MILES Framework for Self-Improving LLM Reasoning Memory

MILES implements a modular instruction memory that dynamically expands step-wise during test-time reasoning. It uses a learnable selection policy optimized for final-answer correctness to compose memory from previous sequential problems, bypassing the need for fixed action spaces.