RL-based Reasoning

• SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution [paper]

  • FAIR at Meta, UIUC, GenAI at Meta, CMU

• Memento: Fine-tuning LLM Agents without Fine-tuning LLMs

  • Time: 25 Aug 2025

    1. Use external case memory to save past experience, learn to read and write cases while keeping LLM fixed

    2. soft Q-learning framework to select the most relevant cases for each situation through experience

Note:

See The State of Reinforcement Learning for LLM Reasoning for a nice overview