MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning

Short summary

MILES is a framework for self-improving LLM reasoning that dynamically expands step-wise memory with learnable selection heads under realistic test-time constraints. It maintains modular memory units of sub-goal embeddings paired with sub-instructions, using a coarse-to-fine retrieval mechanism that trains selection heads from confident samples and applies them to guide reasoning on uncertain ones. Experiments show it matches or outperforms prior methods with superior accuracy-efficiency tradeoffs.

• •Modular memory with learnable selection heads for test-time LLM reasoning improvement
• •Coarse-to-fine retrieval: expand memory from confident samples, guide uncertain ones
• •Matches or outperforms prior methods with better accuracy-efficiency tradeoffs

Generated with AI, which can make mistakes.