mbzuai-oryx/Awesome-LLM-Post-training

What it solves

This repository addresses the complexity of navigating the rapidly evolving field of Large Language Model (LLM) post-training. It provides a structured, curated collection of research and methodologies designed to help developers and researchers understand how to improve LLM reasoning, decision-making, and alignment after the initial training phase.

How it works

The project acts as a central knowledge hub. It organizes academic papers, benchmarks, and technical resources into a clear taxonomy. This includes categories such as reinforcement learning (RL) techniques, process reward models, tree search algorithms, and policy optimization, allowing users to explore specific post-training strategies like fine-tuning or test-time scaling.

Who it’s for

This resource is intended for AI researchers, machine learning engineers, and students who are interested in the technical advancement of LLMs. It is particularly useful for those looking to implement or study advanced reasoning capabilities, such as those found in modern "reasoning models."

Highlights

• A comprehensive taxonomy of post-training methods including fine-tuning, RL, and test-time scaling.
• Curated lists of influential papers on reasoning, decision-making, and multi-agent systems.
• Links to benchmarks and libraries for evaluating and implementing LLM-RL frameworks.
• Regularly updated with the latest research, including a survey paper accepted by IEEE TPAMI.