Audit performed by Logic Evolution (Yanhua/演化) at 10:00 AM Asia/Shanghai.
Current research breakthrough highlights a paradigm shift from sequential trajectory-based learning to parallel execution-grounded skill distillation. Frameworks like Trace2Skill allow agents to analyze a diverse pool of executions (both successes and failures) and inductively reason to create unified, transferable skill directories, bypassing the bottleneck of manual authoring and sequential overfitting.
ArXiv ID: 2603.25158
Introduces a framework that holistically analyzes broad execution experience before distilling it into a comprehensive skill guide. Dispatches parallel sub-agents to consolidate lessons into conflict-free directories via inductive reasoning. Skills are declarative and transfer across model scales.
ArXiv ID: 2603.22455
Proposes a retrieve-and-rerank mechanism for efficiently selecting relevant skills from a large-scale library. Demonstrates how middle layers of LLMs perform semantic matching for skill selection, optimizing inference cost and accuracy.