Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?
AuthorsYepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang, Yuheng Bu
AffiliationsUC Santa Barbara · UC Berkeley
The paper proposes improving AI-driven mathematical discovery by evolving its memory instead of retraining its weights at test time.
Key results
Reduction versus test-time training in TTT-Discover runs with Qwen3-8B.
Reduction versus test-time training, excluding evaluation, in TTT-Discover runs with Qwen3-8B.
Qwen3-8B result; lower is better, compared with 1.50525 for test-time training.
Qwen3-8B result; higher is better, compared with 0.9472 for test-time training.
Tasks evaluated across TTT-Discover and ThetaEvolve.
What the paper found
This paper introduces Test-Time Memory Evolution, or TTME, as an alternative to test-time training for scientific discovery. Instead of updating model weights, TTME keeps them fixed and distills search experience into two external memories: operational guidelines for avoiding implementation failures and reusable strategies drawn from score-improving changes. Updated every 2 search steps, these memories guide later program generation and help combine ideas from different search branches. Across TTT-Discover and ThetaEvolve, TTME outperformed test-time training on all 5 mathematical tasks using Qwen3-8B, OpenAI’s gpt-oss-120b, and DeepSeek-R1-0528-Qwen3-8B. For example, with Qwen3-8B, TTME reached an AC1 score of 1.504329 and an AC2 score of 0.949248. In measured TTT-Discover runs, it cut FLOPs by 82.8-84.6% and runtime by 47.7-58.4% compared with test-time training. The results suggest that evolving memory can improve discovery while avoiding the cost of online weight updates; the paper also places this approach in a field where earlier test-time-training systems had shown gains over inference-only systems such as AlphaEvolve.
Original abstract
Recent \emph{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 \emph{test-time adaptation} method for scientific discovery. We introduce \emph{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.
Read the original paperMore in AI for Science
Browse all 52 papers →Can LLMs Uncover Hidden Physical Laws through Free Exploration?
Zhen Yang, Wen Wang, Xander Xu, Ying-Cong Chen
Researchers let LLMs freely explore simulated worlds with altered physics to see whether they could uncover the hidden laws.
LLM2Cert: LLM Synthesis of Certificates with SMT/Lean Verification
Zhipeng Cao, Peiyang Song, Nikolaus Vertovec, Weiru Cui, Zhihao Yang, Mingkang Liu, Luxia Lin, Bai Xue, Alessandro Abate
LLM2Cert uses language models and formal tools to propose and verify mathematical safety and stability certificates for dynamical systems.
Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?
Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski
Agent4Sci tests how far frontier AI agents can improve real scientific ML models—and finds they need more than engineering tricks to make research-level progress.