arXiv:2608.12522v1 Announce Type: new Abstract: LLM-based program evolution systems such as FunSearch and AlphaEvolve have shown strong ability to discover novel algorithms, but typically optimize each task in isolation
Sign in with a social account
论文指出的关键痛点是：现有 LLM 程序进化系统在优化每个任务时"从零开始"，即便两个任务在结构或求解策略上高度相似，系统也不会自动利用先前任务的经验。这种孤立优化的方式意味着计算资源的重复消耗，也限制了进化效率的进一步提升。ε-MemEvo 的命名暗示其方法围绕"记忆"（Memory）与"进化"（Evolution）展开——通过维护一个可动态更新的记忆库，在不同任务的程序进化过程中识别并迁移有价值的模式。不过，论文摘要仅提供了问题定位与方向性描述，关于记忆库的具体构建方式、迁移触发条件以及 ε 参数的确切含义，原文未在摘要中展开。
将本条论文放在同期 arXiv 预印本中观察，可以看到一个共同趋势：研究者正在从"让 LLM 在单一任务上表现更好"转向"让 LLM 在不同场景间更灵活地适配"。例如同批发布的《Learning to Adapt Cross-Domain Preferences via Meta-LoRA for LLM Personalization》同样处理跨域迁移问题，只不过聚焦于个性化偏好而非程序进化；《What Drives LLM Self-Reflection?》则通过受控消融实验追问自我反思机制中真正起作用的组件。这些工作共同反映出，LLM 研究正从能力展示阶段进入机制理解与效率优化阶段。
需要明确的是，本文目前以 arXiv 预印本形式发布，标注为 v1 版本，尚未经过同行评议。摘要中未披露实验所用的任务集规模、基线对比方法、评测指标或消融实验结果，因此无法从现有信息判断 ε-MemEvo 相比 FunSearch 或 AlphaEvolve 的实际增益幅度。论文标题中的"Adaptive"（自适应）与"Cross-Task Memory Transfer"（跨任务记忆迁移）作为方法主张是清晰的，但支撑这些主张的实证细节有待全文发布后才能评估。读者在引用或比较时应将此视为初步研究结果。
如果 ε-MemEvo 的方法得到验证，其价值在于降低程序进化系统在多任务场景下的计算开销，并可能加速新任务上的算法发现速度。对于将 FunSearch 类系统用于实际算法搜索的研究者而言，跨任务记忆迁移意味着不必为每个新问题重新投入完整的进化计算资源。但这一潜力的兑现程度，取决于记忆迁移的准确性与泛化边界——迁移不当的经验反而可能引导进化走向次优解。这些问题有待论文全文与后续实验给出答案。
arXiv:2608.12522v1 Announce Type: new Abstract: LLM-based program evolution systems such as FunSearch and AlphaEvolve have shown strong ability to discover novel algorithms, but typically optimize each task in isolation, discarding search experience after completion. We introduce -MemEvo, a framework for cross-task knowledge transfer in LLM program evolution. -MemEvo stores prior experience as task-agnostic tactic memories: compact natural-language summaries of successful algorithmic strategies rather than raw code, enabling transfer across tasks with different APIs and evaluators. To avoid negative transfer from semantically mismatched memories, -MemEvo uses an adaptive injection gate that decides whether retrieved memories should be injected, and at what intensity. We evaluate -MemEvo on 8 diverse optimization benchmarks spanning mathematical optimization and systems engineering, using a content-level Leave-One-Out protocol that excludes target-task memory entries. On the primary GPT-5 backbone, -MemEvo improves AUCC over AdaEvolve on all 8 tasks, with a mean relative gain of +8.7%, and improves early-stage convergence by +9.4% on average. Ablations show that naive memory injection can fail catastrophically, while adaptive gating remains safe across all five ablation tasks. The data-updated posterior is interpretable in observed states: it favors skip during improving search and shifts from skip to hint across early and late plateaus. These gains incur less than 1% computational overhead.