Authors:

Major: Computer Science

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Asymmetric Memory Evolution for Multi-Agent Algorithm and Code Design The use of Large Language Models (LLMs) for automated algorithm design and code generation is rapidly shifting from single-agent setups to more complex multi-agent collaborations. While these multi-agent systems offer significant efficiency gains, they currently face a major limitation known as "Imitation Collapse." In standard frameworks, agents share their successful findings freely. This leads to a feedback loop where agents stop exploring unique paths and instead mimic the high-performing peers, causing the system to homogenize and converge prematurely on suboptimal solutions. To address this critical issue, we propose Asymmetric Memory Evolution (AMEvo), a novel framework that regulates how agents share information. Instead of full transparency, AMEvo introduces an asymmetric update mechanism that balances the need for cooperation with the need for individual diversity. The system operates on two core principles: Failure Socialization and Success Privatization. Through Failure Socialization, agents broadcast their errors to the group, allowing the collective to "prune" the search space and avoid repeating known mistakes. Conversely, Success Privatization allows agents to keep their winning strategies to themselves. This prevents the group from rushing to copy a single solution, thereby safeguarding heuristic diversity and encouraging broader exploration. We validate AMEvo across a variety of challenges, including classic combinatorial optimization and complex scientific discovery in both geometric and discrete spaces. Our results demonstrate that by strategically controlling information flow, AMEvo effectively prevents redundancy and significantly boosts the system's collective search capabilities. Poster Session 5 3:30 PM-4:30 PM CT Room C Poster #12: Walker Johnson Majors: Applied Mathematics, Physics

Source:

Texas A&M University / 2026

Topics:

No topics listed

Co-authors:

Major: Computer Science