Recent test‑time training (TTT) frameworks enable open‑weight large language models (LLMs) to surpass inference‑only frontier models on scientific discovery tasks. However, TTT requires online weight updates during discovery, incurring substantial computational overhead, especially when scaling to larger models.
To address this, we introduce a lightweight test‑time adaptation approach called Test‑Time Memory Evolution (TTME). TTME operates as a plug‑and‑play module that continuously updates an external memory by distilling past experiences, guiding subsequent discovery without altering model weights.
We conduct a clean comparison by keeping the discovery pipeline identical and swapping only the test‑time adaptation component. Across five mathematical tasks, TTME consistently outperforms TTT while reducing FLOPs by 82.8%‑84.6% and cutting runtime by 47.7%‑58.4%.
These results establish TTME as an effective and computationally efficient alternative to TTT in the evaluated discovery settings.
Review