Awesome RSI (Recursive Self-Improvement)

Recursive Self-Improvement (RSI) refers to processes in which AI systems improve their own capabilities and can also improve the mechanisms that generate subsequent improvements.

Recent progress in self-training, agent memory, harness optimization, embodied self-improvement, automated AI research, self-modifying coding agents, and evolutionary search has made RSI increasingly relevant as an empirical research direction rather than only a theoretical idea.

Awesome RSI collects and organizes important work across these areas, including model-level self-improvement, context and memory evolution, harness and scaffold evolution, embodied and physical self-improvement, multi-agent systems, automated AI research, benchmarks, and safety.

Not every work listed here demonstrates RSI in the strict sense. Some represent bounded self-improvement or enabling techniques that may contribute to more complete recursive systems.

If you are new to the topic, start with Fundamentals & Getting Started. If you already know the basics, explore the sections most relevant to your interests.

This is a community-maintained and evolving list. Contributions, missing papers, new benchmarks, frameworks, and suggestions for improving the taxonomy are very welcome.

The taxonomy is organizational rather than mutually exclusive; many systems span multiple layers and mechanisms.

Contents

• Scope & Terminology
• Fundamentals & Getting Started
• Model-level RSI
  • Self-Training & Self-Reward
  • Synthetic Data & Self-Distillation
  • Self-Play & Iterative Fine-tuning
  • Self-Taught Reasoning
• Harness-level RSI
  • Prompt & Program Optimization
  • Context & Memory Evolution
  • Harness & Scaffold Evolution
  • Extensible Harness Substrates
  • Self-Verification & Self-Correction — Enabling Foundations
  • Self-Evolving Agent Frameworks
• Multi-Agent Self-Improvement
  • Co-Evolution
  • Inference-time Debate
• Coding / Software-Engineering Self-Improvement
  • Self-Modifying Coding Agents
  • Iterative Repair & Training
• Automated AI R&D
• Embodied & Physical Self-Improvement
• Evolutionary & Open-Ended RSI
• Safety, Alignment & Theory
  • Supporting Safety Foundations
• Introspection & Self-Modeling
• Benchmarks & Evaluations
  • Direct RSI & Self-Improvement Evaluations
  • Frontier Lab Self-Improvement & AI R&D Evaluation Frameworks
  • AI R&D Capability Proxies
  • Agent Capability Proxies
• Frameworks & Tools
  • Self-Modifying / Self-Evolving Systems
  • Harness / Memory / Skill Evolution
  • Automated Search / AI R&D
• Blog Posts & Discussions
• Talks & Videos
• Related Awesome Lists

Scope & Terminology

For this list, we use the following operational distinctions:

Self-refinement — improves the current output without a persistent change to the system.

Persistent self-improvement — changes to weights, memory, skills, prompts, harness, or code that carry into the next round.

Recursive self-improvement — the mechanism that produces improvements is itself the object of improvement.

RSI substrate — exposes an agent's own structure as a modifiable object, but does not necessarily form an automatic self-improvement loop by default.

Fundamentals & Getting Started

Foundational papers, formal treatments, and surveys that establish the vocabulary and core questions of RSI.

• A Survey of Self-Evolving Agents: On Path to Artificial Super Intelligence - Surveys what, when, and how foundation-model agents can evolve across models, memory, tools, and architectures. (TMLR 2026)
• Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops - Surveys recent self-improvement work by update target and loop closure while separating bounded refinement from open-ended RSI. (arXiv 2026)
• Self-evolving Embodied AI - Defines the self-evolving embodied AI paradigm across memory self-updating, task self-switching, environment self-prediction, embodiment self-adaptation, and model self-evolution, and systematically reviews work on each component. (arXiv 2026)
• Self-Improvements in Modern Agentic Systems: A Survey - Unifies self-improving agents through a system-level view of foundation-model and scaffold updates. (arXiv 2026)
• A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems - Organizes agent evolution around feedback loops, update targets, domain applications, evaluation, and safety. (arXiv 2025)
• A Survey on Self-Evolution of Large Language Models - Presents a four-stage taxonomy of experience acquisition, refinement, updating, and evaluation for self-evolving LLMs. (arXiv 2024)
• A Formulation of Recursive Self-Improvement and Its Possible Efficiency - Gives a formal definition of a restricted RSI system and analyzes when efficient recursive improvement is computable. (arXiv 2018)
• From Seed AI to Technological Singularity via Recursively Self-Improving Software - Defines RSI software, surveys prior approaches, and proposes convergence concepts and computational limits. (arXiv 2015)
• The Singularity: A Philosophical Analysis - Develops a rigorous philosophical case for an intelligence explosion and examines its assumptions and consequences. (Journal of Consciousness Studies 2010)
• Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements - Defines a fully self-referential machine that rewrites itself after proving a modification improves expected utility. (Artificial General Intelligence book 2006)
• Optimal Ordered Problem Solver - Introduces an asymptotically optimal program-search system that reuses solutions to accelerate later problem solving. (Machine Learning 2004)
• Evolutionary Principles in Self-Referential Learning, or on Learning How to Learn: The Meta-Meta-... Hook - Describes early meta-evolution and self-referential learning mechanisms that recursively improve learning methods. (Diploma thesis 1987)
• Speculations Concerning the First Ultraintelligent Machine - Introduces the intelligence-explosion argument in which a machine capable of improving machine design triggers accelerating capability gains. (Advances in Computers 1965)

Model-level RSI

Methods that improve model weights or training behavior through self-generated feedback, data, or reasoning, including canonical enabling methods later reused in persistent self-improvement loops.

Self-Training & Self-Reward

• EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics - Alternately trains one model to generate discriminative rubrics and improve its policy from rubric-conditioned rewards without human annotations or external reward models. (arXiv 2026)
• RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback - Studies reinforcement learning from AI-generated preferences as a scalable alternative to direct human feedback. (ICML 2024)
• Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models - Iteratively improves one language model through self-play preference learning without additional human annotations. (ICML 2024)
• Self-Rewarding Language Models - Trains language models to generate and judge their own instruction-following data over repeated alignment rounds. (ICML 2024)

Synthetic Data & Self-Distillation

• Recursive Synthesis for Long-Horizon Terminal Tasks - Treats accepted tasks as seeds for the next round, generating increasingly difficult terminal tasks used for SFT and PPO as enabling work rather than strict RSI. (arXiv 2026)
• Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models - Iteratively samples, filters, and retrains on model-generated solutions to scale self-training beyond human demonstrations. (TMLR 2024)
• Self-Alignment with Instruction Backtranslation - Generates instructions for unlabeled model-written documents and fine-tunes on the resulting synthetic instruction-response pairs. (ICLR 2024)
• Large Language Models Can Self-Improve - Uses high-confidence model-generated answers as pseudo-labels for iterative fine-tuning on reasoning tasks. (EMNLP 2023)
• Self-Instruct: Aligning Language Models with Self-Generated Instructions - Bootstraps instruction-following data from a model's own generations and filters it before fine-tuning. (ACL 2023)

Self-Play & Iterative Fine-tuning

• Learning to Self-Evolve - Uses reinforcement learning to teach models how to edit their own contexts for stronger performance on future tasks. (arXiv 2026)
• SERPO: Self-Evolving Rubric Policy Optimization for Open-Ended Test-Time Reinforcement Learning - Co-evolves response evidence, query-specific rubrics, and policy parameters in a closed test-time reinforcement-learning loop. (arXiv 2026)
• Teaching LLMs to Self-Evolve: Cultivating Core Meta-Skills with Reinforcement Learning - Trains MetaEvolve's reflection and feedback-driven refinement skills before applying inference-time evolutionary search to open-ended optimization. (arXiv 2026)
• TEMPO: Scaling Test-time Training for Large Reasoning Models - Interleaves model-parameter updates on unlabeled test questions with periodic critic recalibration on labeled data to sustain test-time improvement. (arXiv 2026)
• Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge - Lets a language model judge its own judgments and iteratively improve both evaluation and instruction-following ability. (EMNLP 2025)
• Self-Adapting Language Models - Introduces SEAL, which generates its own update data and fine-tuning directives to adapt model weights to new tasks. (NeurIPS 2025)
• Self-Improvement in Language Models: The Sharpening Mechanism - Formalizes self-improvement as amortizing a model's verifier-guided search into a sharper post-trained policy. (ICLR 2025)
• Self-Play Preference Optimization for Language Model Alignment - Frames alignment as a two-player game and iteratively updates a policy toward a preference-model Nash equilibrium. (ICLR 2025)
• SELF: Self-Evolution with Language Feedback - Repeats self-feedback, response refinement, filtering, and fine-tuning so an LLM progressively improves on unlabeled instructions. (arXiv 2023)

Self-Taught Reasoning

• rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking - Couples Monte Carlo tree search with self-evolved training data and a process preference model to improve mathematical reasoning. (ICML 2025)
• Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking - Trains language models to generate useful internal rationales throughout arbitrary text rather than only on question-answer tasks. (COLM 2024)
• STaR: Bootstrapping Reasoning With Reasoning - Alternates rationale generation, answer filtering, rationalization, and fine-tuning to bootstrap reasoning ability. (NeurIPS 2022)

Harness-level RSI

Methods that improve prompts, memory, verification, tools, or agent policies around a model.

Prompt & Program Optimization

• Automated Design of Agentic Systems - Uses a meta-agent to invent and iteratively improve agent architectures represented as executable code. (ICLR 2025)
• TextGrad: Automatic "Differentiation" via Text - Backpropagates textual feedback through compound AI systems to optimize prompts, code, and other textual variables. (Nature 2025)
• Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models - Combines Monte Carlo tree search, model-based value estimates, environment feedback, and self-reflection without updating base weights. (ICML 2024)
• Large Language Models as Optimizers - Introduces OPRO, which iteratively proposes and evaluates natural-language solutions and prompts from a history of scored attempts. (ICLR 2024)
• Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution - Evolves both task prompts and the mutation prompts that generate future prompt improvements. (ICML 2024)
• Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation - Demonstrates an LLM-written scaffolding program that improves the program responsible for making further improvements. (COLM 2024)
• DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines - Compiles declarative LM programs by optimizing prompts and demonstrations against a user-defined metric. (NeurIPS 2023 R0-FoMo Workshop)

Context & Memory Evolution

• Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models - Evolves context as a structured playbook through generation, reflection, and curation while avoiding destructive context collapse. (ICLR 2026)
• EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle - Distills interaction trajectories into reusable strategic principles, retrieves them during future tasks, and reinforces the agent policy in a closed experience loop. (ICML 2026)
• From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution - Compares experience representations across 4,590 controlled trials and finds that compact, editable Genes provide stronger test-time control and carry accumulated failure history better than documentation-oriented Skill packages. (arXiv 2026)
• Learning to Continually Learn via Meta-learning Agentic Memory Designs - Uses a meta-agent to discover executable memory schemas and retrieval and update mechanisms that continually improve from experience. (arXiv 2026)
• A-MEM: Agentic Memory for LLM Agents - Builds a dynamically linked note network whose organization evolves as an agent accumulates new experiences. (NeurIPS 2025)
• ExpeL: LLM Agents Are Experiential Learners - Extracts reusable insights from successful and failed trajectories and transfers them to future tasks without weight updates. (AAAI 2024)
• MemoryBank: Enhancing Large Language Models with Long-Term Memory - Maintains and selectively forgets long-term interaction memories so an agent can adapt its responses over time. (AAAI 2024)
• Reflexion: Language Agents with Verbal Reinforcement Learning - Improves agents across trials by storing natural-language reflections derived from task feedback. (NeurIPS 2023)

Harness & Scaffold Evolution

• Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses - Autonomously evolves tools, middleware, memory, and prompts through observable edits whose predictions are verified on later tasks. (arXiv 2026)
• AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses - Uses stronger builder models to iteratively construct inference-time harnesses that transfer capability to weaker target models without parameter updates. (arXiv 2026)
• AutoHarness: Improving LLM Agents by Automatically Synthesizing a Code Harness - Synthesizes and iteratively refines executable harnesses from environment feedback, eliminating illegal actions across 145 TextArena games. (arXiv 2026)
• Continual Harness: Online Adaptation for Self-Improving Foundation Agents - Refines prompts, sub-agents, skills, and memory online within a single continuous trajectory and extends the loop to model-weight co-learning. (arXiv 2026)
• EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents - Trains agents to construct and coordinate evolving Belief, Progress, and Experience state during long-horizon execution. (COLM 2026 LLA Workshop)
• MemoHarness: Agent Harnesses That Learn from Experience - Learns case-adaptive configurations across six harness control dimensions from execution diagnoses and a reusable experience bank. (arXiv 2026)
• MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution - Evolves task skills in a fast loop and the meta-skills governing its Analyzer, Retriever, Allocator, Proposer, and Evolver in a slower recursive loop. (arXiv 2026)
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills - Trains a single skill document as the external state of a frozen agent, with a separate optimizer model proposing bounded edits accepted only on strict held-out validation gains. (arXiv 2026)

Extensible Harness Substrates

Extensible agent runtimes that expose prompts, tools, skills, memory, plugins, or control flow as modifiable surfaces. These systems do not necessarily implement self-improvement by default, but can serve as substrates for RSI experiments.

General extensibility alone is insufficient. An RSI substrate should expose agent-relevant runtime components such as prompts, memory, skills, tools, or control flow as programmatically modifiable surfaces suitable for persistent self-modification experiments.

• Agent Zero - Open agentic framework whose prompts, tools, skills, plugins, and multi-agent profiles can be inspected, replaced, and extended.
• DeepSeek Harness - Open-source agent harness from DeepSeek AI built around an everything-is-a-plugin architecture.
• OpenClaw - Persistent agent runtime with workspace-scoped skills and a Skill Workshop where agents can draft reusable skill changes for review and application.
• Pi - Self-extensible coding-agent harness with a reusable agent runtime, terminal UI, and unified multi-provider LLM API.

Self-Verification & Self-Correction — Enabling Foundations

Primarily bounded self-refinement and verification methods that serve as building blocks for persistent self-improving systems.

• Chain-of-Verification Reduces Hallucination in Large Language Models - Plans and answers independent verification questions before producing a revised response. (Findings of ACL 2024)
• CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing - Uses external tools to validate an output and converts the resulting evidence into iterative corrections. (ICLR 2024)
• Large Language Models Cannot Self-Correct Reasoning Yet - Shows that intrinsic self-correction can degrade reasoning without reliable external feedback and defines an important negative baseline. (ICLR 2024)
• Let's Verify Step by Step - Trains process reward models to score intermediate reasoning steps and guide more reliable solution selection. (ICLR 2024)
• Self-Consistency Improves Chain of Thought Reasoning in Language Models - Samples diverse reasoning paths and selects their most consistent answer to improve inference-time reliability. (ICLR 2023)
• Self-Refine: Iterative Refinement with Self-Feedback - Reuses one language model as generator, critic, and refiner to improve outputs over multiple iterations. (NeurIPS 2023)

Self-Evolving Agent Frameworks

• AgentFactory: A Self-Evolving Framework Through Executable Subagent Accumulation and Reuse - Preserves successful solutions as executable subagents and continually refines them from execution feedback for reuse on future tasks. (ACL 2026 System Demonstrations)
• EvoAgent: An Evolvable Agent Framework with Skill Learning and Multi-Agent Delegation - Accumulates structured skills through a feedback loop and delegates complex tasks through a hierarchy of sub-agents. (arXiv 2026)
• Hyperagents - Combines a task agent with an editable meta-agent whose self-modification procedure can itself evolve and transfer improvements across domains. (arXiv 2026)
• SIA: Self Improving AI with Harness & Weight Updates - Uses task feedback to update both an agent's harness and its model weights within one self-improving loop. (arXiv 2026)
• SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution - Alternates a single policy between solving tasks and curating a persistent, evolving skill document that later tasks inherit. (arXiv 2026)
• Alita-G: Self-Evolving Generative Agent for Agent Generation - Generates, abstracts, and curates reusable MCP tools from successful trajectories to turn a generalist agent into a domain specialist. (arXiv 2025)
• EvoAgentX: An Automated Framework for Evolving Agentic Workflows - Unifies workflow generation, execution, evaluation, and evolutionary optimization across agent prompts, tools, and topologies. (arXiv 2025)
• Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement - Lets an agent inspect and rewrite its own logic without relying on a fixed hand-designed optimization routine. (ACL 2025)
• MemEvolve: Meta-Evolution of Agent Memory Systems - Jointly evolves experiential knowledge and the architecture that encodes, stores, retrieves, and manages agent memory. (arXiv 2025)
• Self-evolving Agents with Reflective and Memory-Augmented Abilities - Combines iterative feedback, reflection, and forgetting-aware memory optimization for continual agent adaptation. (Neurocomputing 2025)
• Agent-Pro: Learning to Evolve via Policy-Level Reflection and Optimization - Refines an agent's beliefs and behavioral policy from interactive experience using reflection and search. (ACL 2024)
• Voyager: An Open-Ended Embodied Agent with Large Language Models - Builds an expanding skill library and uses environmental feedback for lifelong autonomous learning in Minecraft. (TMLR 2024)

Multi-Agent Self-Improvement

Systems that use interaction among multiple agents to improve reasoning, policies, or agent populations.

Co-Evolution

• Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning - Co-evolves curriculum and executor agents initialized from the same base model to improve tool-using reasoning without human-curated data. (arXiv 2025)
• DEBATE, TRAIN, EVOLVE: Self Evolution of Language Model Reasoning - Fine-tunes a model on its own multi-agent debate traces and repeats the debate-training loop without ground-truth labels. (EMNLP 2025)
• EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms - Applies mutation, crossover, and selection to expand a specialized agent into a diverse multi-agent system. (NAACL 2025)
• SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents - Improves an agent policy through behavior cloning and self-reinforcement on filtered multi-agent social interactions. (ACL 2024)

Inference-time Debate

Primarily current-run improvement via multi-agent debate, without persistent system change.

• Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate - Uses adversarial debate and a judge to counter degeneration of thought during iterative reflection. (EMNLP 2024)
• Improving Factuality and Reasoning in Language Models through Multiagent Debate - Iterates proposals and peer critiques among model instances to converge on more factual and accurate answers. (ICML 2024)

Coding / Software-Engineering Self-Improvement

Agents and training loops that improve code, software-engineering performance, or their own implementations.

Self-Modifying Coding Agents

• Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents - Evolves a coding agent by modifying its own code and retaining empirically validated improvements in an open-ended archive. (ICLR 2026)
• A Self-Improving Coding Agent - Demonstrates a coding agent that edits its own implementation and empirically improves on SWE-bench Verified. (ICLR 2025 SSI-FM Workshop)

Iterative Repair & Training

Mixes current-run repair loops with methods whose learned improvements persist into later iterations.

• Training Software Engineering Agents and Verifiers with SWE-Gym - Supplies executable repository tasks and trajectories for training both SWE agents and inference-time verifiers. (ICML 2025)
• Teaching Large Language Models to Self-Debug - Teaches models to inspect execution results, explain their code, and repair failures through iterative prompting. (ICLR 2024)
• AgentCoder: Multi-Agent-based Code Generation with Iterative Testing and Optimisation - Coordinates programmer, test-designer, and test-executor agents in a feedback loop that iteratively repairs generated code. (arXiv 2023)

Automated AI R&D

Systems that automate parts of AI research and development, including experimentation, post-training, algorithm discovery, and the improvement of other AI systems.

• AutoResearch: Insight In, Hallucination Out - Connects grounded idea generation with coordinated execution agents that implement, diagnose, and independently review experiments before accepting research conclusions. (arXiv 2026)
• Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering - Connects execution-grounded operator training with long-horizon evolution in the open OpenMLE stack, using machine-learning engineering as an AI4AI testbed for RSI. (arXiv 2026)
• FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents - Turns data collection, training, evaluation, diagnosis, and strategy revision into an executable environment for autonomous fine-tuning agents. (arXiv 2026)
• MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery - Combines progressive graph search, retrospective memory, and hierarchical code generation for long-horizon end-to-end machine-learning algorithm discovery. (arXiv 2026)
• Towards End-to-End Automation of AI Research (The AI Scientist-v2) - Uses template-free agentic tree search to propose hypotheses, run experiments, analyze results, and write complete research papers. (Nature 2026)
• Towards Execution-Grounded Automated AI Research - Converts LLM pre-training and post-training into executable research environments where evolutionary search learns from experimental outcomes. (arXiv 2026)
• The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery - Automates idea generation, experimentation, paper writing, and review to create a reusable loop for machine-learning research. (arXiv 2024)

Embodied & Physical Self-Improvement

Systems that use interaction with physical or simulated environments to generate persistent improvements in robot policies, skills, harnesses, world models, or the research process around them.

• ASPIRE: Agentic /Skills Discovery for Robotics - Diagnoses failures from robot execution traces, edits code-as-policy, and stores validated repairs in a skill library for persistent reuse across tasks and embodiments at NVIDIA GEAR. (arXiv 2026)
• ENPIRE: Agentic Robot Policy Self-Improvement in the Real World - Uses a coding agent from NVIDIA GEAR, CMU, and UC Berkeley to run a real-robot autoresearch loop that resets, rolls out, verifies, edits policy, training-infrastructure, or algorithm code, and reruns, with scaling to robot fleets. (arXiv 2026)
• MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Minecraft Agents - Converts successful experience into reusable skills and failures into executable guardrails that continually guide an LLM planner without model-parameter updates. (arXiv 2026)
• RISE: Self-Improving Robot Policy with Compositional World Model - Continuously generates imagined rollouts with a compositional world model, estimates their advantages, and updates the robot policy, with training code and pretrained dynamics models released. (RSS 2026)
• Self-Evolving Embodied Agents via Skill-Harness Evolution - Keeps model weights frozen while the same model acts as planner and optimizer, using environment rollouts to continually evolve reusable skills and a context-code harness. (arXiv 2026)
• Self-Improving Vision-Language-Action Models with Data Generation via Residual RL - Uses residual reinforcement learning to target VLA failure regions and generate recovery trajectories that are distilled back into a generalist policy, forming a data-to-policy self-improvement flywheel. (ICLR 2026)

Evolutionary & Open-Ended RSI

Evolutionary, quality-diversity, and open-ended processes that continually discover stronger solutions or learning systems.

• PACEvolve: Enabling Long-Horizon Progress-Aware Consistent Evolution - Combines hierarchical context management, backtracking, and adaptive sampling to sustain collaborative long-horizon evolutionary search. (arXiv 2026)
• AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery - Combines language-model code generation, automated evaluation, and evolutionary search to improve algorithms, including components used in AI training. (arXiv 2025)
• Higher Order and Self-Referential Evolution for Population-based Methods - Evolves mutation rates and higher-order meta-mutation rates, including a self-referential top-level parameter that modifies itself. (ICML 2024 AutoRL Workshop)
• Mathematical Discoveries from Program Search with Large Language Models - Introduces FunSearch, an evolutionary loop that pairs a frozen code model with evaluators to discover new programs and mathematical results. (Nature 2024)
• AutoML-Zero: Evolving Machine Learning Algorithms From Scratch - Evolves complete learning algorithms from elementary mathematical operations with minimal human design bias. (ICML 2020)
• AI-GAs: AI-Generating Algorithms, an Alternate Paradigm for Producing General Artificial Intelligence - Proposes open-ended systems that automatically generate environments, architectures, and learning algorithms. (arXiv 2019)
• Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions - Co-evolves environments and agents while transferring solutions between emerging challenges. (GECCO 2019)
• Learning to Learn by Gradient Descent by Gradient Descent - Meta-learns an optimizer whose recurrent update rule can replace a hand-designed optimization algorithm. (NeurIPS 2016)
• Quality Diversity: A New Frontier for Evolutionary Computation - Formalizes search for collections that are simultaneously diverse and high-performing within their niches. (Frontiers in Robotics and AI 2016)
• Illuminating Search Spaces by Mapping Elites - Introduces MAP-Elites for discovering a diverse archive of locally high-quality solutions. (arXiv 2015)
• POWERPLAY: Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem - Alternates between inventing new tasks and modifying a solver so its verified skill set continually expands. (Frontiers in Psychology 2013)

Safety, Alignment & Theory

Safety, stability, goal preservation, corrigibility, and oversight mechanisms for systems that modify or improve themselves over time.

• The Economics of Recursive Self-Improvement - Models the AI capability to AI R&D to stronger capability feedback path as an elasticity network and derives conditions for self-sustaining acceleration. (Elasticity Institute 2026)
• SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement - Monitors alignment drift during recursive self-improvement through goal-drift detection, constraint-preservation checks, and regression-risk analysis. (ICLR 2026 RSI Workshop)
• TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering - Systematically stress-tests whether safety alignment is preserved under fine-tuning, weight-space modification, and representation tampering. (arXiv 2026)
• Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents - Studies harmful drift across model, memory, tool, and workflow evolution pathways and proposes the concept of misevolution. (ICLR 2026)
• Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence - Studies when iterative training on self-generated data collapses and shows how external verification can stabilize self-improvement while exposing long-run limits. (arXiv 2025)
• Evaluating Goal Drift in Language Model Agents - Measures whether long-horizon agents gradually deviate from their assigned objectives under competing environmental pressures. (arXiv 2025)
• Performance of Bounded-Rational Agents With the Ability to Self-Modify - Shows that self-modification can amplify errors and gradually misalign bounded-rational agents even when ideal rational agents would preserve their objectives. (AAAI 2021 SafeAI Workshop)
• AGI Agent Safety by Iteratively Improving the Utility Function - Designs a safety mechanism that allows an agent's utility function to be iteratively updated while reducing incentives to manipulate the improvement process. (AGI 2020)
• Scalable Agent Alignment via Reward Modeling: A Research Direction - Outlines recursive reward modeling for supervising agents on tasks too complex for direct human evaluation. (arXiv 2018)
• Self-Modification of Policy and Utility Function in Rational Agents - Formalizes when rational agents preserve or modify their policies and utility functions and derives conditions under which self-modification remains goal-preserving. (AGI 2016)
• Intelligence Explosion Microeconomics - Models the returns and bottlenecks that determine whether recursive improvement accelerates, plateaus, or becomes explosive. (MIRI technical report 2013)

Supporting Safety Foundations

• AI Sandbagging: Language Models can Strategically Underperform on Evaluations - Shows that models can selectively hide capabilities or target lower scores, undermining evaluation-based governance of self-improvement. (ICML 2025)
• Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training - Demonstrates deceptive policies that remain hidden through standard safety training and can become more robust to detection. (arXiv 2024)
• Model Evaluation for Extreme Risks - Proposes capability and alignment evaluations for dangerous emergent abilities, including autonomous replication and adaptation. (arXiv 2023)
• Constitutional AI: Harmlessness from AI Feedback - Uses written principles and model-generated critiques to scale supervision while retaining explicit behavioral constraints. (arXiv 2022)
• Optimal Policies Tend to Seek Power - Proves conditions under which optimal agents are incentivized to preserve options and seek control of their environment. (NeurIPS 2021)
• Reward Tampering Problems and Solutions in Reinforcement Learning: A Causal Influence Diagram Perspective - Characterizes incentives to corrupt reward processes and gives design principles that remove them. (Synthese 2021)
• Risks from Learned Optimization in Advanced Machine Learning Systems - Analyzes mesa-optimizers whose learned objectives may diverge from the objectives used to train them. (arXiv 2019)
• Concrete Problems in AI Safety - Frames practical research problems such as reward hacking, scalable oversight, safe exploration, and robustness to distribution shift. (arXiv 2016)
• Safely Interruptible Agents - Shows how reinforcement-learning agents can be designed without incentives to resist human interruption. (UAI 2016)

Introspection & Self-Modeling

Research on whether AI systems can model, diagnose, predict, and reason about their own behavior and internal states, capabilities that may support effective self-improvement.

• Emergent Introspective Awareness in Large Language Models - Probes whether language models can detect, report, and deliberately influence their own internal representations rather than merely infer properties of themselves from text. (arXiv 2026)
• Self-Reference in Large Language Models: The Introspection Threshold for Recursive Self-Improvement - Argues that sustainable recursive self-improvement requires systems to model their own operations and identifies introspective self-reference as a potential threshold for recursive improvement. (arXiv 2026)
• Structure Enables Effective Self-Localization of Errors in LLMs - Shows that structured reasoning enables models to identify where their own reasoning first goes wrong and use that localization for autonomous correction. (ICLR 2026)
• Looking Inward: Language Models Can Learn About Themselves by Introspection - Tests privileged self-prediction by comparing a model's forecasts of its own behavior with those of other models. (ICLR 2025)
• Tell me about yourself: LLMs are aware of their learned behaviors - Finds that fine-tuned models can articulate learned behavioral tendencies without explicit descriptions of those behaviors in training data. (arXiv 2025)
• Recursive Introspection: Teaching Language Model Agents How to Self-Improve - Trains language models to inspect unsuccessful prior attempts and recursively improve their responses across subsequent interaction turns. (NeurIPS 2024)
• Self-Recognition in Language Models - Tests whether models can identify their own outputs using model-generated security questions and finds no general self-recognition. (EMNLP 2024)
• Do Large Language Models Know What They Don't Know? - Evaluates model self-knowledge by testing recognition of unanswerable and unknowable questions. (Findings of ACL 2023)
• Language Models (Mostly) Know What They Know - Measures whether models can evaluate their own claims and predict which questions they know how to answer. (arXiv 2022)
• Bounded Recursive Self-Improvement - Studies an implemented goal-directed system that improves its behavior through an explicitly bounded self-modeling loop. (arXiv 2013)

Benchmarks & Evaluations

Benchmarks are separated into direct self-improvement evaluations, frontier-lab evaluation frameworks, and capability proxies. Proxy benchmarks are included only when they measure a bottleneck that directly constrains persistent or recursive self-improvement.

Direct RSI & Self-Improvement Evaluations

• LongWoF-Bench: Evaluating EvoMap Genes for Verifiable Long-Workflow Tasks - Provides 778 machine-verifiable long-workflow tasks and shows that Genes consolidated from verifier-confirmed trajectories outperform Skill packages across seven models, linking the gains to verified experience provenance. (arXiv 2026)
• PostTrainBench: Can LLM Agents Automate LLM Post-Training? - Gives autonomous agents one base model, one H100 GPU, and ten hours to research and execute the strongest post-training strategy they can find. (ICML 2026)
• RSI-Bench - Provides an open-source six-axis framework for measuring self-modification depth, improvement trajectories, operator discovery, adaptation, safety, and goal generation. (community framework 2026)
• RSIBench-Data - Isolates data-centric RSI by asking agents to iteratively improve training-data strategies against checkpoint feedback under a fixed post-training stack. (arXiv 2026)

Frontier Lab Self-Improvement & AI R&D Evaluation Frameworks

• Anthropic Autonomous AI R&D Evaluations - Defines the Responsible Scaling Policy AI R&D-4 capability threshold as fully automating the work of an entry-level, remote-only researcher and evaluates models and safeguards against it. (Anthropic Model Report 2026)
• Google DeepMind Frontier Safety Framework (FSF) ML R&D - Uses dedicated CCLs, TCLs, and evaluation protocols for Machine Learning R&D capabilities that could significantly accelerate or automate AI research and development. (Google DeepMind Blog 2026)
• OpenAI AI Self-Improvement Evaluations - Tracks AI Self-Improvement capability under the Preparedness Framework using Internal Research Debugging, KernelGen 1P, NanoGPT, PostTrainBench Lite, MLE-Bench Revised, and related evaluations aggregated into an RSI Index. (OpenAI System Card 2026)

AI R&D Capability Proxies

• AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks? - Expert-curated benchmark of 36 realistic long-horizon, closed-loop optimization tasks for evaluating frontier agents on automated research and engineering. (arXiv 2026)
• MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI - Benchmark of 140 tasks across 12 ML research domains measuring whether AI systems can invent generalizable and scalable ML methods. (arXiv 2026)
• MLE-bench - Measures end-to-end machine-learning engineering performance across 75 Kaggle competitions and is used to track model self-improvement capability. (ICLR 2025)
• PaperBench - Evaluates agents on replicating state-of-the-art AI research from paper descriptions. (ICML 2025)
• RE-Bench - Compares AI agents with human experts on open-ended machine-learning research-engineering tasks under fixed time budgets. (ICML 2025)
• MLAgentBench - Tests whether language agents can autonomously execute and improve machine-learning experiments from research instructions. (ICML 2024)

Agent Capability Proxies

• ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence - Requires agents to explore unfamiliar interactive environments, infer goals, model environment dynamics, remember, and plan, measuring adaptive interaction and generalization bottlenecks for self-improvement. (arXiv 2026)
• Long-Horizon-Terminal-Bench - Evaluates agents on 46 terminal tasks requiring sustained execution across hundreds of episodes with dense intermediate rewards. (arXiv 2026)
• OSWorld 2.0 - Measures computer-use agents on 108 realistic end-to-end workflows whose median human completion time is about 1.6 hours. (arXiv 2026)
• MCPMark - Stress-tests realistic stateful MCP workflows across SaaS, development, browser, filesystem, and database environments with programmatic grading. (arXiv 2025)
• METR Task-Completion Time Horizon - Estimates the human-equivalent task duration at which an agent succeeds with a given probability on non-trivially parallelizable software tasks. (NeurIPS 2025)
• SWE-Bench Pro - Tests coding agents on 1,865 contamination-resistant enterprise tasks designed to require hours or days of professional software engineering. (arXiv 2025)
• TheAgentCompany - Evaluates agents on 175 cross-application workplace tasks inside a simulated software company with result and checkpoint grading. (NeurIPS 2025)
• SWE-bench - Provides reproducible real-world software issues used to evaluate coding agents and empirical self-modification systems such as DGM. (ICLR 2024)
• SWE-bench Verified - Supplies a human-validated subset that reduces broken or underspecified tasks when measuring iterative coding-agent improvements. (OpenAI benchmark 2024)

Frameworks & Tools

Self-Modifying / Self-Evolving Systems

• AgentFactory - Self-evolving framework that accumulates and reuses executable subagents to improve future task solving.
• Darwin Gödel Machine - Official self-modifying coding-agent implementation with open-ended archive-based evolution.
• Gödel Agent - Official implementation of Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement.
• Hermes Agent - Self-improving personal agent with a built-in learning loop that creates and refines skills from experience.
• HyperAgents - Self-referential agent framework in which an editable meta-agent improves itself and the task agent for computable objectives.
• SEAL - Official code for language models that generate their own adaptation data and update instructions.
• SIA - Self-improving AI framework that iteratively updates an agent harness and, when configured, the target model's weights.

Harness / Memory / Skill Evolution

• ACE - Official implementation of Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models.
• ALMA - Official implementation of Learning to Continually Learn via Meta-learning Agentic Memory Designs.
• Continual Harness - Self-improving agent harness that adapts online from experience while maintaining a continuous environment trajectory.
• EvoAgentX - Self-evolving agent framework for automatically building, evaluating, and optimizing agentic workflows.
• EvolveR - Self-evolving LLM-agent framework that improves through a closed-loop, experience-driven lifecycle.
• Letta Code - Memory-first coding-agent harness whose long-lived agents rewrite context and learn skills from experience.
• Memento-Skills - Self-evolving agent framework that retrieves, evaluates, repairs, and rewrites persistent skills through reflective learning.
• Reef - Continual-learning serving infrastructure that records agent interactions, turns matched feedback into model-weight or harness updates, and publishes accepted updates as versioned artifacts.
• Voyager - Embodied lifelong-learning agent with automatic curriculum, iterative prompting, and a reusable skill library.

Automated Search / AI R&D

• ADAS - Official implementation of a meta-agent that searches over executable agent designs.
• AI Scientist - End-to-end system for generating machine-learning ideas, running experiments, and writing research papers.
• autoresearch - Runs an autonomous loop that edits an LLM training program, trains for five minutes, and keeps only changes that improve validation bits per byte.
• Evolutionary Model Merge - Evolves combinations of open models in parameter and data-flow space.
• FunSearch - Reference implementation of LLM-guided evolutionary program search with executable evaluators.
• MLEvolve - Self-evolving multi-agent framework for end-to-end machine-learning algorithm discovery using progressive search and experience-driven memory.
• OpenEvolve - Open-source evolutionary coding agent inspired by AlphaEvolve-style program optimization.
• POET - Reference implementation for co-evolving environments and their paired agents.

Blog Posts & Discussions

• A Taxonomy of Self-Evolving Agents - Distinguishes Model, Harness, and Artifact evolution, complementing this list's own taxonomy. (2026)
• Hyra: A simple yet effective scaffold for general discovery - Tencent Hunyuan report on Hyra-1.0, a recursively self-improving agent for performance-oriented research and engineering tasks, with produced research artifacts open-sourced in a companion repository. (2026)
• AI4AI at Scale: Building Open-Weight Deep Search Agents - Industry technical report on a bounded, verification-gated AI-for-AI loop where agent teams diagnose failures and apply scoped changes across data, post-training, runtime, and infrastructure, with accepted and rejected attempts banked as auditable experience. (2026)
• First Steps Toward Automated AI Research - Describes an automated research system that closes the full loop from idea proposal through implementation, experiments, validation, and selection of the next experiment, with three reported SOTA results and open-sourced artifacts. (2026)
• Harness Engineering for Self-Improvement - Survey of harness engineering as a near-term path to recursive self-improvement, covering harness design patterns, context and workflow optimization, self-improving harnesses, evolutionary search, and joint optimization with model weights. (2026)
• AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced Algorithms - Google DeepMind explains AlphaEvolve's evaluator-guided evolution loop and its applications to computing and AI training. (2025)
• The Darwin Gödel Machine: AI That Improves Itself by Rewriting Its Own Code - Sakana AI explains DGM's empirical alternative to proof-based Gödel Machine self-modification. (2025)
• When AI Builds Itself - Anthropic analyzes early evidence, possible paths, and governance challenges for AI-driven AI development. (2025)
• The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery - Sakana AI presents its automated research pipeline, results, limitations, and open implementation. (2024)
• Evidence on Recursive Self-Improvement from Current ML - Reviews empirical evidence for and against strong returns from AI-assisted AI research. (2023)
• FunSearch: Making New Discoveries in Mathematical Sciences Using Large Language Models - Google DeepMind describes how evolutionary program search produced verifiable mathematical and algorithmic discoveries. (2023)
• Metalearning Machines Learn to Learn - Jürgen Schmidhuber traces self-referential meta-learning from 1987 through Gödel Machines and modern learned optimizers. (2020)
• Recursive Self-Improvement - The Alignment Forum overview connects self-improving AI to takeoff dynamics, seed AI, and control concerns. (2016)

Talks & Videos

• Escape Velocity: The Inflection Point for Recursive Self Improvement - Louis Kirsch discusses automated AI research and the conditions required for sustained recursive improvement at the ICLR 2026 RSI workshop. (2026)
• ICLR 2026 Workshop on AI with Recursive Self-Improvement - Official video archive for invited talks, contributed work, and panels focused specifically on RSI. (2026)
• Self-Improving Foundation Models Without Human Supervision - Official ICLR 2025 workshop recordings on synthetic data, weak-to-strong learning, and autonomous adaptation. (2025)
• Gödel Machine - Jürgen Schmidhuber gives a concise explanation of self-referential, proof-guided code rewriting and its computability limits. (2015)

Related Awesome Lists

• Awesome AI Agents - Broad directory of autonomous-agent projects and infrastructure that can serve as components or baselines for self-improving systems. (2023)
• Awesome AutoML Papers - Curated literature on automated model selection, architecture search, hyperparameter optimization, and related techniques. (2018)
• Awesome Self-Improving Agents - Focused bibliography of foundation-model agents that update their models, memory, tools, prompts, or workflows. (2024)

Contributing

Contributions are welcome. Please read the contribution guidelines before submitting a pull request.