Abstract
Evolutionary search processes that use large language models may exhibit instability and inefficiency, which can lead to stagnation, context pollution, and suboptimal collaboration in parallel runs. A progress-aware framework can be used to govern these searches. The framework may integrate components such as hierarchical context management for organizing the search into a structured idea pool, momentum-based backtracking for adaptively escaping local minima by tracking a progress metric, and self-adaptive collaborative evolution for dynamically managing knowledge transfer between parallel agents. Functioning as a data-driven control system, the framework can make adaptive decisions regarding, for example, pruning, backtracking, and collaboration. This approach may improve the stability and efficiency of the search process and can support self-improvement in complex optimization problems.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yan, Minghao; He, Zhankui; Wang, Chi; Chi, Ed; Coleman, Ben; Chen, Ziqi; Xie, Zhouhang; Chen, Shuo; Sachdeva, Noveen; Wang, Beidou; Ye, Isabella; Cheng, Derek; Wang, Weili; Kang, Wang-Cheng; Pereira, Fernando; and Peng, Bo, "Progress-Aware Framework for Governing LLM-Driven Evolutionary Search", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10043