This repository serves as a curated collection of papers, code, and resources for "Latent Reasoning," stemming from the survey "A Survey on Latent Reasoning." It targets researchers and practitioners seeking to understand and advance models that perform reasoning implicitly within a latent space, rather than through explicit, step-by-step methods. The project offers a structured overview of the latest research, facilitating quicker adoption and development in this specialized AI domain.

How It Works

The project organizes research into a detailed taxonomy of latent reasoning approaches. This includes categorizing works by their core mechanisms, such as activation-based recurrence, architectural recurrence, training-induced recurrence, hidden-state methods, optimization-based state evolution, and mechanistic interpretability. Each entry links to the original paper and, where available, its associated code, providing a comprehensive map of the field's methodologies and advancements.

Quick Start & Requirements

This repository is a research compilation, not a runnable software project. It does not offer a unified installation or execution command. Users are directed to the individual papers and their respective codebases for practical implementation and experimentation.

Highlighted Details

• Features a comprehensive taxonomy covering diverse latent reasoning paradigms, from recurrent methods to diffusion models and mechanistic interpretability.
• Aggregates over 100 research papers with direct links to publications and code repositories, facilitating access to cutting-edge work.
• Actively maintained, with recent updates in July 2025 reflecting ongoing research in the field.
• Includes links to arXiv, Hugging Face papers, and various conference proceedings.

Maintenance & Community

The repository actively encourages community contributions, feedback, and suggestions for missed papers or updates via email or pull requests. A list of contributors is acknowledged.

Licensing & Compatibility

No specific license is declared for the repository itself. Users must consult the licenses of individual linked papers and code projects for usage restrictions and compatibility.

Limitations & Caveats

As a curated bibliography, this repository does not provide a unified framework or tool for latent reasoning. Its utility is as a reference and starting point for exploring individual research projects, rather than a deployable solution.