This repository curates papers, libraries, and resources focused on constrained decoding for Large Language Models (LLMs). It aims to help researchers and developers build more reliable, controllable, and efficient LLM generation systems by providing a comprehensive overview of techniques and tools for enforcing specific output formats and structures.

How It Works

The project addresses the challenge of LLM output control by cataloging methods that guide the generation process. This includes libraries implementing techniques like Context-Free Grammars (CFGs), regular expressions, and JSON schema adherence, as well as papers exploring novel decoding algorithms, benchmarking frameworks, and theoretical underpinnings of constrained generation. The core advantage lies in providing a centralized, up-to-date resource for a rapidly evolving field, enabling users to quickly find relevant tools and research for their specific LLM control needs.

Quick Start & Requirements

This is a curated list, not a runnable project. Users should refer to individual library repositories for installation and usage. Requirements vary by library but may include specific Python versions, Hugging Face Transformers, or CUDA for GPU acceleration. Links to relevant libraries, papers, and benchmarks are provided within the README.

Highlighted Details

• Features a wide array of libraries supporting various constraints (CFG, Regex, JSON Schema) and LLM frameworks (Transformers, vLLM, llama.cpp).
• Includes a comprehensive, chronologically ordered list of research papers on constrained decoding, categorized by application (e.g., code generation, diffusion models) and evaluation.
• Highlights benchmarking efforts and datasets specifically designed for evaluating constrained decoding performance.

Maintenance & Community

The list is actively maintained, with a clear invitation for contributions via issues or pull requests. The README acknowledges the significant contributions from the Outlines team and links to a related "awesome-llm-json" list.

Licensing & Compatibility

The repository itself is a list and does not have a specific license. Individual libraries mentioned will have their own licenses, which users must consult for compatibility and usage restrictions, especially for commercial applications.

Limitations & Caveats

The list is explicitly stated as not exhaustive and subject to change. The features mentioned for libraries are not exhaustive, and users are strongly encouraged to consult the respective repositories for complete details. The primary limitation is that this is a resource list, not a unified tool, requiring users to integrate individual libraries themselves.