skill-evolution

CausalEvo: Causal Skill Self-Evolution for Health LLM Agents 提出了一种因果技能自进化方法，通过在固定模型权重和规则数量下对比不同技能集的效果来决定保留或替换技能，在 MedCalc-Bench-Verified 六项任务上达到 60.00% 准确率，优于 SkillOpt（55.83%），且训练 token 消耗减少 55.6%。

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。

SkillEvoReg introduces a regularization framework to prevent overfitting in the skill evolution of language-model agents, combining dropout, local regularization, and causal validation to maintain performance while controlling skill growth.

SkillDRE introduces a dual-stage feedback loop for autonomously evolving malicious skill packages in agents, achieving high attack success rates while maintaining benign task functionality.

This paper introduces SkillPivot, a deviation-guided framework for skill self-evolution in LLM agents that improves skills by contrasting failed and successful trajectories from a common prefix.

Sentient's new EvoSkill v2 is an open-source framework that evolves agent skills from failed attempts, demonstrating how AI coaches can exploit reward hacking and highlighting the need for separation of powers and strong sandboxing in evaluation.

This paper introduces EvoSkill-GUI, a training-free framework that allows GUI agents to improve skills through in-execution reflection, revision, and reuse, demonstrating performance gains on multiple benchmarks without retraining.

WikiSkill is a framework that co-evolves agent skills with a persistent knowledge base, consistently outperforming state-of-the-art methods and enabling effective skill transfer across models.

A paper from Google discusses separating skill-evolution systems into components like raw execution traces and a persistent wiki of knowledge for AI agents, related to self-improvement concepts.

This article introduces how Warp uses Claude to build self-improving AI agents, automatically optimizing skills through human feedback, and summarizes best practices for effective agent development.

A Google paper discusses maintaining agent skills by separating raw execution traces from persistent knowledge in a wiki-like structure.

WikiSkill introduces a framework that co-evolves agent skills with a persistent knowledge base to systematically accumulate experience and improve performance across models, demonstrating benefits over state-of-the-art skill-evolution methods.

This article introduces the Hermes Must-Install List, featuring plugins like the all-in-one plugin suite, Web Command Center, model-switching wizard, and skill self-evolution tools, aimed at enhancing Hermes' autonomy and productivity.

This paper introduces SkillMisevo-Gym and SkillMisevo-Bench to study how self-improving LLM agents can evolve unsafe skills from compromised experience, plus SafeEvolve as a mitigation wrapper. Experiments across 25 agent-method configurations show skill misevolution is widespread and can persist across sessions, though SafeEvolve reduces fresh-session harm significantly.

The paper introduces DIVE, a diversity-driven framework that enables frozen LLMs to self-improve by evolving persistent natural-language skills from task experience and verifier feedback, without parameter updates. It outperforms existing methods on math and logical reasoning tasks and transfers across model scales.

SkillEvo introduces a method to continuously improve AI agent skills by using multi-turn interaction feedback and governance layers to maintain evolution gradients, surpassing self-reflection and single-turn QA-driven approaches.

This paper introduces SESA, a self-evolving skill-augmented search agent that co-evolves task generation and skill memory via tool-augmented search self-play. It improves accuracy across seven QA benchmarks over baselines while supporting memory-free deployment.

Introduces KnowAct-GUIClaw, a framework for personal GUI assistants with self-evolving memory and skill, achieving state-of-the-art performance on the MobileWorld benchmark and outperforming closed-source models like GPT-5.5.

This paper proposes a method for co-evolving evaluation metrics and skills in self-improving LLM agent systems, demonstrating that metrics can be evolved and that a co-evolution approach recovers most of the performance of a ground-truth-driven oracle across code generation, text-to-SQL, and report generation tasks.

ComfyClaw is an agentic skill evolution framework for ComfyUI image generation workflows, using typed graph editing and region-level VLM verifiers to translate visual failures into repair suggestions, outperforming baselines across multiple configurations.