Test-time memory evolution beats test-time training on math discovery
Original title: Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?
TTME updates external memory instead of weights and cuts FLOPs by 82.8–84.6% and runtime by 47.7–58.4%.
The paper introduces Test-Time Memory Evolution (TTME), a plug-and-play alternative to test-time training for scientific discovery with open-weight language models. TTME distills experience into external memory to guide later discovery without online weight updates. Under the same discovery framework, the authors report that TTME beat TTT on all five mathematical tasks while lowering compute and runtime.
Unconfirmed: Performance and cost figures are the authors' reported results on the evaluated settings.
- TTME avoids online weight updates during discovery.
- It outperformed TTT on all five mathematical tasks evaluated.
- FLOPs fell 82.8–84.6% and runtime 47.7–58.4% versus TTT.
- Only the test-time adaptation method was swapped in the comparison.
From the original article
1 min read at huggingface.co
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.
By Yepeng Liu, Shengmai Chen, Xuandong Zhao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.