Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?
Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang, Yuheng Bu
UC Santa Barbara · UC Berkeley
Published October 23, 2026 · on the feed October 28, 2026 · ▲ 30 · arXiv 2610.32046
The abstract
Recent test-time training (TTT) frameworks have enabled open-weight large language models (LLMs) to outperform inference-only frameworks powered by frontier models on scientific discovery tasks. However, TTT introduces additional computational cost for online weight updates during the discovery process, making computational efficiency an important consideration when scaling to larger models. This motivates us to explore an efficient test-time adaptation method for scientific discovery. We introduce Test-Time Memory Evolution (TTME), a plug-and-play test-time adaptation method, as an alternative to TTT. TTME continually updates the external memory by distilling the experience to guide subsequent discovery without updating model weights. To have a clean comparison, we evaluate TTME and TTT under the same discovery framework, replacing only the test-time adaptation method. Our evaluation shows that TTME outperforms TTT on all five mathematical tasks, while reducing computational cost by 82.8-84.6% in FLOPs and 47.7-58.4% in runtime. These results establish TTME as an effective and computationally efficient alternative to TTT in the evaluated discovery settings.
The Brief
2:00A short podcast summary of the paper. Written and voiced by AI.
Transcript
Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?. By Yepeng Liu and six others, from UC Santa Barbara and UC Berkeley. Published on the twenty-third of October.
Test-time training frameworks have let open-weight large language models outperform inference-only frameworks that run on frontier models, at least on scientific discovery tasks. That success has a price. Test-time training updates the model's weights online while discovery is under way, and those updates add computational cost. As the models being adapted get larger, the cost grows with them, which makes efficiency an important consideration rather than an afterthought. The question the authors set themselves is whether adaptation at test time needs to touch the weights at all.
Their answer is Test-Time Memory Evolution, or TTME. It is a plug-and-play method for test-time adaptation, offered as an alternative to test-time training. Instead of changing the model, TTME changes what the model is given. It keeps an external memory and continually updates it by distilling experience from the discovery process, and that memory then guides the discovery that follows. The model's weights stay exactly as they were.
The comparison is set up to be clean. TTME and test-time training are run inside the same discovery framework, and the only thing swapped between the two runs is the test-time adaptation method itself. Everything else is held fixed, so any difference in outcome or in cost belongs to the choice between updating memory and updating weights.
On all five mathematical tasks evaluated, TTME outperforms test-time training. It also costs less to run. Measured in floating point operations, the reduction is between eighty-two point eight and eighty-four point six percent. Measured in runtime, the saving is smaller but still large: between forty-seven point seven and fifty-eight point four percent.
Within the discovery settings they evaluated, the authors conclude that TTME is an effective and computationally efficient alternative to test-time training. The question in the title gets a qualified yes. On these five tasks, distilled experience held outside the model did more than weight updates did, and it did so for a fraction of the compute.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Learning while searching works, but it is expensive. Here is the bill.
You are running an automated discovery loop: a language model proposes programs for a maths construction task, an evaluator scores them, and the evidence should make later attempts better.
The road ahead
- 1 · The problem ~7 min A discovery loop should learn from its own attempts, but learning by retraining the model's weights during the search is slow and costly.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Keep the model's weights fixed and write what the search learns into an external memory that the model reads in its context.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each search step generates candidates with a fixed model, evaluates them, distills the results into two memories, and feeds the memories into later steps.
- 6 · The experiment ~6 min You’ll guess: Swapping test-time training for TTME inside the same discovery framework, how much runtime was saved?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.