New AI 'Trace' Turns Botched Attempts Into Reliable Step-by-Step Skills

Researchers built a smarter way for AI agents to learn from their own past attempts. Their system, Trace, takes sparse-reward trajectories full of failures, loops, and detours and compiles them into executable Walkthroughs: compact, state-conditioned, verifiable procedures that encode entry conditions, ordered state-action-effect steps, and completion and failure predicates. It detects progress anchors from rewards and persistent state changes, propagates credit to find valuable transitions, and estimates action prerequisites from cross-episode success and failure evidence, then slices backward through dependencies to keep only dependency-consistent action chains. On J-TTL, WebShop, and ScienceWorld with three open-source LLMs, Trace consistently outperformed eight test-time learning and memory baselines, improving average AUC by 30.0% and Final-3 by 40.5% over the strongest baseline while using fewer inference tokens. The takeaway from the authors: long-horizon interaction benefits more from state-conditioned executable procedures than from complete trajectories or abstract summaries.