aifollow.news Search
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

UC Santa Barbara paper uses evolving memory instead of test-time training for math discoveryMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

The paper proposes TTME, which keeps model weights fixed and stores coding rules and ideas behind score gains from earlier attempts to guide later ones. The source reports that TTME beat the reported test-time training scores across five math tasks, two frameworks and three models. With Qwen3-8B on three tasks, it used 82.8%–84.6% fewer FLOPs and 47.7%–58.4% less runtime.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

加州大学圣塔芭芭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,因为持续演化的记忆取得了更高的评分,且使用的算力少得多。

也就是说,在数学发现任务中,持续演化的记忆胜过测试时训练,同时所需算力少得多。

测试时训练让模型在搜索过程中对自己经过评分的尝试进行强化学习,从而不断改进。这样的权重更新既慢又昂贵,因此该方法很难扩展到更大的模型。

研究人员构建了TTME:它保持权重不变,将经验教训存入记忆。记忆中保存着从先前尝试中获得的实用编码规则,以及促使评分提高的思路,用来指导后续尝试。

在5项数学任务、2种框架和3个模型中,TTME每次都超过了已报告的测试时训练得分。在使用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配有自动检查器,不妨先试试存储经验教训的记忆,再考虑为搜索过程中的权重更新付出成本。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction