Original post · AI translation
– arxiv. org/abs/2610.32046
标题：“测试时记忆演化能否在科学发现中胜过测试时训练？”
加州大学圣巴巴拉分校的一篇新论文表明，大语言模型进行数学发现时可以跳过测试时训练，因为演化式记忆取得了更高的分数，且使用的算力少得多。
也就是说，在数学发现任务中，演化式记忆胜过测试时训练，而且所需算力少得多。
测试时训练让模型在搜索过程中，对自己经过评分的尝试运行强化学习，从而不断改进。这样的权重更新缓慢且成本高，因此该方法很难扩展到更大的模型。
他们构建了 TTME，它保持权重不变，并将经验存入记忆。记忆中保存着从先前尝试中总结出的实用编码规则，以及分数提升背后的思路，用来指导后续尝试。
在 5 项数学任务、2 种框架和 3 个模型上，TTME 每次都超过了已报告的测试时训练分数。使用 Qwen3-8B 完成其中 3 项任务时，与测试时训练相比，它使用的 FLOPs 减少了 82.8% 至 84.6%，运行时间缩短了 47.7% 至 58.4%。
如果你的大语言模型搜索循环有自动检查器，可以先尝试用记忆保存经验，再考虑为搜索过程中的权重更新付出成本。
Original
New UC Santa Barbara paper shows that LLM math discovery can skip test-time training, since an evolving memory scored better and used far less compute. i.e. evolving memory beats test-time training on math discovery, with far less compute. Test-time training lets a model improve during a search by running RL on its own scored attempts. Those weight updates are slow and costly, so the method is hard to scale to bigger models. They built TTME, which keeps the weights fixed and stores lessons in memory. The memory holds practical coding rules from earlier tries and the ideas behind score gains, which guide later attempts. Across 5 math tasks, 2 frameworks, and 3 models, TTME beat the reported test-time training scores every time. With Qwen3-8B on 3 tasks, it used 82.8% to 84.6% fewer FLOPs and 47.7% to 58.4% less runtime than test-time training. If your LLM search loop has an automatic checker, try a memory of lessons before paying for weight updates mid-search.