待翻译：The Path to Recursive Self-Improving Agents

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

Uh oh! There was an error while loading. Please reload this page. Notifications You must be signed in to change notification settings Fork 1 Star 13 BranchesTags Open more actions menu Folders and files NameName Last commit message Last commit date Latest commit History 18 Commits 18 Commits assets assets README.md README.md README.zh-CN.md README.zh-CN.md google2215cfd46f5372a6.html google2215cfd46f5372a6.html index.html index.html index.zh-CN.html index.zh-CN.html papers.yaml papers.yaml Repository files navigation Shuaiqi Liu*, Zhengkai Lin*, Yuxiang Zhang*, Yuanyi Ren*, Yue Wu, Yongbin Li, Zheng Wang, Zhihang Fu*, Jieping Ye Alibaba Group English | 中文版 Full Text A living survey and structured literature map for self-improving agent systems: systems that transform experience and evaluation feedback into persistent updates to their own components, moving toward recursive self-improving agents. Overview This repository accompanies the survey “The Path to Recursive Self-Improving Agents: Foundation, Framework, and Future Directions.” The survey studies agent systems that can autonomously convert execution traces, training signals, validation results, and other feedback into durable changes to their own components. It formalizes the agent system as: $$ x_t = (\mathcal{M}_t, \mathcal{H}_t, \mathcal{D}_t, \mathcal{T}_t, \mathit{Imp}_t) $$ where (\mathcal{M}) is the foundation model, (\mathcal{H}) is the agent harness, (\mathcal{D}) is the agent data system, (\mathcal{T}) is the agent trainer, and (\mathit{Imp}) is the improvement mechanism. A self-improving system follows the update rule: $$ x_{t+1} = \mathit{Imp}_t(x_t) $$ Recursive self-improvement (RSI) appears when (\mathit{Imp}t) is itself part of the modifiable system state and can become (\mathit{Imp}{t+1}), allowing the system to improve its own future improvement process. Why This Survey Matters The paper makes four main contributions: Formal foundation: It defines agent-system self-improvement and recursive self-improvement, then introduces a five-level capability grading standard from manual improvement to general recursive self-improvement. Unified framework: It models the foundation model, agent harness, agent data system, agent trainer, and improvement mechanism as a coupled evolving system. Taxonomy of existing work: It organizes research into agent harness self-improvement, agent data system self-improvement, agent trainer self-improvement, and cross-component co-improvement. Roadmap for future research: It identifies open problems around long-horizon evaluation, modifiable infrastructure, bounded-to-general RSI, safety, and human-agent co-improvement. Core Framework The unified framework emphasizes two questions: What can be improved? The autonomous improvement scope may include the model, harness, data system, trainer, and eventually the improvement mechanism itself. How does improvement proceed? The improvement mechanism process includes diagnosing bottlenecks, proposing candidate modifications, evaluating and selecting candidates, and integrating accepted changes into persistent system state. L1–L5 Capability Grading The survey distinguishes self-improving systems by autonomy in the improvement loop, whether the improvement mechanism is mutable, and whether the capability generalizes across domains. Level Category Name Autonomous Proposal Autonomous Implementation Autonomous Validation Improve the Improvement Mechanism Domain Generality L1 Manual Improvement L2 Assisted Improvement ✓ L3 Programmatic Self-Improvement ✓ ✓ ✓ L4 Bounded Recursive Self-Improvement ✓ ✓ ✓ ✓ L5 General Recursive Self-Improvement ✓ ✓ ✓ ✓ ✓ L1 — Manual Improvement: This type of agent system has no autonomous improvement capability. It is deployed and executed in a fixed form, and every change requires a manual development and deployment process. L2 — Assisted Improvement: This type of system can propose candidate modifications or provide diagnostic evidence, but humans remain responsible for validating and applying substantive changes. The improvement mechanism Imp is human-maintained, so L2 is a precursor to self-improvement rather than a full instance of it. L3 — Programmatic Self-Improvement: This type of system can autonomously propose, apply, and validate modifications to operational components such as the foundation model, agent harness, data system, or trainer. However, Imp remains fixed or externally maintained. L4 — Bounded Recursive Self-Improvement: This type of system can not only propose, validate, and apply candidate modifications, but also rewrite its own improvement mechanism, making the improvement process self-referential. The process supports sustained long-term progress within a bounded domain. L5 — General Recursive Self-Improvement: L5 retains the autonomy, self-reference, and long-term progress of L4, while transferring improvement capability across broad and evolving task domains rather than remaining limited to a fixed benchmark or narrow operational setting. Living Survey Index Taxonomy Agent Harness Self-Improvement The harness is the execution layer that determines what the model observes, what actions it can invoke, and how observations and actions are organized into task workflows. Harness self-improvement means converting traces, evaluation feedback, and failures into persistent changes to memory, skills, prompts, workflows, tools, or executable scaffolds. Module-Level Self-Improvement: Memory This category of work primarily modifies experience memory, retrieval rules, and memory architecture, turning execution traces into reusable, persistent state that conditions later behavior. Paper Link Year Level Reflexion link 2023 L3 SWE-Exp link 2025 L3 ReasoningBank link 2025 L3 FLEX link 2025 L3 Decocted Experience link 2026 L3 Live-Evo link 2026 L3 Memory Transfer Learning link 2026 L3 Module-Level Self-Improvement: Skill & Tool This category of work primarily modifies skill libraries, tool libraries, or the shared skill-tool ecosystem that agents retrieve, compose, and reuse. Paper Link Year Level Tool Makers link 2023 L3 ToolCoder link 2025 L3 SkillWeaver link 2025 L3 SkillX link 2026 L3 Workflow-to-Skill link 2026 L3 SkillFoundry link 2026 L3 OpenSkill link 2026 L3 SkillOS link 2026 L3 SkillComposer link 2026 L3 Programmatic Skill Networks link 2026 L3 SkillAxe link 2026 L3 SkillAudit link 2026 L3 SkillGen link 2026 L3 CoEvoSkills link 2026 L3 SkillSmith link 2026 L3 Confucius Code Agent link 2025 L3 Module-Level Self-Improvement: Prompt & Context This category of work primarily modifies prompt programs, context playbooks, guideline documents, or decision-rule libraries; a small subset also makes the prompt optimizer itself mutable. Paper Link Year Level Dynamic Cheatsheet link 2025 L3 ACE (Context Engineering) link 2025 L3 SCOPE link 2025 L3 Reflective Context Learning link 2026 L3 Unified Context Evolution link 2026 L3 KACE link 2026 L3 MEMO link 2026 L3 SEEK-SQL link 2026 L3 GEPA link 2025 L3 Trace2Policy link 2026 L3 Learning to Self-Evolve link 2026 L3 SePO link 2026 L4 Orchestration & Architecture Search This category of work primarily modifies trajectory orchestration, agent composition, or workflow/communication topology, so the harness spends compute where uncertainty is high and reuses useful intermediate results. Paper Link Year Level AgentGA link 2026 L3 SWE-Replay link 2026 L3 Log-Augmented Generation link 2025 L3 FailureMem link 2026 L3 EvoRepair link 2026 L3 EvoAgent link 2024 L3 ADAS link 2024 L3 EvoMAS link 2026 L3 EVOCHAMBER link 2026 L3 MermaidFlow link 2025 L3 EvoAgentX link 2025 L3 EvoFlow link 2025 L3 SEW link 2025 L3 HyEvo link 2026 L3 AdaptFlow link 2025 L3 JudgeFlow link 2026 L3 Lean4Agent / LeanEvolve link 2026 L3 EvoFSM link 2026 L3 ScoreFlow link 2025 L3 Learning to Compose link 2026 L3 Workflow-R1 link 2026 L3 Learning to Hand Off link 2026 L3 AutoTTS link 2026 L3 Self-Referential Code Modification This category of work primarily modifies harness policy, runtime scaffold, agent source code, or the improvement mechanism itself, making the modified agent part of the mechanism that produces further changes. Paper Link Year Level Life-Harness link 2026 L3 HarnessFix link 2026 L3 Milkyway link 2026 L3 Self-Harness link 2026 L3 POLARIS link 2026 L3 DemoEvolve link 2026 L3 SIGA link 2026 L3 HarnessForge link 2026 L3 Agentic Harness Engineering link 2026 L3 Meta-Harness link 2026 L3 AgentFlow link 2026 L3 Adaptive Auto-Harness link 2026 L3 Group-Evolving Agents link 2026 L3 Gödel Agent link 2025 L4 SICA link 2025 L4 Darwin Gödel Machine link 2026 L4 Huxley-Gödel Machine link 2025 L4 Red Queen Gödel Machine link 2026 L4 HyperAgents link 2026 L4 Harnessing Agentic Evolution link 2026 L4 EvoX link 2026 L4 ANCHOR link 2026 L3 Statistical Gödel Machine link 2025 L3 ANNEAL link 2026 L3 Agent Data System Self-Improvement The agent data system manages the lifecycle of training and evaluation data. It includes data production and data utilization, connecting inference-time interaction records with persistent learning in the trainer. Environment Generation & Simulation This category of work primarily modifies environment configurations, simulators, or the tool/skill libraries used to construct environments, keeping training environments aligned with the agent's evolving capability boundary. Paper Link Year Level EnvGen link 2024 L3 Adaptive Environment Generation link 2026 L3 ACCEL link 2022 L3 DRED link 2024 L3 SimWorld Studio link 2026 L3 DreamGym link 2025 L3 Task Synthesis This category of work primarily modifies task instructions, constraints, or evaluation criteria, using diagnostic feedback, fixed quality-control signals, or competitive feedback to keep tasks relevant and appropriately difficult. Paper Link Year Level WebRL link 2024 L3 CoEvolve link 2026 L3 SeRL link 2026 L3 Self-CriTeach link 2026 L3 SAGE link 2026 L3 Trajectory Synthesis This category of work primarily modifies solution trajectories, self-play data, or preference data that record agent experience for a given task. Paper Link Year Level SPIN link 2024 L3 Arena Learning link 2024 L3 EVOLVE link 2025 L3 DNPO link 2026 L3 PLD link 2025 L3 Verification & Quality Assurance This category of work primarily modifies judge models, validators, or quality filters that determine which synthesized data is retained. Paper Link Year Level Self-Improving VLM Judges link 2025 L3 Curriculum Adaptation This category of work primarily modifies the curriculum scheduler, task allocation, or difficulty adaptation that determines how produced data are selected and ordered for training. Paper Link Year Level AMC-TSI link 2026 L3 EvoCurr link 2025 L3 TRUSTEE link 2026 L3 Actor-Curator link 2026 L3 Multi-Module Co-Improvement (within the Data System) This category of work primarily couples two or more data-production or data-utilization modules through feedback loops, so multiple modules improve collaboratively rather than in isolation. Paper Link Year Level Agent0-VL link 2025 L3 ACE (Adversarial Code Evolution) link 2026 L3 Agent-World link 2026 L3 Agent0 link 2025 L3 R-Zero link 2025 L3 Agent Trainer Self-Improvement The trainer converts agent experience into persistent model updates. Trainer self-improvement revises the persistent state that determines supervision, optimization, infrastructure, or the mechanism that improves the trainer itself. Inner-Loop Trainer Adaptation This loop modifies persistent trainer state, such as a reward model, verifier, rubric memory, process reward model, batch-size controller, or infrastructure configuration, within an active training lineage while the improvement mechanism remains fixed. Paper Link Year Level ACE [truncated for AI cost control]