Title: Dual-Cluster Memory Agent: Tackling Multi-Paradigm Ambiguity in Optimization Tasks

Abstract:

Large Language Models (LLMs) frequently encounter difficulties when addressing structural ambiguity in optimization problems. Such issues arise when a single problem allows for multiple modeling paradigms that are related yet contradictory, thereby obstructing the generation of effective solutions. To overcome this challenge, we introduce the Dual-Cluster Memory Agent (DCM-Agent), a method designed to boost performance by utilizing historical solutions without requiring additional training.

The core of this approach is Dual-Cluster Memory Construction. The agent categorizes past solutions into distinct modeling and coding clusters. It then condenses the information within each cluster into three specific, structured formats: Approach, Checklist, and Pitfall. This condensation process yields guidance knowledge that is broadly generalizable.

Additionally, the agent employs Memory-augmented Inference to dynamically steer solution pathways. This mechanism enables the detection and correction of errors while allowing for adaptive switching between reasoning paths, guided by the structured knowledge base. Evaluations across seven optimization benchmarks indicate that DCM-Agent delivers an average performance gain ranging from 11% to 21%. Furthermore, our analysis identifies a "knowledge inheritance" effect: memory banks generated by larger models can effectively direct smaller models to achieve higher performance levels, underscoring the framework’s scalability and efficiency.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC