Long-horizon Reasoning Agent for Olympiad-Level Mathematical Problem Solving

A new framework called Mistake Notebook Learning (MNL) has been introduced to enhance the performance of large language models (LLMs) by utilizing a persistent knowledge base of abstracted error patterns. This approach allows for batch-wise error abstraction, enabling models to learn from multiple failures and retain only effective guidance, achieving performance close to supervised fine-tuning on benchmarks like GSM8K.

The introduction of Saturn, a SAT-based reinforcement learning framework, aims to enhance the reasoning capabilities of large language models (LLMs) by addressing key limitations in existing RL tasks, such as scalability, verifiability, and controllable difficulty. Saturn utilizes Boolean Satisfiability problems to create a structured learning environment for LLMs.