Reasoning models boost math problem solving with external memory
Reasoning with Context Management and External Memory
Computation and Language
Summary
Solving hard problems can be tricky because computers often forget useful details when their memory is too small. This paper explains how adding an extra memory that works alongside the main reasoning model can help keep important information safe without slowing the system down. The authors show their approach speeds up calculations for problems that need many repeated steps and helps solve more complex math puzzles better. They tested this on several small to medium-sized models and also used the popular GPT-2 model.
What this means in practice
- •For machine learning engineers: Implement external memory to speed up complex problem-solving tasks in models without increasing computational cost.
Authors
Man-Ling Sung, William Hartmann, Jiangying Zhou
Abstract
Reasoning models maintain information about the status of their reasoning process through in-context memory. Complex problems require larger context windows and memory requirements. In this scenario, context management without caching risks discarding useful information for reasoning. We propose RCMEM, an approach that extends context management with an external memory to maintain additional information without burdening the reasoning model with additional context. RCMEM significantly reduces computation time on problems that require repeated computation of intermediate steps, while achieving better performance on complex problems. We demonstrate our approach on a few mathematics problems using model sizes ranging from 0.4 to 25 million parameters, and GPT-2.