递归自改进论文索引 Prism-Shadow/awesome-rsi

我们的综述

From Existing Methods to New Research Ideas: Two Perspectives on Recursive Self-Improvement

RSI 方法的九维分类体系，加上来自传统深度学习的洞见，帮你找到新的研究方向。

观看讲解视频 · 1 分钟

https://github.com/user-attachments/assets/70f488af-925c-4b93-ba03-5c84711b8dbf

从这里开始

Awesome RSI 汇集递归自我改进（Recursive Self-Improvement）的基准、方法与系统。我们关注这样一条回路：智能体执行任务，从轨迹和反馈中学习，更新自身状态，再把更新后的状态用于后续工作。这个状态可以是模型参数、harness 代码、上下文、记忆或技能。

115 篇方法论文 · 33 篇基准论文 · 3 个系统

| 你想找… | 从这里开始… |
| RSI 入门介绍 | 理解 RSI —— 图文指南，提供中英文版本。 |
| 可以继续做的方法 | 方法与系统 —— 按改进对象、拓扑、反馈等条件筛选。 |
| 衡量改进效果的方式 | 基准 —— 对比改进对象，以及在线、离线、混合三类评测协议。 |
| 论文之间的联系 | 引用图 —— 在方法与基准之间切换，搜索节点，缩放和平移。 |

论文地图

本 README 中的每份列表都沿一个维度组织：方法与系统按 RSI 改进对象划分，基准按 RSI 模式划分。想用更多方式浏览这个合集，请访问 Awesome RSI 网站，在那里可以对比更多维度、组合筛选条件，找到你感兴趣的工作。

智能体改进的是什么？ 下面这棵树按 RSI 改进对象维度展开。

模型参数 · Harness 代码 · 上下文 · 记忆 · 技能 · 其他产物 · 基准测试

这棵树展示的是合集里的示例。下面的列表沿用网站现有的标签；同一项工作可能出现在多个产物下。收录范围既包括显式的自我修改，也包括长程任务中由经验驱动的改进。

方法与系统

在网站上对比所有维度 →

按 RSI 产物分组，组内按首次发表或项目发布时间排序，最新的在前。会议标签对应合集里记录的会议；项目则链接到各自的仓库。

模型参数

更新学生的模型参数，以提升任务表现。这既包括教师借助学生的表现反馈来改进后续各轮训练——即使每轮都从初始模型重新开始，也包括更新后的学生继续从后续任务中学习。

| 论文或项目 | 发表信息 |
| ReSAIL: Mitigating Collapse in Iterative Agent Self-Distillation | arXiv preprint |
| Gödel Forest: Balancing Search Depth and Breadth for Data-Centric Recursive Self-Improvement | arXiv preprint |
| Recursive Self-Improvement via On-Policy Distillation for Reasoning | arXiv preprint |
| TISD: On-Policy Self-Distillation with Trajectory Intervention | arXiv preprint |
| Experience Funnel: A State-Policy Alternating Loop for Self-Evolving Agents | arXiv preprint |
| NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness | arXiv preprint |
| RISE: Recursive Improvement via Self-Extrapolating Policy Distillation | arXiv preprint |
| Reef: Continual Learning Infrastructure for Self-Improving Agents | Project · v0.0.2 |
| EnvHarness: Awakening Static Worlds for Agent Learning | arXiv preprint |
| HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry | arXiv preprint |
| ExpGraph：模型无关的经验学习，用图结构记忆服务 LLM 智能体 | arXiv preprint |
| Continual Harness：面向自我改进基础智能体的在线适应 | arXiv preprint |
| G-Zero：从零数据出发的开放式生成自博弈 | arXiv preprint |
| P²O：策略与提示词联合优化 | arXiv preprint |
| MM-Zero：从零数据自我演进的多模型视觉语言模型 | EMNLP 2026 |
| TTCS：面向自我演进的测试时课程合成 | arXiv preprint |
| 在极少人工监督下引导 LLM 自我演进 | arXiv preprint |
| VisPlay：从图像出发自我演进的视觉语言模型 | CVPR 2026 |
| R-Zero：从零数据自我演进的推理 LLM | ICLR 2026 |
| 自适应语言模型 | NeurIPS 2025 |

Harness 代码

改动可执行智能体、harness、控制流、自我改进或工具代码。

| 论文或项目 | 发表信息 |
| HarnessPro：递归式 Harness 自我改进，用 Pareto 优化权衡性能与成本 | arXiv preprint |
| Harness²：面向开放世界的递归智能体 Harness | arXiv preprint |
| RoboRSI：复杂真实环境中稳定、高效、可复用的机器人自我演进 | arXiv preprint |
| MESH-Harness：用 Bandit 引导的组合式演进实现自我改进的智能体 Harness | arXiv preprint |
| VERSE：面向智能体 Harness 的可验证自我演进优化器 | arXiv preprint |
| 从研究中学习：走向终身智能体 Harness 演进 | arXiv preprint |
| MoMHa：在准确率、安全性与 token 上对 LLM Harness 做多目标优化 | NeurIPS 2026 |
| SkillEvoReg：正则化智能体技能演进，防止过拟合 | arXiv preprint |
| HarnessPAI：面向物理 AI 的演进式 Harness | arXiv preprint |
| 长 Harness，而非长上下文：从无策略脚手架到可复用专家智能体 | arXiv preprint |
| Harness 即语言：极简却极具表达力的智能体框架 | arXiv preprint |
| AI 研究智能体的递归自我改进 | arXiv preprint |
| Harness-Zero：以智能体为 Harness 的 Harness 蒸馏 | arXiv preprint |
| RRSI：智能体运行框架的正则化递归自我改进 | arXiv 预印本 |
| 借助快速树搜索实现自我改进 | arXiv 预印本 |
| SoL-Pi：递归扩展自动研究循环，打造高效智能体运行框架 | arXiv 预印本 |
| EvoOntology：面向数据智能体的自演化本体层 | arXiv 预印本 |
| ModularRSI：模块化、可泛化的递归运行框架自我改进 | arXiv 预印本 |
| Dream-RSI：通过演化世界实现递归自我改进 | 预印本 |
| Ecdysis：高效训练 LLM 智能体运行时框架 | arXiv 预印本 |
| Code-to-Harness：从自我对弈中蒸馏黑盒优化器 | arXiv 预印本 |
| SimSkill：面向交通仿真的自演化 LLM 智能体，用于积累技能与知识 | arXiv 预印本 |
| Reef：面向自我改进智能体的持续学习基础设施 | 项目 · v0.0.2 |
| HarnessEvolve：从参考轨迹中学习，实现可靠的智能体自我演化 | arXiv 预印本 |
| Proteus：面向 AI 智能体的运行框架无关自我演化框架 | 项目 · v0.3.0 |
| ForeDreamer：面向未来事件预测的自演化双智能体记忆架构 | Findings of EMNLP 2026 |
| Mendel Gödel Machine：通过比较演化实现递归自我改进的编码智能体 | arXiv 预印本 |
| EvolveNet：面向智能体自我改进的协作式运行框架演化 | arXiv 预印本 |
| DarwinX：通过自然选择演化智能体运行框架 | arXiv 预印本 |
| HarnessBank：面向智能体运行框架自我演化的语义基因库搜索与门控验证 | arXiv 预印本 |
| The Red Queen Gödel Machine：协同演化智能体及其评估器 | arXiv 预印本 |
| HarnessX：可组合、自适应、可演化的智能体运行框架工坊 | arXiv 预印本 |
| Self-Harness：自我改进的运行框架 | arXiv 预印本 |
| 从失败轨迹到可靠的 LLM 智能体：诊断并修复运行框架缺陷 | arXiv 预印本 |
| OpenSkill：面向 LLM 智能体的开放世界自我演化 | arXiv 预印本 |
| SkillSmith：协同演化技能与工具，构建自我改进的智能体系统 | arXiv 预印本 |
| PANDO: Efficient Multimodal AI Agents via Online Skill Distillation | arXiv preprint |
| DemoEvolve: Overcoming Sparse Feedback in Agentic Harness Evolution with Demonstrations | arXiv preprint |
| MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems | arXiv preprint |
| Continual Harness: Online Adaptation for Self-Improving Foundation Agents | arXiv preprint |
| LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling | arXiv preprint |
| Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses | arXiv preprint |
| Escher-Loop: Mutual Evolution by Closed-Loop Self-Referential Optimization | arXiv preprint |
| Mem²Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation | ACL 2026 |
| Meta-Harness: End-to-End Optimization of Model Harnesses | arXiv preprint |
| Hyperagents | arXiv preprint |
| Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing | arXiv preprint |
| Meta Context Engineering via Agentic Skill Evolution | ICML 2026 |
| Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine | ICLR 2026 Oral |
| Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents | ICLR 2026 |
| Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement | ACL 2025 |

上下文

提示词或工作上下文会更新。在连续运行中，观察、反馈和经验可以不断累积或重新组织，用来指导后续动作或任务。

| 论文或项目 | 发表 |
| HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs | arXiv preprint |
| Harness²: Recursive Agent Harnessing for an Open World | arXiv preprint |
| EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution | arXiv preprint |
| EvoBelief: Falsifiable World Models for Self-Improving LLM Agents | arXiv preprint |
| RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments | arXiv preprint |
| Learning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents | arXiv preprint |
| EviSkill: Grounding Skill Evolution in Replayable Evidence | arXiv preprint |
| MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution | arXiv preprint |
| VERSE: Verified Self-Evolving Optimizer for Agent Harnesses | arXiv preprint |
| Learning from Research: Toward Lifelong Agent Harness Evolution | arXiv preprint |
| NavHarness: Towards Lifelong Embodied Navigation | arXiv preprint |
| MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens | NeurIPS 2026 |
| SkillEvoReg: Regularizing Agent Skill Evolution Against Overfitting | arXiv preprint |
| SkillRefine: Cross-Source Skill Induction and Execution Validation for LLM Agents in Refinery Planning Software | arXiv preprint |
| A Wrong Turn Does Not Ruin the Journey: Deviation-Guided Skill Self-Evolution for LLM Agents | arXiv preprint |
| HarnessPAI: An Evolving Harness for Physical AI | arXiv preprint |
| SEEK: Skill-Routed Evaluation with Evolvable Knowledge for Industrial Search | arXiv preprint |
| Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents | arXiv preprint |
| Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity | arXiv preprint |
| Recursive self-improvement of AI research agents | arXiv preprint |
| Harness-Zero: Harness Distillation via Agent-as-Harness | arXiv preprint |
| RRSI: Regularized Recursive Self-Improvement of Agent Harnesses | arXiv preprint |
| TTSE: A Two-Track Online Self-Evolution Framework | arXiv preprint |
| Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design | arXiv preprint |
| Self Improvement via Fast Tree-search | arXiv preprint |
| SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback | arXiv preprint |
| SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness | arXiv preprint |
| EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents | arXiv preprint |
| Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents | arXiv preprint |
| AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery | arXiv preprint |
| AnchorGUI: Asymmetric Memory for Dual-Scale Learning in GUI Navigation | ECCV 2026 |
| EvoOntology: A Self-Evolving Ontology Layer for Data Agents | arXiv preprint |
| ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement | arXiv preprint |
| RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments | arXiv preprint |
| Dream-RSI: Recursive Self-Improvement through Evolving Worlds | Preprint |
| COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization | arXiv preprint |
| Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents | arXiv preprint |
| Code-to-Harness: Distilling Black-Box Optimizers from Self-Play | arXiv preprint |
| Experience Funnel: A State-Policy Alternating Loop for Self-Evolving Agents | arXiv preprint |
| Procedural Graphs: Self-Evolving Execution Structures for LLM Agents | arXiv preprint |
| SkillAdam: Stable and Efficient Skill Evolution for Agents | arXiv preprint |
| SimSkill: A Self-Evolving LLM Agent for Skill and Knowledge Accumulation in Traffic Simulation | arXiv preprint |
| CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning | arXiv preprint |
| Reef: Continual Learning Infrastructure for Self-Improving Agents | Project · v0.0.2 |
| SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams | arXiv preprint |
| Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement | arXiv preprint |
| HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution | arXiv preprint |
| PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents | arXiv preprint |
| WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution | arXiv preprint |
| Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses | arXiv preprint |
| MediSkill-Evo：面向循证临床交互的过程约束自我进化 | arXiv preprint |
| Prime Agent：可自我改进的 RLM Harness | arXiv preprint |
| Proteus：面向 AI 智能体的 Harness 无关自我进化框架 | Project · v0.3.0 |
| TRACE：面向一致、限额感知 LLM 智能体的自我进化技能库 | arXiv preprint |
| ForeDreamer：面向未来事件预测的自我进化双智能体记忆架构 | Findings of EMNLP 2026 |
| EnvHarness：唤醒静态世界，服务智能体学习 | arXiv preprint |
| HyperSkill：基于超图结构技能记忆的自我进化 LLM 智能体 | arXiv preprint |
| Evo-Harness：面向自我进化智能体的上下文到 Harness 技能编译 | arXiv preprint |
| DIVE：通过多样性驱动技能进化，释放冻结语言模型的自我改进能力 | arXiv preprint |
| Mendel Gödel Machine：通过比较进化实现递归自我改进的编码智能体 | arXiv preprint |
| SkillHEX：通过假设驱动的自主探索与利用改进智能体技能 | arXiv preprint |
| EvolveNet：面向智能体自我改进的协作式 Harness 进化 | arXiv preprint |
| DarwinX：通过自然选择进化智能体 Harness | arXiv preprint |
| 以知识为中心的自我改进 | arXiv preprint |
| 递归式 Harness 自我改进 | arXiv preprint |
| HarnessBank：面向智能体 Harness 自我进化的语义基因库搜索与门控验证 | arXiv preprint |
| SAGEAgent：面向多模态生存预测中成本感知模态获取的自我进化智能体 | arXiv preprint |
| 红皇后 Gödel Machine：协同进化智能体及其评估器 | arXiv preprint |
| 当规则学会学习：面向法律案例检索的自我进化智能体 | ACL 2026 |
| HarnessX：可组合、自适应、可进化的智能体 Harness 铸造厂 | arXiv preprint |
| Self-Harness：自我改进的 Harness | arXiv preprint |
| 从失败轨迹到可靠 LLM 智能体：诊断并修复 Harness 缺陷 | arXiv preprint |
| OpenSkill：面向 LLM 智能体的开放世界自我进化 | arXiv preprint |
| SePO: Self-Evolving Prompt Agent for System Prompt Optimization | arXiv preprint |
| SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision | Findings of EMNLP 2026 |
| ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents | arXiv preprint |
| PANDO: Efficient Multimodal AI Agents via Online Skill Distillation | arXiv preprint |
| DemoEvolve: Overcoming Sparse Feedback in Agentic Harness Evolution with Demonstrations | arXiv preprint |
| Continual Harness: Online Adaptation for Self-Improving Foundation Agents | arXiv preprint |
| LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling | arXiv preprint |
| Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses | arXiv preprint |
| Mem²Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation | ACL 2026 |
| Meta-Harness: End-to-End Optimization of Model Harnesses | arXiv preprint |
| P²O: Joint Policy and Prompt Optimization | arXiv preprint |
| Hyperagents | arXiv preprint |
| Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing | arXiv preprint |
| Meta Context Engineering via Agentic Skill Evolution | ICML 2026 |
| Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine | ICLR 2026 Oral |
| Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models | ICLR 2026 |
| ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory | ICLR 2026 |
| Memp: Exploring Agent Procedural Memory | Findings of ACL 2026 |
| GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning | ICLR 2026 Oral |
| Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents | ICLR 2026 |
| Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory | EACL 2026 |
| Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement | ACL 2025 |
| Cradle: Empowering Foundation Agents Towards General Computer Control | ICML 2025 |

记忆

信息或经验会跨步骤、轨迹或任务存储、更新和检索。

| 论文或项目 | 发表 |
| Harness²: Recursive Agent Harnessing for an Open World | arXiv preprint |
| EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution | arXiv preprint |
| EvoBelief: Falsifiable World Models for Self-Improving LLM Agents | arXiv preprint |
| RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments | arXiv preprint |
| Learning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents | arXiv preprint |
| EviSkill: Grounding Skill Evolution in Replayable Evidence | arXiv preprint |
| Learning from Research: Toward Lifelong Agent Harness Evolution | arXiv preprint |
| NavHarness: Towards Lifelong Embodied Navigation | arXiv preprint |
| SkillRefine: Cross-Source Skill Induction and Execution Validation for LLM Agents in Refinery Planning Software | arXiv preprint |
| HarnessPAI: An Evolving Harness for Physical AI | arXiv preprint |
| Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity | arXiv preprint |
| Recursive self-improvement of AI research agents | arXiv preprint |
| Harness-Zero: Harness Distillation via Agent-as-Harness | arXiv preprint |
| RRSI: Regularized Recursive Self-Improvement of Agent Harnesses | arXiv preprint |
| TTSE: A Two-Track Online Self-Evolution Framework | arXiv preprint |
| Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design | arXiv preprint |
| SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback | arXiv preprint |
| Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents | arXiv preprint |
| AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery | arXiv preprint |
| AnchorGUI: Asymmetric Memory for Dual-Scale Learning in GUI Navigation | ECCV 2026 |
| EvoOntology: A Self-Evolving Ontology Layer for Data Agents | arXiv preprint |
| ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement | arXiv preprint |
| ReMe：面向 AI 智能体的本地优先、自进化个人知识库 | 项目 · v0.4.1.12 |
| RSIAgent：在新环境中自主探索递归式自我改进 | arXiv 预印本 |
| Dream-RSI：通过不断演化的世界实现递归式自我改进 | 预印本 |

| Experience Funnel: A State-Policy Alternating Loop for Self-Evolving Agents | arXiv preprint |
| Procedural Graphs: Self-Evolving Execution Structures for LLM Agents | arXiv preprint |
| SkillAdam: Stable and Efficient Skill Evolution for Agents | arXiv preprint |
| SimSkill: A Self-Evolving LLM Agent for Skill and Knowledge Accumulation in Traffic Simulation | arXiv preprint |
| CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning | arXiv preprint |
| SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams | arXiv preprint |
| Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement | arXiv preprint |
| PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents | arXiv preprint |
| WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution | arXiv preprint |
| Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses | arXiv preprint |
| MediSkill-Evo: Process-Constrained Self-Evolution for Evidence-Grounded Clinical Interaction | arXiv preprint |
| Prime Agent: A Self-Improving RLM Harness | arXiv preprint |
| Proteus: A Harness-Agnostic Self-Evolution Framework for AI Agents | Project · v0.3.0 |
| ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction | Findings of EMNLP 2026 |
| EnvHarness: Awakening Static Worlds for Agent Learning | arXiv preprint |
| HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory | arXiv preprint |
| Knowledge-Centric Self-Improvement | arXiv preprint |
| SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction | arXiv preprint |
| ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning | ICML 2026 AI for Math Workshop |
| HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry | arXiv preprint |
| Self-Harness: Harnesses That Improve Themselves | arXiv preprint |
| From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws | arXiv preprint |
| SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems | arXiv preprint |
| ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents | arXiv preprint |
| PANDO: Efficient Multimodal AI Agents via Online Skill Distillation | arXiv preprint |
| Continual Harness: Online Adaptation for Self-Improving Foundation Agents | arXiv preprint |
| LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling | arXiv preprint |
| Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses | arXiv preprint |
| HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents | ACL 2026 |
| Mem²Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation | ACL 2026 |
| Meta-Harness: End-to-End Optimization of Model Harnesses | arXiv preprint |
| Hyperagents | arXiv preprint |
| Meta Context Engineering via Agentic Skill Evolution | ICML 2026 |
| Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models | ICLR 2026 |
| ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory | ICLR 2026 |
| Memp: Exploring Agent Procedural Memory | Findings of ACL 2026 |
| Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory | EACL 2026 |
| A-MEM: Agentic Memory for LLM Agents | NeurIPS 2025 |
| Agent Workflow Memory | ICML 2025 |
| Cradle: Empowering Foundation Agents Towards General Computer Control | ICML 2025 |

Skill

Reusable strategies, procedures, workflows, or skill resources are created and revised.

| Paper or project | Publication |
| Skill∞: Recursive Self-Improvement for Agent Skill Optimization | arXiv preprint |
| HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs | arXiv preprint |
| Harness²: Recursive Agent Harnessing for an Open World | arXiv preprint |
| EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution | arXiv preprint |
| RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments | arXiv preprint |
| Learning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents | arXiv preprint |
| EviSkill: Grounding Skill Evolution in Replayable Evidence | arXiv preprint |
| Learning from Research: Toward Lifelong Agent Harness Evolution | arXiv preprint |
| NavHarness: Towards Lifelong Embodied Navigation | arXiv preprint |
| SkillEvoReg: Regularizing Agent Skill Evolution Against Overfitting | arXiv preprint |
| SkillRefine: Cross-Source Skill Induction and Execution Validation for LLM Agents in Refinery Planning Software | arXiv preprint |
| A Wrong Turn Does Not Ruin the Journey: Deviation-Guided Skill Self-Evolution for LLM Agents | arXiv preprint |
| HarnessPAI: An Evolving Harness for Physical AI | arXiv preprint |
| SEEK: Skill-Routed Evaluation with Evolvable Knowledge for Industrial Search | arXiv preprint |
| Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity | arXiv preprint |
| Harness-Zero: Harness Distillation via Agent-as-Harness | arXiv preprint |
| RRSI: Regularized Recursive Self-Improvement of Agent Harnesses | arXiv preprint |
| TTSE: A Two-Track Online Self-Evolution Framework | arXiv preprint |
| Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design | arXiv preprint |
| SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback | arXiv preprint |
| EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents | arXiv preprint |
| Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents | arXiv preprint |
| AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery | arXiv preprint |
| AnchorGUI: Asymmetric Memory for Dual-Scale Learning in GUI Navigation | ECCV 2026 |
| ReMe: A Local-First, Self-Evolving Personal Knowledge Base for AI Agents | Project · v0.4.1.12 |
| RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments | arXiv preprint |
| COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization | arXiv preprint |
| Code-to-Harness: Distilling Black-Box Optimizers from Self-Play | arXiv preprint |
| Experience Funnel: A State-Policy Alternating Loop for Self-Evolving Agents | arXiv preprint |
| Procedural Graphs: Self-Evolving Execution Structures for LLM Agents | arXiv preprint |
| SkillAdam: Stable and Efficient Skill Evolution for Agents | arXiv preprint |
| SimSkill: A Self-Evolving LLM Agent for Skill and Knowledge Accumulation in Traffic Simulation | arXiv preprint |
| Reef: Continual Learning Infrastructure for Self-Improving Agents | Project · v0.0.2 |
| SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams | arXiv preprint |
| HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution | arXiv preprint |
| PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents | arXiv preprint |
| WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution | arXiv preprint |
| Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses | arXiv preprint |
| MediSkill-Evo: Process-Constrained Self-Evolution for Evidence-Grounded Clinical Interaction | arXiv preprint |
| Prime Agent: A Self-Improving RLM Harness | arXiv preprint |
| Proteus: A Harness-Agnostic Self-Evolution Framework for AI Agents | Project · v0.3.0 |
| TRACE: A Self-Evolving Skill Bank for Consistent, Limit-Aware LLM Agents | arXiv preprint |
| ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction | Findings of EMNLP 2026 |
| EnvHarness: Awakening Static Worlds for Agent Learning | arXiv preprint |
| HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory | arXiv preprint |
| Evo-Harness: Context-to-Harness Skill Compilation for Self-Evolving Agents | arXiv preprint |
| DIVE: Unlocking Self-Improvement in Frozen Language Models Through Diversity-Driven Skill Evolution | arXiv preprint |
| SkillHEX: Improving Agent Skills via Hypothesis-Driven Autonomous Exploration and Exploitation | arXiv preprint |
| DarwinX: Evolving Agent Harnesses Through Natural Selection | arXiv preprint |
| SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction | arXiv preprint |
| ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning | ICML 2026 AI for Math Workshop |
| When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval | ACL 2026 |
| OpenSkill: Open-World Self-Evolution for LLM Agents | arXiv preprint |
| SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision | Findings of EMNLP 2026 |
| SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems | arXiv preprint |
| ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents | arXiv preprint |
| PANDO: Efficient Multimodal AI Agents via Online Skill Distillation | arXiv preprint |
| Continual Harness: Online Adaptation for Self-Improving Foundation Agents | arXiv preprint |
| Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses | arXiv preprint |
| Mem²Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation | ACL 2026 |
| Meta Context Engineering via Agentic Skill Evolution | ICML 2026 |
| Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models | ICLR 2026 |
| Memp: Exploring Agent Procedural Memory | Findings of ACL 2026 |
| Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory | EACL 2026 |
| Agent Workflow Memory | ICML 2025 |
| Cradle: Empowering Foundation Agents Towards General Computer Control | ICML 2025 |

Other artifacts

Other artifacts evolve, such as data strategies, experiment configurations, or task solutions.

| Paper or project | Publication |
| Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement | arXiv preprint |
| EnvHarness: Awakening Static Worlds for Agent Learning | arXiv preprint |

Benchmarks

Compare benchmark dimensions on the website →

Grouped by RSI mode, newest first. Benchmarks that support multiple protocols appear in each relevant group. Artifact labels describe what evolves in the evaluated workflow.

Online · Offline · Offline → Online

Online

Experience accumulates during the task stream or repeated interaction, and later work can use it.

| Benchmark paper | Year | RSI artifact |
| RSIGym: A Flexible Environment for Recursive Self-Improvement | 2026 | Model parameters · Harness code · Context · Other artifacts |
| Agent Plasticity: Measuring Self-Improvement Through Experience | 2026 | Harness code · Context · Memory · Skill |
| Aspire: Can Models Self-Evolve from Vague Goals? | 2026 | Model parameters · Harness code · Context · Memory |
| FinEvo-Bench: A Longitudinal Benchmark for Self-Evolving Agents in Professional Financial Workflows | 2026 | Memory · Skill |
| ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities? | 2026 | Context · Skill |
| PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents | 2026 | Memory · Skill |
| PATH-Bench: Path-Dependent Evaluation of Lifelong Agents | 2026 | Context · Memory · Skill |
| AgentStream: How Well Do Self-Evolving LLM Agents Perform Under Streaming Tasks? | 2026 | Context · Memory · Skill |
| RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement | 2026 | Context · Other artifacts |
| EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments | 2026 | Context · Other artifacts |
| Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments | 2026 | Context · Memory · Other artifacts |
| AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks? | 2026 | Context · Other artifacts |
| Can Generalist Agents Automate Data Curation? | 2026 | Context · Other artifacts |
| EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective | 2026 | Context · Memory |
| MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI | 2026 | Context · Other artifacts |
| SkillLearnBench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real-World Tasks | 2026 | Skill |
| SkillFlow: Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents | 2026 | Context · Skill |
| Agent² RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training? | 2026 | Context · Other artifacts |
| PostTrainBench: Can LLM Agents Automate LLM Post-Training? | 2026 | Context · Other artifacts |
| Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory | 2025 | Context · Memory |
| Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark | 2025 | Context · Memory · Skill |
| LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners | 2025 | Context · Memory |

Offline

Evolution precedes a separate held-out evaluation.

| Benchmark paper | Year | RSI artifact |
| Agent Plasticity: Measuring Self-Improvement Through Experience | 2026 | Harness code · Context · Memory · Skill |
| HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? | 2026 | Harness code · Context |
| Aspire: Can Models Self-Evolve from Vague Goals? | 2026 | Model parameters · Harness code · Context · Memory |
| S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement? | 2026 | Model parameters · Context · Memory |
| Evo-Bench: Can Language Models Improve Agent Harness? | 2026 | Harness code · Context |
| HarnessOpt-Bench: Evaluating LLMs at Harness Optimization | 2026 | Harness code · Context |
| GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks | 2026 | Skill |
| EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer | 2026 | Skill |
| SEAGym: An Evaluation Environment for Self-Evolving LLM Agents | 2026 | Harness code · Context · Memory · Skill |
| The Meta-Agent Challenge: Are Current Agents Capable of Autonomous Agent Development? | 2026 | Harness code · Context |
| SkillFlow: Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents | 2026 | Context · Skill |
| VeRO: A Harness for Agents to Optimize Agents | 2026 | Harness code · Context |
| MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems | 2025 | Memory |
| Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions | 2025 | Context · Memory |

Offline to online

An artifact is built offline and continues to evolve during online use.

| Benchmark paper | Year | RSI artifact |
| SkillFlow: Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents | 2026 | Context · Skill |

Use the explorer

1. Choose a collection: Benchmark papers or Methods & Systems.
2. Narrow the question: select multiple tags within a row, then combine dimensions across rows. For example, look for methods that update Skill, run Online, and use Environment feedback.
3. Read and compare: search titles, sort by recency or citations, and expand abstracts and taxonomy profiles.
4. Follow the connections: open Citation graph, choose methods or benchmarks, and search for a paper. Drag and zoom to explore its neighborhood.

The site supports English and Chinese, with the language switch in the upper-right corner.

Citation

If you find this collection useful, please cite our survey:

@misc{zhou2026existingmethodsnewresearch,
title = {From Existing Methods to New Research Ideas: Two Perspectives on Recursive Self-Improvement},
author = {Leijun Zhou and Yanke Yu and Zhihao Liu and Qu Xiang and Tiancheng Hu and Yiru Dong and Yihan Wu and Yangxi Xiang and Xuejun Wu and Xi Chen and Buyue Qian and Yaowei Zheng and Junhao Hu},
year = {2026},
eprint = {2610.14625},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.14625},
}

WeChat Group

Scan the QR code to join the Awesome RSI discussion group.

Back to top ↑