Tag
AtomSD 提出了一种原子引导的自蒸馏框架，让 LLM 智能体从自身成功轨迹中提取细粒度的"过程原子"（Intent、When、Do、Avoid）并蒸馏进不断演化的策略中，实验显示其平均性能比最先进基线高出最多 9.31 分。
论文提出 MAGE 框架，利用用户手册作为可扩展的程序性知识来源，同时改进计算机使用代理（CUA）的推理与训练，在六个应用上提升了任务成功率并减少了对人工标注数据的依赖。
SkillNEO 提出非精英进化优化方法来改进 LLM agent 技能，通过允许低奖励版本继续迭代并分析版本间差异来保留有效步骤、修复失败，使 GPT-5.4 Mini 在 86 项 SkillsBench 任务上平均奖励从 49.1% 提升至 66.4%，且改进后的技能还能提升更强模型的表现。
本文提出 EVISKILL，一个证据驱动的框架，将 LLM 智能体的执行观察组织为可重放的证据卡（Replayable Evidence Cards），并通过定向重放验证技能编辑，从而在不更新模型参数的情况下实现可持续的技能进化。实验在三个交互式基准和六个 LLM 主干上验证了其有效性。
Code2Skill is an automated pipeline that extracts verifiable skills from source code to improve AI agent performance, producing a large grounded skill bank that enhances agents before interaction experience.
This paper introduces Scientific Agent Skills, an open-source library of 163 procedural knowledge skills across 16 scientific domains to help AI research agents perform defensible analyses.
CaSKG proposes a counterfactual-causal skill graph framework to calibrate procedural relations, enhancing compact and executable skill retrieval for LLM agents, and demonstrates improved performance on benchmarks like ALFWorld and ScienceWorld.
The paper presents ACES, a framework for continuous evaluation of AI agent skills through live trials, measuring Skill Lift to quantify added value, and demonstrating its effectiveness on enterprise repositories compared to scan-only gates.
FinSkillBench is an evaluation suite for AI agents in investment management, showing that curated procedural skills improve performance compared to self-generated skills.
This paper presents SkillJuror, a framework for evaluating how different organizational paradigms for Agent Skills affect LLM agent runtime behavior. A study on 82 tasks shows that Progressive Disclosure increases resource diversity and uptake events compared to a flat baseline, with task-dependent outcome improvements.
SKIM is an adaptive multi-resolution soft token compression framework that compresses procedural skills for LLMs, maintaining task performance while reducing prefill cost and latency.
This paper introduces SkeMex, a self-evolving framework that enhances medical agents by distilling interaction trajectories into structured skill memory, enabling better long-term clinical reasoning through context-dependent utility estimation and governance.
This paper introduces W2S, a framework that automatically constructs executable Skills for LLM agents from historical interaction traces using a Skill-IR intermediate representation, improving behavioral replay consistency by 10.5% over baselines.
This paper presents a prototype framework for managing uncertainty in LLM-generated procedural knowledge for virtual laboratory planning, using structured domain representations to repair uncertain procedural steps.
This paper introduces MMSkills, a framework for representing, generating, and using multimodal procedural knowledge for visual agents, combining textual procedures with visual state cards and keyframes, and demonstrates improvements in GUI and game-based visual agent benchmarks.