Adaptive Solving with Episodic Memory and Serial Multi-Agent Reasoning

Abstract

Large language models (LLMs) have demonstrated strong capabilities on complex reasoning tasks. Existing studies mainly enhance the complex reasoning ability of LLMs through chain-of-thought reasoning, tool augmentation, experience knowledge reuse, and multi-agent collaboration. However, existing methods still suffer from clear limitations in adaptive scheduling for tasks of different difficulty levels and in the organization of collaborative reasoning. Most frameworks apply a uniform heavyweight reasoning pipeline to all problems, lack a dynamic path selection mechanism based on task complexity, and often organize multi-agent collaboration as parallel answer generation followed by competition or voting, which is not well suited to complex reasoning tasks with strong step dependencies. To address these issues, we propose AdaSolve, an adaptive episodic-knowledge-enhanced multi-agent reasoning framework. AdaSolve builds an adaptive routing mechanism based on a complexity classifier: simple problems are solved through direct reasoning, whereas complex problems activate an enhanced path. For complex problems, the framework further introduces experience knowledge retrieval and model self-recall mechanisms to provide historical reasoning experience as references for the current problem. Then, a planning agent decomposes the problem into ordered subtasks, and multiple agents cooperate in a serial manner to perform subtask reasoning, result synthesis, and answer verification. In this way, AdaSolve improves the stability and interpretability of complex problem solving. Experiments on GSM8K, MATH-500, and AIME show that AdaSolve achieves substantial and consistent performance gains over the base model across difficulty levels, and remains competitive with strong experience-enhanced baselines. Ablation studies further verify the effectiveness of complexity classification and serial multi-agent collaboration. Token-consumption analysis also shows that the method improves reasoning performance while maintaining system efficiency, demonstrating the advantage of combining adaptive episodic knowledge enhancement with serial multi-agent reasoning for complex reasoning tasks.