System 2 Research Links is maintained by OpenThought on GitHub and is a collection of research materials on reasoning and cognition in AI systems. The term "System 2" in the repository title is taken from the slower, more deliberative and logical mode of thinking described by Daniel Kahneman in Thinking, Fast and Slow. The goal of the repository is to bring together related books, papers, blog posts, and other materials, specifically covering agents, cognitive architectures, general problem-solving strategies, and self-improvement.
The repository organizes materials in a directory format, with core categories including: cognitive architectures, papers related to cognitive architectures, LLM reasoning improvement and synthetic data training, reward models (ORM/PRM), reinforcement learning, Monte Carlo tree search, Minecraft agents, large-scale sampling/generation and testing, world models, neuro-symbolic methods, active inference, prompting techniques, and more. Under each category, papers, code repositories, or blog links are listed by time or theme.
The cognitive architectures section includes systems such as SOAR, ACT-R, SPAUN, ART, CLARION, EPIC, LIDA, Sigma, OpenCog, NARS, Icarus, MicroPsi, Thousand Brains Theory & HTM, SPH, Leabra, CogNGen, KIX, ACE, Iterative Updating of Working Memory, and links to their related papers or codebases.
In the LLM reasoning direction, the repository includes many representative works, such as ReAct, SWE-agent, The AI Scientist, OpenDevin, TextGrad, Agentless, AlphaCode, AI Agents That Matter, Sibyl, Trial and Error, Self-Rewarding Language Models, ArchCode, MedAgent-Zero, Cognitive Architectures for Language Agents, ADAS, IDEA, LAW, GenRM, CodeAct, PLANSEARCH, Agent S, and others, and lists in chronological order new papers from 2024 to 2025, such as LLaVA-o1, Multi-agent Architecture Search via Agentic Supernet, START, LADDER, and others.
The reward model section covers QwQ PRM, Process Reinforcement through Implicit Rewards, Implicit PRM, Math-Shepherd, Training Verifiers to Solve Math Word Problems, and others. The reinforcement learning and search section includes GRPO, PPO, REINFORCE++, OREO, DQO, RLEF, R^3, ReFT, ARES, Monte Carlo Tree Diffusion, rStar-Math, Agent Q, LightZero, MuZero, and others.
The world models and neuro-symbolic section includes GameNGen, GAIA-1, the Dreamer series, World Models, AlphaGeometry, DreamCoder, SymbolicAI, VerityMath, and others. The prompting techniques section systematically organizes methods such as Chain-of-Thoughts, Tree-of-Thoughts, Graph-of-Thoughts, Algorithm of Thoughts, Chain-of-Verification, Mixture-of-Agents, Program of Thoughts, Buffer of Thoughts, Chain of Code, Thought of Search, RE2, ECHO, Divergent CoT, Iteration of Thought, Logic-of-Thought, Forest-of-Thought, Chain-of-Associated-Thoughts, and others, along with surveys and guides on prompt engineering.
The repository also includes links to several open reasoning datasets and reproduction projects, such as SYNTHETIC-1, Open Reasoning Data, OpenThoughts-114k, reasoning-gym, open-r1 datasets, Unsloth GRPO, TinyZero, DeepSeek R-1, OpenR, O1-Journey, Open-O1, g1, toyberry, etc.
Usage: browse or clone the repository on GitHub, and find materials by category; if resources need to be supplemented, you may refer to the How to contribute guide in the repository to participate in contributions. This repository is mainly intended for researchers and developers in the directions of AI reasoning, cognitive architectures, and agents. The repository itself is a collection of links and does not provide a unified API or a directly installable and runnable software package; the specific functions, usage steps, and limitations of various methods, models, or tools must be checked in their corresponding original materials.
