AI 程式設計／演化搜尋
EvoMem Turns Cross-Task Optimization Experience Into Persistent Memory, Accelerating Evolutionary Search by 5.93× on Average
EvoMem extracts provenance-tracked strategy cards from successful program mutations, then retrieves up to three cards to guide the LLM on subsequent tasks. Across nine test categories, it improved the target metric by an average of 6.40%, though results varied widely and the experiments remain limited to small, research-oriented workloads.
LLM-driven evolutionary program search typically discards the exploration process after each run and retains only the best program. As a result, the next similar task must rediscover the same caching, sparsification, or kernel fusion strategies. EvoMem, released on August 11, turns these intermediate results into a memory layer that persists across runs.
After an evolutionary run completes, the system compares child programs with their parents, sibling variants, and subsequent elite lineages to extract evidence-backed “idea cards.” Each card contains a strategy description, source task, program lineage, applicable contexts, and usage statistics. Semantically similar abstract strategies can be merged, while concrete programs remain separate. Before generating a new mutation, the system first uses an LLM to identify potentially relevant source tasks, then combines textual and embedding similarity to retrieve up to three cards for inclusion in the mutation prompt. Program selection, correctness validation, and fitness calculation remain unchanged, so memory serves only as a search bias and does not bypass the evaluator.
The study used Gemini 3 Flash to power GigaEvo and evaluated it on geometric packing, multi-hop question answering, AlgoTune scientific programs, and KernelBench GPU kernels. To reduce leakage, each test excluded memories generated by the same benchmark. Across nine result categories, the target metric improved by an average of 6.40%, while the number of candidates required to match the memory-free baseline’s best score fell to roughly one-sixth, representing a 5.93× average speedup. KernelBench showed an average gain of 16.89%, but individual results ranged from 0% to 65.93%, and some runs were slower. Child programs actually adopted strategies from relevance-based retrieval only 4.05% of the time, although that was higher than the 0.25% adoption rate for randomly selected cards.
The key engineering contribution is not any single average result, but the elevation of evolutionary trajectories into auditable, retrievable assets. Further work is still needed to validate the approach on large repositories, cross-language modifications, and long-term memory contamination. For now, the memory store is small, the ablation studies are incomplete, and high-temperature sampling produces substantial rerun variance, so the 5.93× figure should not be treated as a stable deployment speedup.