Resources on recursive language models, recursive inference, recursive reasoning architectures, and self-calling AI systems.
Recursive systems matter because they allow models and agents to decompose context, call sub-processes, refine intermediate outputs, verify assumptions, and organise longer-horizon computation.
- Taxonomy
- Core Papers
- Recursive Language Models
- Recursive Reasoning Architectures
- Inference-Time Recursion
- Recursive Agents and Tool Environments
- Recursive Evaluation and Verification
- Recursive Planning and Search
- RL and Self-Improving Systems
- Simulation Recursion and Social Simulation
- Benchmarks and Evaluation Tasks
- Open-Source Implementations
- Maintainer Notes
| Pattern | Description | Example resources |
|---|---|---|
| Self-recursive inference | Model calls itself on bounded sub-contexts. | RLM, λ-RLM |
| Tree recursion | Expands and evaluates branches. | ToT, LATS, MCTS methods |
| Graph recursion | Reasoning states form reusable graph structures. | GoT |
| Iterative refinement | Output-feedback-revision loops. | Self-Refine, Reflexion |
| Recursive architectures | Recurrent or weight-shared reasoning modules. | HRM, TRM, RSM |
| Recursive retrieval | Hierarchical abstraction over context. | RAPTOR |
| Recursive self-improvement | System modifies policy, code, prompts, or search process. | STOP, Gödel Agent, Polaris |
| Simulation recursion | Agents recursively plan, reflect, update beliefs, and interact. | Generative Agents, AgentSociety, SOTOPIA |
- Recursive Language Models - Introduces RLMs as an inference strategy where an LLM uses an external environment to inspect, decompose, and recursively call itself over long prompts. (2025)
- Reasoning Language Models: A Blueprint - Provides a modular blueprint for reasoning language models, covering chains, trees, graphs, nested reasoning, search, reinforcement learning, supervision, and test-time compute. (2025)
- Recursive Models for Long-Horizon Reasoning - Formalises recursive model self-invocation for long-horizon reasoning under bounded context and evaluates recursive reasoning on Boolean satisfiability. (2026)
- Recursion of Thought: A Divide-and-Conquer Approach to Multi-Context Reasoning with Language Models - Introduces recursive multi-context reasoning where model outputs can trigger subproblem contexts beyond a single sequence window. (2023)
- Tree of Thoughts: Deliberate Problem Solving with Large Language Models - Frames reasoning as search over intermediate "thought" states with generation, self-evaluation, and selection. (2023)
- Self-Refine: Iterative Refinement with Self-Feedback - Uses the same LLM as generator, feedback provider, and refiner in an inference-time improvement loop. (2023)
- Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation - Studies a scaffolding programme that uses a language model to modify code that can call itself for further improvement. (2023)
See the core papers above for recursive self-invocation and multi-context recursion.
- Recursive Language Models - Project blog post explaining RLMs, REPL environments, recursive subcalls, and early long-context experiments. (2025)
- The Y-Combinator for LLMs: Solving Long-Context Rot with Lambda-Calculus - Introduces λ-RLM, a typed functional approach to recursive language-model inference with explicit control flow, termination guarantees, and cost bounds. (2026)
- RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval - Builds a recursive tree of clustered summaries so retrieval can operate over both local chunks and higher-level abstractions. (2024)
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions - Interleaves retrieval and chain-of-thought steps so each retrieval depends on previously derived reasoning in a mutually recursive loop. (2022)
- Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context - Studies when recursion actually helps RLM-style long-context inference and shows uncertainty-guided self-reflective program search can match or exceed recursive self-calls. (2026)
- Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE - Provides a tuning-free RoPE extension for preserving short-context behaviour while scaling accumulated traces in recursive and repository-scale workflows. (2026)
See Tree of Thoughts in the core papers for the canonical tree-search framing.
- Graph of Thoughts: Solving Elaborate Problems with Large Language Models - Represents LLM-generated units of reasoning as graph nodes that can be transformed, scored, aggregated, and looped. (2023)
- Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models - Prompts models with algorithmic reasoning trajectories to encourage structured exploration with recurrent dynamics. (2023)
- Cumulative Reasoning with Large Language Models - Coordinates proposer, verifier, reporter, and halter roles in an iterative reasoning process. (2023)
- Everything of Thoughts: Defying the Law of Penrose Triangle for Thought Generation - Combines thought generation with reinforcement learning and Monte Carlo Tree Search for multi-solution problem solving. (2023)
- THREAD: Thinking Deeper with Recursive Spawning - Frames generation as threads that recursively spawn child threads for task solving and question answering. (2025)
- Hierarchical Reasoning Model - Proposes a recurrent reasoning architecture with high-level and low-level modules operating at different timescales for sequential reasoning tasks. (2025)
- Less is More: Recursive Reasoning with Tiny Networks - Proposes Tiny Recursive Model, a small-network architecture that recursively refines latent state and answers on puzzle tasks. (2025)
- Form Follows Function: Recursive Stem Model - Introduces a recursive reasoning model trained as a stable depth-agnostic transition operator, with test-time recursion and convergence behaviour as a reliability signal. (2026)
- Least-to-Most Prompting Enables Complex Reasoning in Large Language Models - Decomposes problems into progressively harder subproblems whose solutions feed later steps, enabling easy-to-hard generalisation. (2022)
- Decomposed Prompting: A Modular Approach for Solving Complex Tasks - Routes sub-tasks to dedicated prompt handlers that can recursively decompose a task into smaller instances of itself. (2022)
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters - Studies how search, revision, and verifier-guided strategies allocate extra inference-time compute. (2024)
- Large Language Monkeys: Scaling Inference Compute with Repeated Sampling - Analyses repeated sampling as a test-time compute strategy whose value depends on verification and selection. (2024)
- Test-time Recursive Thinking: Self-Improvement without External Feedback - Proposes an inference-time recursive thinking framework that uses rollout strategies, accumulated knowledge, and self-generated verification signals for self-improvement. (2026)
- Self-Evaluation Guided Beam Search for Reasoning - Integrates self-evaluation into stochastic beam search for multi-step reasoning. (2023)
- What, How, Where, and How Well? A Survey on Test-Time Scaling in Large Language Models - Surveys test-time scaling methods, including search, verifier-guided reasoning, and adaptive deliberation. (2025)
- A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond - Surveys inefficient reasoning patterns in large reasoning models, including redundant traces, over-analysis, and methods for improving reasoning efficiency. (2025)
- Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models - Iteratively aggregates subsets of a population of reasoning chains so each round bootstraps improved solutions from partially correct ones. (2025)
- ReAct: Synergizing Reasoning and Acting in Language Models - Interleaves reasoning traces, actions, and observations in a loop for tasks that require external interaction. (2022)
- Reflexion: Language Agents with Verbal Reinforcement Learning - Converts task feedback into verbal reflection stored in memory for later attempts. (2023)
- Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models - Combines reflection, acting, and planning through Monte Carlo Tree Search over agent trajectories. (2024)
- Voyager: An Open-Ended Embodied Agent with Large Language Models - Uses environment feedback, self-verification, and an expanding skill library for lifelong embodied exploration. (2023)
- Toolformer: Language Models Can Teach Themselves to Use Tools - Trains models to decide when and how to call tools using self-supervised API-call annotations. (2023)
- Tree Search for Language Model Agents - Applies inference-time best-first tree search over agent actions for exploration and multi-step planning in interactive web environments. (2024)
- Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations - Builds a recursive tree of explanations and solves consistency as a satisfiability problem. (2022)
- Let's Verify Step by Step - Compares process supervision with outcome supervision and releases step-level feedback data for mathematical reasoning. (2023)
- V-STaR: Training Verifiers for Self-Taught Reasoners - Iteratively trains reasoners and verifiers from self-generated reasoning data. (2024)
- Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement - Examines self-bias risks in self-refinement loops and the role of external feedback. (2024)
- CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing - Validates and progressively revises model outputs through an iterative loop of tool-based critique and correction. (2023)
- Generative Verifiers: Reward Modeling as Next-Token Prediction - Trains verifiers as generators so that verification can itself use chain-of-thought reasoning and majority voting. (2024)
- Agent-as-a-Judge: Evaluate Agents with Agents - Uses agentic systems to evaluate other agents with intermediate feedback across the whole task-solving trajectory. (2024)
- SciReasoner - Uses addressable protein, molecular, and crystal-structure tokens as inspectable evidence units for transparent structure-property reasoning. (2026)
- Reasoning with Language Model is Planning with World Model - Uses an LLM as both agent and world model inside Monte Carlo Tree Search for reasoning and planning. (2023)
- LiteSearch: Efficacious Tree Search for LLM - Proposes a tree-search method that allocates expansion based on history and value guidance. (2024)
- Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning - Uses MCTS to collect step-level preference data for iterative reasoning improvement. (2024)
- Mutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers - Couples Monte Carlo Tree Search over human-like reasoning actions with a second model that verifies trajectories in a self-play generation-discrimination loop. (2024)
- rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking - Combines MCTS deep thinking with a process reward model and iterative self-evolution rounds that improve both policy and reward models. (2025)
See STOP in the core papers for recursively self-improving code-generation scaffolds.
- Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing - Combines generated candidate reasoning paths, search, and critique for self-improvement without extra annotations. (2024)
- ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement - Uses self-synthesised reasoning paths as training data to improve reasoning generalisation. (2024)
- Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement - Proposes a self-referential agent framework for recursively modifying its own behaviour. (2024)
- Polaris: A Gödel Agent Framework for Small Language Models through Experience-Abstracted Policy Repair - Proposes an auditable recursive self-improvement loop where an agent analyses failures, abstracts experience, and applies minimal policy repairs. (2026)
- LADDER: Self-Improving LLMs through Recursive Problem Decomposition - Uses recursive generation of easier problem variants to create a difficulty gradient for self-guided learning and test-time reinforcement learning. (2025)
- A Simple Framework for Intrinsic Reward-Shaping for RL using LLM Feedback - Iteratively generates and refines intrinsic reward functions with LLM feedback for reinforcement learning environments. (2025)
- AlphaEvolve: A coding agent for scientific and algorithmic discovery - Describes an evolutionary coding agent that uses LLMs, automated evaluators, and iterative programme improvement. (2025)
- Self-Rewarding Language Models - Uses the model as its own judge in iterative DPO rounds that improve both instruction following and the quality of self-generated rewards. (2024)
- Recursive Introspection: Teaching Language Model Agents How to Self-Improve - Fine-tunes models to detect and correct their own previous mistakes over successive turns of introspection. (2024)
- Automated Design of Agentic Systems - Introduces Meta Agent Search, where a meta agent iteratively programs ever better agents in code against a growing archive of discoveries. (2024)
- Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents - Evolves an archive of coding agents that iteratively modify their own code, validating each self-modification on coding benchmarks. (2025)
- Generative Agents: Interactive Simulacra of Human Behavior - Introduces LLM-based agents with memory, reflection, and planning for simulating believable individual and emergent social behaviour. (2023)
- AgentSociety - Introduces a large-scale simulator for LLM-driven social agents, supporting population-level interactions and computational social experiments. (2025)
- SocioVerse - Proposes an LLM-agent-driven world model for large-scale social simulation, with alignment components for users, environments, interactions, and behavioural patterns. (2025)
- SOTOPIA-S4: a user-friendly system for flexible, customizable, and large-scale social simulation - Presents a scalable system for customisable multi-turn and multi-party LLM-based social simulation with APIs and a web interface. (2025)
- GASim: A Graph-Accelerated Hybrid Framework for Social Simulation - Proposes a graph-accelerated hybrid framework for scalable social simulation with LLM-driven core agents, graph memory, and graph message passing. (2026)
- Designing Reliable Experiments with Generative Agent-Based Modeling: A Comprehensive Guide Using Concordia by Google DeepMind - Provides a practical guide to designing, running, and validating generative agent-based modelling experiments using Concordia. (2024)
- PRM800K - OpenAI's process-supervision dataset of step-level feedback labels for mathematical reasoning. (2023)
- SWE-bench - A software-engineering benchmark frequently used to study agent loops, retries, verification, and inference-time scaling. (2023)
- WebShop - A simulated e-commerce environment used by ReAct and LATS for action-observation agent evaluation. (2022)
- BrowseComp-Plus - A fixed-corpus benchmark for evaluating deep-research agents, retrieval, citation accuracy, and context-engineering choices under controlled conditions. (2025)
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents - Provides an open-ended environment and evaluation framework for social intelligence in language agents across collaborative, competitive, and goal-driven interactions. (2023)
- LIFELONG SOTOPIA: Evaluating Social Intelligence of Language Agents Over Lifelong Social Interactions - Evaluates language agents across multi-episode social interactions, measuring social intelligence over longer interaction histories. (2025)
- princeton-nlp/tree-of-thought-llm - Official implementation of Tree of Thoughts with tasks, prompts, and trajectories. (2023)
- spcl/graph-of-thoughts - Official Graph of Thoughts implementation for graph-structured reasoning operations. (2023)
- lapisrocks/LanguageAgentTreeSearch - Official LATS repository with reasoning, acting, and planning experiments. (2024)
- noahshinn/reflexion - Official Reflexion implementation for verbal reinforcement learning in language agents. (2023)
- parthsarthi03/raptor - Official RAPTOR implementation for recursive tree-organised retrieval. (2024)
- maitrix-org/llm-reasoners - Library for advanced reasoning algorithms, including planning-style and tree-search methods. (2023)
- alexzhang13/rlm - Official RLM inference library with REPL and sandbox environments for recursive language-model calls. (2026)
- alexzhang13/rlm-minimal - Minimal reference implementation of RLMs with REPL environments for experimentation. (2026)
- lambda-calculus-LLM/lambda-RLM - Official implementation of λ-RLM for typed recursive language-model inference. (2026)
- Jasmine0201/GASim - Official implementation of GASim for graph-accelerated hybrid social simulation. (2026)
Thrilled to have you here.
Whether it's a quick typo fix, a fresh resource,
a doc polish, or a sweeping overhaul — every contribution helps this list grow.
Jump in and join the community — PRs of every size are welcome.
Read the contributing guide good first issues
This list is maintained with an agent-assisted curation pipeline that ships with the
repository: paper-scout discovers candidate resources, curate-entry vets and formats
them, entry-reviewer re-verifies every change, and taxonomy-auditor keeps sections and
coverage healthy. Every change is gated by formatting checks, Markdown lint, awesome-lint,
and live link validation. See AGENTS.md for how the pipeline works.
This work is released under CC0 — to the extent possible under law, the maintainers have waived all copyright and related or neighbouring rights.