Meta FAIR 推出 Self-Improving Memory,Continual Learning Bench 得分 56.0
Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Meta FAIR 让 Kimi-K2.6 自己写记忆提示词再蒸馏进权重,56.0 分比人手写的 44.6 高出 11.4 分。做 agent 记忆或持续学习的人可以看看这条自动化提示词优化的路子,思路比分数本身更值得借鉴。
Meta FAIR 发布 Self-Improving Memory 方法,让 Kimi-K2.6 模型自动编写自己的 LOAD 和 OFFLOAD 记忆提示词,再通过蒸馏把这些提示词固化进模型权重。在 Continual Learning Bench 上,优化后的提示词拿到 56.0 分,比最好的人工手写 harness 的 44.6 分高出 11.4 分。
技术路线分两步:模型先针对记忆的加载与卸载两个环节生成并迭代优化提示词,找到比人工设计更有效的记忆管理策略;随后用蒸馏把提示词学到的行为写入权重,使模型在后续任务中无需重新调用提示词也能保留这些能力。
这套方法把提示词工程从人工环节变成模型自我改进的环节,对比的是手写 harness,改进幅度来自自动化搜索而非更大参数量。做持续学习、记忆机制或 agent harness 研究的开发者可以直接参考这条蒸馏路径。
Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Meta FAIR’s Self-Improving Memory lets Kimi-K2.6 write its own LOAD and OFFLOAD memory prompts, then distills them into its weights. Optimized prompts score 56.0 versus 44.6 for the best hand-written harness on Continual Learning Bench. The post Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench appeared first on MarkTechPost .
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Meta FAIR, Kimi-K2.6, continual learning, prompt engineering.
Primary source verification: published directly by marktechpost with first-party authority.