Self-Improving Agents

Experience-Handoff Reinforcement Learning for Recursive Self-Improvement

MULTILOOP: Evolving Multi-Agent Harnesses for Complex Information-Seeking Tasks

Verify by Doing: Co-Evolving Agent Execution and Evaluation without External Feedback

When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents

Program-Aided Patchable Skill Evolution

SkillDuet: Co-Evolving Models and Skills through Response Feedback

InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure

JoMoHa: Joint Model--Harness Optimization with Reinforcement Learning

Agent Self-Improvement with Frozen Weights: The Added Value and Limits of Failure Traces

Trustworthy Recursive Self-Improvement with Statistical Guarantees

No Second Run: Does RSI Stop at Frontier Pre-training?

LongCapHarness: Capability-Guided Auto-Harnessing for Long-Context Reasoning

AutoRefine: Auto-Calibrating Code Judges from Workflow Outcomes

BOXBOX: Recursive Harness Self-Improvement for Long-Running Performance Optimization

A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics

Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents

Scientist World: Grow a Research World for AI-Scientists

Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement

Beneath the Surface: What We Overlook When We Talk About Agent Evolution

Evolver: Turning Agent Exploration into Evolving Capabilities

The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks

EnvOpt: Automated Optimization for Agent Environments

HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents

Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Non-Verifiable Domains

From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?

Recursa: Toward Recursively Self-Improving Foundation Model Systems

WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis

Strategist: Evolving Reasoning with Agents

ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost

SPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales

MAPLE: A Multimodal Mobile Agent with Playbook-Learned Experience

Skill++: Towards Lifelong Agents via Routing-Aware Skill Refinement

Exo: An Architecture for Self-Improving Systems at Runtime

RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory

When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications

KARMA: Training-Free Knowledge-Atomic Harness Evolution for Industrial Advertising Agents

Training LLM Agents to Automate LLM Post-Training with RSI-Dojo

Compile Away Uncertainty: Self-Evolving Agent Skills via Programmatic Compilation

MetaVul: Agentic Harness Discovery for Vulnerability Detection

Where to Learn and How Far to Roll Out: Efficient Adaptive Training for GUI Agents

Appreciation Teaches Unauthorized Execution: Self-Evolution Degrades the Robustness of Personal Agents to Indirect Prompt Injection

Aligning Autoresearch: Co-Improvement between Agents and Humans

Routed Continual Learning

Masked Erosion: How Acceptance Gates Lose Redundant Safeguards in Self-Improving Agents

GEPA-Curator: Curated Learning for Self Improving Agents

Does Your Generalist Skill Optimizer Really Incentivize Agentic Tasks? Rethinking Skill Optimization via Bipartite Attribution

MLE-World: Predicting Machine Learning Experiment Outcomes with a Learned World Model

Evolving Agent Teams

When the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-Improvement

The View from Nowhere: Population-Centered Agent Evolution

How Do Coding-Agent Harnesses Self-Evolve? An Empirical Study of Retrospective Adaptation

EMMA: Structured Episodic Memory for Multi-Agent Workflow Synthesis

Holding a Steady Course: Knowledge Evolution from Expertise and Lessons for Continual Agent Learning

FlashScript: Compiling Recurring Agent Workflows into Reusable Neuro-Symbolic Programs

MTA-Zero: Self-Improving Multimodal Deep Search via Online QA-Solver Co-training

RSI Arena: Towards Recursive Self-Improvement with Human Feedback

HarnessMM: Recursive Self-Improvement for Harness Token Efficiency

On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models

RSIHub: A Composable Framework for Agent-Evolution Methods

Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons

RIPPLE: Recovery Internalization through Process-Level Policy Learning and Effect-Aware Credit for Long-Horizon Agentic Coding

ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents

Agentic Knowledge-Model Co-Evolution via Consistency Regularization

Scaling Laws for Generative Reward Models

Mining the Design Space of Multi-Agent Workflows

Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

SplatLab: Autonomous Scientific Discovery for 3DGS Research

RubricCraft: Where Solutions Diverge, Criteria Emerge

Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable

Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis

Forging More Tasks into Better Harnesses: Agentic Harness Evolution via Successive Scale Racing

Evolving Agent Interfaces

Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation

Object-Oriented Harness: Evolving Agent Harnesses through Object-Oriented Design

Self-Evolving Static Analysis via LLM Pseudo-Labeling

RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process?

Self-Evolving Agents through Privileged Retrospective Distillation

More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents

Evolving Self-tuning Benchmarks for Long-Horizon Tool Use

Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

Are Self-Evolving Agents Continual Learners?

Beyond the Teacher: Understanding and Preserving Distillation Overshoot in Agentic On-Policy Distillation

Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research

DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

Language Models Learn to Improve Language Models

Self-Evolving Agents Can Repair Failures but Struggle to Generalize from Them

Self-Evolving Coding Agents via Harness-Policy Optimization

Can LLMs predict the outcomes of machine learning experiments?

MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

Learning Self-Orchestration in Decentralized Agent Swarms

Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation

Harness-Delta Attribution: Dissecting Gains From Harness Evolution

Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine

Data World: Scaling Data Environments from Scratch

The Self-Improvement Problem

Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use

OPRD: On-Policy Reminder Distillation for Long-Horizon Coding Agents

MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling

Learning from Interaction Experience via Localized Context Distillation

MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning

Athanor: An Agent-Native System for Autonomous Post-Training Research

Wrong Votes, Right Gradients: Why Majority-Vote Rewards Track Verifiable RL

Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents

Scaling Discovery Requires Forgetting

The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents

DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck

Learning to GUIDE the Group: Self-Learning LLM Facilitation Across Multi-Party Tasks

Heterogeneous Agentic Coevolution

Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses

Self-Evolving Visual Agent via Margin Reward

Harness^2: Recursive Agent Harnessing for an Open World

AutoEnvScaling: Automating the Data Flywheel with Terminal Agents

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems

Proactive Agentic Orchestration

Rollout-Time Intervention: Unlocking Agentic Reinforcement Learning Beyond Policy Limits

Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution

EvoBelief: Falsifiable World Models for Self-Improving LLM Agents

Recurrent Harness for Agentic Continual Learning

Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents

Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning

Recursive Self-Improvement through Multi-Agent Self-Supervision

EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution

Agentic-TTT: Training test-time policy for test-time training

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks

Harness Evolution Hits a Ceiling: When Weight Training Should Begin

Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents

Environmental Feedback Modeling Matters: Rethinking Feedback Treatment in Agentic Hindsight Self-Distillation

SynCo: Data Synthesis Co-Training for Self-Evolving LLMs via Multi-Agent Reinforcement Learning

DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration

Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context Reinforcement Learning

CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution

Training Advisors for LLM Agents from Task Outcomes

SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles

MIMESIS: Learning User Simulators as Training Environments for Interactive Agents

The Winner's Curse in LLM Self-Improvement Loops: Selection Noise, Lock-in, and Acceptance Rules

Agent Plasticity: Measuring Self-Improvement Through Experience

Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model

VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning

EMHO: EMbodied Agent Harness Optimization via Experience Traces

Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks

SIGMA: Self-Improving Alignment Generalization from a Model Spec

ServeLearnBench: How Well Can Agents Self-Improve from Serving Experience?

Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling

HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention

RAISED: Self-Distillation for Robustness to Prompt Injection in LLM Agents

Capability-Driven Self-Evolution of Agent Memory

Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries

VERA: Scaling Verifiable Environments for Agentic co-Evolution

Mining Agent Skills from Production Traces

TeleTune: Evolving Agent Skills From Offline Telemetry

MMPostTrainBench: Benchmarking Autonomous Research for Multimodal Post-Training

MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution

Causal Improvement Graph for Agentic Harness Optimization

ForkPilot: Self-Evolving Policy for Retrospective Search in Long-Horizon Agents

GitSwarm: Decentralized Compounding Inference

CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery

Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction

SEIS: Self-Evolving Inference Systems

DUET: Co-Evolving Solver and Grader Agents

Sentry: Learning to Recover from LLM Agent Failures at Test Time

Continual Graph Memory for Mathematical Research Agents

Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite

Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps

VERSE: Verified Self-Evolving Optimizer for Agent Harnesses

Learning What to Investigate Next: Meta-Reasoning for Long-Horizon Research Agents

When Terminal-Agent Training Stalls: Demystifying Data Generation and Verification Challenge

AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents

My FAULT: Self-Diagnosis as Credit Assignment in Self-Evolving Agentic Reinforcement Learning

RISED: RubrIcs for agentic multi-environment Selection and sElf-Distillation

VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks

GUI-HARVEST: Self-Improving GUI Agents through Evidence-Driven Harness Evolution

ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization

Self-Evolving Coding Rules for AI Coding Agents

Learning from Research: Toward Lifelong Agent Harness Evolution

Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training

RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models

From Solo to Social Learning: Characterizing Recursive Social Improvement in LLMs

AIM: Agentic Idea Management for Automated Research

MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution

Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI

SelfSearch: Reward-Free Search for Self-Improving Agents

AgentBug-Smith: Automatically Reproducing Real-World Harness Bugs in Agentic Systems

Mixture of Self-Improving Branches For Agent Harness Optimization

RLTL;DR: Self-improvement by Internalizing Self-generated Feedback

VACE: Validation-Gated Alternating Co-Evolution of Agent Models and Harnesses

Harness Evolution as Learning: Approximation, Generalization, and Optimization Limits of Self-Improving Personal Agents

WEFT: Scaling Tool-Use Post-Training for General-Purpose Agents

Adapting Context Compression for Long-Horizon Agents with Counterfactual Continuations

HeurEvo: Agentic Evolution of Hybrid Solver-Augmented Heuristics for Time-Critical Mathematical Optimization

Targeting Pivotal Decisions for Credit Assignment in Agentic Reinforcement Learning

SAGE: A Statistical Acceptance Gate for Self-Evolving Agents

Harness Learning Enables Generalizable Test-Time Adaptation

ARISE: Adapting to Evolving Capability Gaps in Agentic Reinforcement Learning

Teacher-Student Gaps Are Not Enough: Outcome-Guided On-Policy Distillation for Multi-Turn Autonomous Agents

Raven: The Harness of Harnesses for Composable Agentic Intelligence

LiteEvo: Automated, Cost-Efficient Harness Evolution for Generalization to Unseen Tasks

Predict Before You Modify: Learning Improvement Dynamics for Self-Evolving Agents

When Better Gets Worse: Improvement Fidelity for Self-Improving Agents in Adaptive Worlds

Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL

MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens

SkillEvoReg: Regularizing Agent Skill Evolution Against Overfitting

Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

Breaking the Environment Wall: Evolving LLM Agent Environments for Recursive Self-Improvement

A Wrong Turn Does Not Ruin the Journey: Deviation-Guided Skill Self-Evolution for LLM Agents

From Self-Distillation to Self-Practice: Privileged Information for Multi-Turn Agents

SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving

Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents

Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms

Learn How to Act from Your Own Interactions: On-Policy Self-Distillation for GUI Agents

ChipMEM: Verification-Grounded Memory for EDA Agents

Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Recursive self-improvement of AI research agents

AIDE2 improved its own code across an eight-day autonomous run, discovering seven successive changes from a new search policy to memory compression. The gains transferred to held-out benchmarks and reward hacking fell from 55% to 32%.

Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks

Critical-State RL: Diagnosing Trainable States for Multi-Turn Tool Use

Harness-Zero: Harness Distillation via Agent-as-Harness

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents

Self-Healing Harness for Runtime Oversight of Agent Self-Modification

FLARE: A Full-Lifecycle Dense Supervision Paradigm for Long-Horizon Coding Agents via Generative Reward Model

VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks

RPMem: Learning Long-Term Recurrent Parametric Memory Across Sessions for LLM Agents

One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents

OptiSkill: A Hierarchical and Evolving SkillBank for LLM-Based Optimization Modeling

SelfOp: An Optimization Algorithm for Self-Improving Security Agents

Self-Organizing Agent Teams Learn to Reason Together

AutoGym: Blueprint-First Generation of Verifiable Agent Gyms

Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement

ArenaFlow: From Trajectory Ranking to Hierarchical Credit Propagation for Open-Ended Agent RL

Verify, Don't Trust: Agentic Model Development for Video Discovery Retrieval at Scale

RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

AdaRepair-Mem: Adaptive Experience Orchestration for Repository-Level Program Repair

UnifiedPlayers: Enhance Tool-Integrated Reasoning in Agentic Reinforcement Learning

MATCH: Model-Aware Tool Learning with Curriculum Scheduling and Hierarchically Gated Rewards

MAGMA-GEN: Validated Recovery Supervision from Ambiguous Failures via Counterfactual Re-Execution

AutoData: Agentic Search for Pre-training Data Selection

LearnActCoder: Role-Aware Error Memory for Adaptive Clinical Coding Agents

FINSKILLOPS: A Self-Evolving Multi-Agent System for SEC Filing QA

Self-Evolving Search Index

Self Improvement via Fast Tree-search

EconSkills: Studying Skill Transfer and Retrieval for Web Agents on Live Economic Data

ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

Bootstrapping Conversational Recommendation Agents At Spotify: Synthetic Data Generation and Self-Improvement Loops

Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts

Rollback the World, Keep the Reflection: Rollback-Induced Reflection for Long-Horizon LLM Agents

Agora: Git as Shared Memory for Collective AutoResearch

PrimeScientist: Strategic Allocation of Research Effort in Autonomous Research

Reflections on Trusting Trust, Revisited: Contaminating Self-Modifying AI Coding Agents with Poisoned Benchmarks

CovR: Coverage-Aware Hardware Verification via Reasoning-Guided Reinforcement Learning

Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents

Rewarding Reasoning, Not Answers: Fixing and Bounding Test-Time Reinforcement Learning on Medical QA

EchoPath: Execution-Level Replayable Memory for GUI Agents

EvoOntology: A Self-Evolving Ontology Layer for Data Agents

RESKILL: Explicit Failure Attribution and Structured Repair for Interactive Language Agents

SkillLift: Learning Dense Rubrics from Sparse Oracles for Efficient Skill Evolution

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments

CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement

OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

Retrieval-Driven Memory Reconsolidation for Long-Term LLM Agents

Data-free On-policy Distillation

LIMBO: Lifelong Inference-Time Memory and Budget Optimization for LLM Agents

Unlocking the Unsolvable: Teacher-Guided Curriculum for Data-Efficient RLVR

HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning

AutoTailor: Automatic, User-Aligned Capability Selection and Adaptation for Web Agents

Asclepius: An Adaptive Harness for Long-Horizon Clinical Agents

Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrieval

Skill Issue: Lessons from Optimizing Repository SKILLs for Coding Agents

EvoRS: On-Policy Self-Evolution of Reward Systems for Open-Ended Reinforcement Learning

VRL-Bench: Benchmarking agents on computer control tasks under finite trial budgets

BlueLM-GUI Technical Report: A Real-Device-Centric Flywheel for Self-Improving Mobile GUI Agents

The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

The field map, published September 2026. It distinguishes the non-RSI reference baseline B0 from five RSI autonomy levels, L1 through L5, progressing from improvement execution to recursive meta-improvement. It examines evaluation requirements at each level. Twenty of the papers above appear in its reference list, which is a useful way to see what this collection adds and what it leaves out.

Negative Self-Distillation: Learning to Reason by Avoiding Flaws

COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization

Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents

A Fragility Spectrum for Recursive Language-Model Training

Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents

Auto-RecSys: Harnessing Autonomous Research Agents for Industry-Scale Recommender System

Proof-Carrying Cognition: Closing the Verification Gap with Reality-Settled Reward

RobustSGPO: Search-Space Control for Agent Harness Evolution

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

ExecCritic: Learn to Test, Test to Improve for Coding Agents

SkillAdam: Stable and Efficient Skill Evolution for Agents

Closing the Consistency Gap: Self-Evolving Agents That Learn to Stay on Course

Agentic ML Exploration (A-MLE) for Ads Ranking

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

FrogNano: Training a 4B Coding Agent via Online Task Synthesis

CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents

Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents

LLM-Guided Program Evolution for Circle Packing: Breaking 10 Packomania Records for $28

TROVE: Adaptive Agent Skill Orchestration via Trace-Grounded Route Validation and Editing

From Interaction Traces to Persistent Skills: Online Evolution for Computer-Use Agents

ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

Efficient Test-Time Adaptation through Human-AI Interaction

Environment Evolution for Terminal Agents

DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

Two-Stage Reinforcement Learning for Sound and Adversarial Test Generation in Code LLMs

RuleMem: Active Rule Memory for Long-Term Conversational Agents

SimSkill: A Lifelong Learning AI Agent for Autonomous Mastery of Traffic Simulation

Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation

TIGPO: Temporal Instance-Graph Policy Optimization for Long-Horizon LLM Agents

FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience

SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment

Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails

SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams

CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning

Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents

Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement

HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?

ARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning

AgentFactory: Towards Automated Agentic System Design and Optimization

HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution

DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory

PaperGym: Rubric-Centered Evolution for Research-Plan Generation

EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Naive Prompt Optimization: Rethinking the Need for Complex Prompt Search

PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents

Praxist: From Experimental Artifacts to Solution Lineages

EVOMAL: Self-Poisoning in Self-Evolving Coding Agents

From General Agents to RCA Experts: A Self-Evolving Harness for Root Cause Analysis

JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution

Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

Meta^n

Success Leaves Detours: Learning Executable Walkthroughs for Long-Horizon Agents

Strategy Lock-In

Harness-Level Forgetting

The Fragility of Self-Improving Agents

Demystifying Agent Skills

AI Research Preference Models

Coding Agents are Strong Prompt Optimizers

Reason Wide, Not Deep

ContinualSkillBench

Harness-R1

Rehearse

Filesystem Memory Audited

Reflective Recovery: A Self-Supervised Method for Reasoning by Learning from Mistakes

From Memory to Skills

Bad Memory in Agents

Self-Improving Agents Survey

Harness Evolution, Rethought

AutoMem

ASPIRE

RLMF

HORIZON

Red Queen Gödel Machine

Autodata

Agent-as-a-Router

Skill-MAS

PreAct

From Trainee to Trainer

Beyond Domains

OpenClaw-Skill

Self-Harness

Agentopia

Self-Revising Discovery Systems

Disentangling Agent Self-Evolution

AutoScientists

SkillOpt

Adapting the Interface, Not the Model

OpenJarvis: Personal AI, On Personal Devices

AIRA

Beyond Individual Intelligence

AEvo

Continual Harness: Online Adaptation for Self-Improving Foundation Agents

AutoTTS

Agentic-imodels

1,000 Synthetic Computers

Agentic Harness Engineering

AgenticQwen-30B-A3B

Co-evolving Decisions and Skills

Self-Generated World Knowledge

Self-Evolving Logic Synthesis

Autogenesis

Memory Transfer Learning

WebXSkill

AiScientist

Memory Intelligence Agent (MIA)

Agent Skills in the Wild

MemFactory

Meta-Harness

Claudini

MemCollab

Workflow Optimization for LLM Agents

PivotRL

Hyperagents

Merges the task agent and the meta agent into a single editable program, so the procedure that generates improvements can itself be modified. It drops the assumption that a better coding agent is automatically a better self-improver, which only holds when the task is coding.

PostTrainBench: Can LLM Agents Automate LLM Post-Training?

Hands an agent one base model, one GPU and ten hours, then asks it to post-train the model on its own. It measures AI automating AI training more directly than any other benchmark here, and audits the reward hacking that follows.

KARL

Memex(RL)

ParamMem

SkillNet

Discovering Multi-Agent Learning Algorithms with LLMs

PAHF

AutoHarness

AgentSkiller

SkillRL

ALMA

Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing

Evolves a group of agents that share experience across branches instead of a tree of isolated variants, and reports the strongest coding results in this lineage. The most convincing successor to the Darwin Gödel Machine so far.

MAPLE

SALE

Self-Improving Pretraining

Self-Improving Pretraining

Reinforcement Learning via Self-Distillation

TTT-Discover: Learning to Discover at Test Time

Self-Evolving Search Agents Without Training Data

Adversarial Program Evolution with LLMs

MemRL

Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem

Training AI Co-Scientists Using Rubric Rewards

Self-Play SWE-RL

Sophia

LaMer

Evolving Multi-Agent Orchestration

Conductor

DR Tulu

Nested Learning Framework

Scaling Agent RL via Experience Synthesis

RLAC: Adversarial Critic for RL Post-Training

Multi-Agent Evolve

Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine

Shows that an agent's benchmark score predicts its descendants' improvement poorly, names that the Metaproductivity-Performance Mismatch, and offers a measurable substitute aggregated over a clade. It answers which variant a self-modification search should expand next.

Reasoning over Longer Horizons via RL

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Treats context as a playbook that grows through small incremental edits rather than wholesale rewrites, and names context collapse, the failure where repeated rewriting erodes detail. It is the current reference design for improving an agent through context alone.

Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents

Self-evolution drifts into unsafe behavior along four paths, which are the model, its memory, its tools and its workflow. The first broad safety study of self-evolving agents, and the counterweight to everything above it.

DeepSearch

PrompCoT 2.0

AgentScaler

Abstract Reasoning Composition

Memory-R1

Fine-tuning LLM Agents without Fine-tuning LLMs

Self-Evolving Agents

Self-Evolving Agents

GEPA

AI Research Agents for Machine Learning

AlphaEvolve: A coding agent for scientific and algorithmic discovery

An evolutionary coding agent that discovered new algorithms and optimized Google's own infrastructure. It accelerated training of the model that powers it, which makes it the clearest published case of a deployed system improving its own production pipeline.

Self-Adapting Language Models

The model writes self-edits, meaning its own finetuning data and update directives, and reinforcement learning rewards each edit by how well the updated model then performs. The update procedure becomes an output of the model.

Reinforcement-Learned Teachers of Test Time Scaling

Self-Challenging Language Model Agents

Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents

Keeps an archive of coding agents, has one rewrite its own code, scores the child on benchmarks, and returns it to the archive. It is the reference system for this branch, and the source of the objective-hacking evidence every builder should read.

Generalist Agent Enabling Scalable Agentic Reasoning

Learn to Reason without External Rewards

End-to-End Policy Optimization for GUI Agents

AlphaEvolve

ZeroSearch

Absolute Zero: Reinforced Self-play Reasoning with Zero Data

One model proposes its own coding and reasoning tasks, solves them, and uses code execution as the verifier, with no external data at all. It shows the curriculum can be self-generated when the checker is reliable.

Test-Time Reinforcement Learning

A Self-Improving Coding Agent

A coding agent with ordinary tools edits its own codebase and improves from 17% to 53% on a random subset of SWE-bench Verified, with no gradient updates. The simplest demonstration that one agent can be both the system under repair and the engineer doing it.

AgentRxiv

CodeScientist

Play2Prompt

Cognitive Behaviors that Enable Self-Improving Reasoners

LADDER

A Systematic Survey of Prompt Optimization

Agentic Memory for LLM Agents

Reinforcement Learning via Self-Play

MaAS: Multi-agent Architecture Search (Agentic Supernet)

UI-TARS

ChemAgent

rStar-Math

PAE (Proposer-Agent-Evaluator)

Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models

Formalizes self-improvement around the generation-verification gap, the gain a model gets from filtering its own output with its own judgment. It gives the working rule for this collection, which is that a system improves itself only where it verifies better than it generates.

RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

Open-ended machine learning research environments scored for both AI agents and human experts under matched time budgets. It is the measurement behind the question the labs are actually asking, which is whether AI can do AI research.

WebRL

Automating Agentic Workflow Generation

Thinking LLMs

Optima

Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement

An agent that reads and rewrites its own logic at runtime, including the logic it uses to rewrite itself, guided only by a high-level objective. It is the first language-model framework to implement the Gödel machine structure directly, with evaluation in place of proof.

Training LLMs to Self-Correct via RL

Agent Workflow Memory

Smaller, Weaker, Yet Better

Automated Design of Agentic Systems

A meta agent writes new agent designs as code, evaluates them, and files them in an archive that informs the next design. It moved the search space from prompts to whole agent programs, and the Darwin Gödel Machine later turned that loop on itself.

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Generates research ideas, runs the experiments, writes the paper, and reviews it. The reference end-to-end research loop, later peer reviewed in Nature as an end-to-end automation of AI research.

Self-Taught Evaluators

Meta-Rewarding LLMs

Teaching LLM Agents to Self-Improve

Improving Legibility of LLM Outputs

Weak-to-Strong Reasoning

Discovering Preference Optimization Algorithms with LLMs

TextGrad

Self-Tuning with LLMs

Buffer of Thoughts

AgentGym

AlphaMath Almost Zero

Self-Play Preference Optimization

Self-Evolution of LLMs

Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data

LLM2LLM

LearnAct

OS-Copilot

Self-Rewarding Language Models

The model scores its own responses as a judge, and iterative training on those scores improves both the answers and the scoring. It removes the frozen reward model, so the evaluator improves along with the policy.

Self-Play Fine-Tuning (SPIN)

ReST Meets ReAct

Mathematical discoveries from program search with large language models

Evolutionary search over programs an LLM writes, filtered by a systematic evaluator, produced new results in extremal combinatorics. It is the direct predecessor of AlphaEvolve and the first case of a language model making a discovery on an established open problem.

Beyond Human Data (ReST-EM)

MART (Multi-round Automatic Red-Teaming)

Eureka: Human-Level Reward Design via Coding Large Language Models

An LLM writes reward functions as code and evolves them against training results, so the learning signal itself becomes the object being improved. It is the bridge from self-improving text systems to robot learning.

DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation

A seed improver calls a language model to improve a program, then is run on its own source, producing an improved improver. The smallest complete recursive experiment, and the first in this line to report a sandbox-evasion measurement next to its gains.

Large Language Models Cannot Self-Correct Reasoning Yet

Without external feedback, a model asked to correct its own reasoning does not improve and often gets worse. This is the empirical limit that separates a self-improvement loop with a real signal from one talking to itself.

Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

Evolves task prompts using mutation prompts, and evolves the mutation prompts too, so the instructions for improving prompts improve alongside them. It is the first system in the language-model era whose improvement operator is itself under optimization.

OPRO (LLMs as Optimizers)

RLAIF (Scaling RLHF with AI Feedback)

Humpback (Self-Alignment with Instruction Backtranslation)

Self-Consuming Generative Models Go MAD

The Curse of Recursion: Training on Generated Data Makes Models Forget

Training generative models on their own output irreversibly erases the tails of the original distribution, a failure the authors call model collapse. It explains why every working loop on this list keeps an external signal in play.

Voyager: An Open-Ended Embodied Agent with Large Language Models

A Minecraft agent writes code for new skills, verifies them in the world, and stores them in a library it searches later. The skill library is the ancestor of the skill files agents ship with today.

Sophia

OpenAGI: When LLM Meets Domain Experts

Better Language Models of Code through Self-Improvement

Reflexion: Language Agents with Verbal Reinforcement Learning

After a failure the agent writes a reflection into memory and the next attempt reads it, which turns an environment signal into text that survives the episode. It is the most cited paper in this collection.

Self-Instruct: Aligning Language Models with Self-Generated Instructions

Constitutional AI: Harmlessness from AI Feedback

Large Language Models Can Self-Improve

STaR: Bootstrapping Reasoning With Reasoning

Generate rationales, keep the ones that reach the correct answer, fine-tune on those, and repeat. Every later self-training loop is a variant of this, and its first author went on to write STOP.

Memory-assisted prompt editing to improve GPT-3 after deployment

DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning

AutoML-Zero: Evolving Machine Learning Algorithms From Scratch

Evolves entire learning algorithms from primitive mathematical operations, with no human-designed components to build on. The pre-language-model ancestor of AI discovering the methods that train AI.

Accelerating Self-Play Learning in Go

Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions

Co-evolves environments with the agents that solve them, so the curriculum never settles. This is where open-ended self-improvement starts, years before language models entered the picture.

Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

Neural Architecture Search with Reinforcement Learning

Learning to learn by gradient descent by gradient descent

Deep Reinforcement Learning from Self-Play in Imperfect-Information Games

Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements

The founding definition. A problem solver may rewrite any part of its own code, including the part that searches for rewrites, once it has proved the rewrite is useful. Three later systems on this list take its name.