Fetching from the wire…
Vibe Coding2026-09-04 · source-backed
Tool-using agents lose wall-clock to serial action-observation turns, not just inference. SMC runs a large authoritative actor producing the official trajectory while a faster drafter continuously predicts and executes future action chains on an isolated environment snapshot, matching against a macro library of recurring skeletons mined from training traces. When the actor's next call matches the first drafted action, SMC commits the remaining pre-executed steps and their observations. Qwen3.5-27B INT4 actor with Qwen3.5-4B drafter matches sequential accuracy at 18.59% lower latency on Tau^2-Bench Telecom and 44.9% lower wall time on AppWorld with a small completion drop. Code is public. arXiv 2609.03236
Each link below shares sources, entities, or timing with this story.
SkillForge (arXiv 2608.24747) notes that skill-extraction approaches like SkillRL never verify whether a stored skill still works against the current environment, so the bank grows monotonically while quality rots. It makes skill usage explicit during interaction so RL optimiz...
When an agent consolidates an external observation into long-term memory, attach platform-controlled metadata recording the source's trust level, then gate tool execution by matching action risk against supporting-memory authority. Laundered memories hit a 1.000 attack success...
Agent Lightning v1.0 (arXiv 2608.17528) inverts the standard agentic RL architecture, and the inversion is the whole point. Normally the training engine owns the environment loop. It drives the agent, collects trajectories, computes rewards. Which means your training setup and...
This inverts speculative decoding. A small open-weight draft model scores a black-box agent's already-generated trajectory in one forward pass, needing no logits, weights, activations or repeated sampling (arXiv 2609.05274). Phase-aware features separating reasoning spans from...
Two thirds. Not two thirds of a contrived jailbreak set. Two thirds of realistic malicious issue requests, against the exact three tools most of the people reading this run daily. Ankur Singh, Jinqiu Yang, and Tse-Hsun Chen built IssueTrojanBench across four attack categories...
arXiv 2607.08010 replaces the inference-time coding loop with a pipeline that collects execution traces, observes live backend schemas and values, synthesizes candidate tools, and repairs them against labeled cases. Runtime calls the compiled tool instead of re-deriving it. Th...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.