HippoSpark: An On-Demand Experience System for LLM Reasoning

2026-06-29

AI Summary

Existing large language models typically reuse experience only at the task level, which proves insufficient for providing precise, state-dependent guidance to address localized bottlenecks in complex reasoning. This work proposes the first state-level experience reuse framework, which employs a state-aware retrieval mechanism to dynamically extract relevant experiences from distilled historical reasoning trajectories and integrates them via prompt engineering to deliver immediate, actionable guidance at critical decision points. By moving beyond conventional task-level paradigms, the method enables fine-grained, context-sensitive application of prior knowledge. Extensive experiments across mathematical, scientific, and programming benchmarks demonstrate substantial improvements over standard prompting and task-level experience reuse approaches, confirming the efficacy of state-level guidance in enhancing reasoning performance.

Abstract

Distilling historical trajectories into reusable experience to enhance future problem-solving has become a focal point of recent LLM research. However, existing methods predominantly operate at the task level, leveraging general summaries or rules under the assumption that analogous tasks share universal solution patterns. This approach often fails in complex reasoning, which typically falters at local bottlenecks that require precise, state-specific guidance rather than broad heuristics. We introduce HippoSpark, a state-level experience system that performs on-demand retrieval tailored to the immediate needs of the current reasoning state. Across mathematical, scientific, and programming benchmarks, HippoSpark consistently outperforms both standard prompting and task-level experience baselines. Our findings reveal that the most effective experience systems are those that provide actionable guidance at critical bottlenecks rather than serving as generic task-level context. Our code is available at https://github.com/DanlingMeng/HippoSpark.

Problem

• LLM reasoning
• experience distillation
• reasoning bottlenecks
• state-level guidance
• complex reasoning

Innovation

• state-level experience
• on-demand retrieval
• LLM reasoning
• bottleneck guidance
• experience distillation