Paper Finds Memory Evolution Outperforms Test-Time Training

A research paper published on Hugging Face on Oct 23 introduces Test-Time Memory Evolution (TTME), a method for helping large language models conduct scientific discovery by updating external memory instead of changing model weights. The paper compares TTME with test-time training (TTT), which updates model weights during discovery and adds computational cost.

The researchers evaluated both methods within the same discovery framework, changing only the test-time adaptation method. TTME outperformed TTT on all five mathematical tasks in the evaluation. The authors report that TTME reduced computation by 82.8-84.6% in FLOPs and runtime by 47.7-58.4%.

The paper describes TTME as a plug-and-play alternative that distills experience into external memory to guide later discovery steps. The reported results apply to the evaluated discovery settings. The paper lists Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang and Yuheng Bu as authors, and identifies the University of California, Santa Barbara affiliation. It says the work was submitted on Oct 28. The abstract notes that earlier TTT frameworks had enabled open-weight large language models to outperform inference-only frameworks powered by frontier models on scientific discovery tasks; it presents their additional online weight-update cost as the motivation for testing TTME's efficiency. The article provides no results beyond the five mathematical tasks.

A research paper published on Hugging Face on Oct 23 introduces Test-Time Memory Evolution (TTME), a method for helping large language models conduct scientific discovery by updating external memory instead of changing model weights. The paper compares TTME with test-time training (TTT), which updates model weights during discovery and adds computational cost.

The researchers evaluated both methods within the same discovery framework, changing only the test-time adaptation method. TTME outperformed TTT on all five mathematical tasks in the evaluation. The authors report that TTME reduced computation by 82.8-84.6% in FLOPs and runtime by 47.7-58.4%.

The paper describes TTME as a plug-and-play alternative that distills experience into external memory to guide later discovery steps. The reported results apply to the evaluated discovery settings. The paper lists Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang and Yuheng Bu as authors, and identifies the University of California, Santa Barbara affiliation. It says the work was submitted on Oct 28. The abstract notes that earlier TTT frameworks had enabled open-weight large language models to outperform inference-only frameworks powered by frontier models on scientific discovery tasks; it presents their additional online weight-update cost as the motivation for testing TTME's efficiency. The article provides no results beyond the five mathematical tasks.