ε-MemEvo: Cross-Task Memory Transfer for LLM Program Evolution
The introduction of a new framework named ε-MemEvo aims to facilitate the transfer of knowledge across various tasks in LLM-based program evolution systems. Unlike previous systems like FunSearch and AlphaEvolve, which treat each task separately and disregard previous search experiences, ε-MemEvo retains past experiences as task-agnostic tactic memories—concise natural-language descriptions of effective algorithmic methods—instead of raw code. This enables task transfer involving different APIs and evaluators. To prevent negative transfer caused by mismatched memories, ε-MemEvo incorporates an adaptive injection gate that determines the appropriateness and intensity of memory injection. The framework was tested on eight varied optimization benchmarks, including mathematical optimization and systems engineering, utilizing a content-level Leave-One-Out protocol. The full paper can be found on arXiv with the identifier 2608.12522.
Key facts
- ε-MemEvo is a framework for cross-task knowledge transfer in LLM program evolution.
- It stores prior experience as task-agnostic tactic memories: natural-language summaries of successful strategies.
- An adaptive injection gate controls memory injection to prevent negative transfer.
- Evaluated on 8 optimization benchmarks in mathematical optimization and systems engineering.
- Uses a content-level Leave-One-Out protocol for evaluation.
- Paper available on arXiv:2608.12522.
- Contrasts with FunSearch and AlphaEvolve, which optimize tasks in isolation.
- Memory format enables transfer across different APIs and evaluators.
Entities
Institutions
- arXiv