats self-repair at 2.5–5.5x lower token costarXiv 2607.26117Stack layer / Update threadDomain-conditioned prompting does not make generated code more secure — pick a different model insteadarXiv 2607.25225Stack layer / ContrastCoRT Gets Token-Level Credit From Rubric-Guided RL Without Training a Second Model, Adding 4.4 Points Over Response-Level GRPOarXiv / HuggingFace Daily Papers