Found first: a primary source the press has not covered yet.
A paper submitted to arXiv on October 23 introduces Test-Time Memory Evolution (TTME), a plug-and-play adaptation method that outperforms test-time training on all five mathematical discovery tasks the researchers evaluated. The paper reports FLOPs reductions of 82.8 to 84.6% and runtime reductions of 47.7 to 58.4% compared to TTT baselines, with results scoped explicitly to the evaluated discovery settings.
What the source says
The work is by Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang, and Yuheng Bu. Where test-time training updates model weights at inference to adapt to new problems, TTME updates an external memory by distilling accumulated experience, leaving model weights untouched. The method is tested on open-weight LLMs across five mathematical discovery tasks. On all five, TTME outperforms TTT while requiring 82.8 to 84.6% fewer FLOPs and 47.7 to 58.4% less wall-clock time. The paper does not name the five tasks in the abstract.
Why it matters
Test-time training has attracted interest as a way to boost LLM performance on hard problems without retraining, but gradient updates at inference carry real cost. These results show that, on the mathematical discovery tasks tested, weight updates are not necessary to achieve equivalent or better gains. For practitioners building LLM-based scientific discovery pipelines, memory-based adaptation becomes a lower-cost option worth evaluating. How far this transfers beyond the five evaluated tasks is not established by the paper.