This repository serves as a curated collection of research papers focused on self-correcting Large Language Models (LLMs) that utilize automated feedback mechanisms. It is intended for researchers and practitioners in the field of Natural Language Processing and AI safety who are investigating methods to improve LLM accuracy, reliability, and adherence to instructions. The primary benefit is a structured overview of the rapidly evolving landscape of LLM self-correction techniques.

How It Works

The collection categorizes self-correction strategies into three main phases: Training-Time, Generation-Time, and Post-hoc. Training-Time methods focus on improving the model during its learning phase, often through reinforcement learning from human feedback (RLHF) or self-training. Generation-Time strategies enhance output quality during inference, employing techniques like re-ranking or feedback-guided generation. Post-hoc correction methods refine outputs after generation, utilizing approaches such as self-refinement, external feedback, or model debates. This categorization provides a systematic framework for understanding and comparing different self-correction paradigms.

Quick Start & Requirements

This repository is a collection of research papers and does not involve code execution. No installation or specific software requirements are necessary to access the information. Links to the papers are provided within the README.

Highlighted Details

• Comprehensive categorization of self-correction strategies across training, generation, and post-hoc phases.
• Includes a survey paper by the contributors, offering a structured overview of the field.
• Extensive list of seminal and recent research papers with direct links to their sources.
• Covers a wide range of techniques including RLHF, self-training, re-ranking, self-refine, and model debate.

Maintenance & Community

The repository is maintained by contributors, with Liangming Pan and Xinyuan Lu noted. The README encourages community contributions to fill any gaps and provides contact information for issues.

Licensing & Compatibility

The licensing is not explicitly stated in the README. As it is a collection of research papers, users should refer to the individual licenses of the linked papers for usage and distribution rights. Compatibility for commercial use or closed-source linking would depend on the licenses of the cited works.

Limitations & Caveats

This repository is a curated list of papers and does not provide code implementations or benchmarks. The completeness of the collection is subject to ongoing community contributions, and some important works may be missing.