External memory helps models find better mathematical solutions
Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery? · arXiv:2610.32046
In plain words
When a computer searches for a mathematical construction, it can learn from each attempt. One approach updates the computer model itself during the search, which takes extra computation.
This paper introduces Test-Time Memory Evolution, which instead saves practical lessons and useful ideas in memory while leaving the model unchanged.
Across the tested mathematical tasks, the method found better-scoring valid candidates than the reported weight-updating approach. The cost measurements also favored memory on three tasks with one model, but they do not cover every test in the paper.
The research
When a model searches for a mathematical construction, feedback from its tested candidates can help guide later attempts. Test-time training uses that feedback to update model weights, but the online updates add computation.
Can a different way of retaining search experience find stronger candidates with less measured computation?
TTME stores lessons and useful search strategies in external memory, guiding a fixed-weight model to better-scoring candidates than reported TTT results in the tested settings.
Test-time training learns from evaluated candidates by updating model weights, but those online updates add computation. Can a search instead store useful experience outside the model and still find strong mathematical constructions?
The authors compare TTME with reported TTT results within two discovery frameworks, using each task’s verifier to score candidates. TTME has better best-found scores in every reported setting across five tasks.
This supports memory as an alternative in those settings; a separate control also finds better scores than fixed-weight search without adaptation.
The evidence covers two discovery frameworks, three models and five mathematical construction tasks.
The TTT score comparisons use results reported in the original papers, and the source does not establish identical checkpoints or serving conditions across those comparisons; the cost measurements cover only Qwen3-8B on three TTT-Discover tasks.
What the evidence shows
- A different way to adapt
Test-Time Memory Evolution (TTME) is a method for adapting a model during scientific discovery without changing its weights.
The paper asks whether a model can use feedback from evaluated candidate solutions as effectively as test-time training (TTT), which updates weights during the search, while avoiding some of that computational work.
- Two memories guide candidates
TTME keeps two kinds of external memory. Guideline Memory turns outcomes and failures into practical implementation lessons.
Strategy Memory extracts algorithmic ideas from valid changes that improve a task score, groups related ideas, and adapts selected strategies into instructions for later candidates. The fixed-weight model uses this material during generation; its weights remain unchanged.
- Scores across tested settings
Task verifiers score the generated candidates, and the authors compare the best valid score in each reported setting, following each task’s objective direction. TTME scores better than the reported TTT results across two discovery frameworks, three models and five mathematical construction tasks.
The TTT scores come from original papers, and identical checkpoints and serving conditions are not established for every comparison.
- Memory against no adaptation
A separate control compares TTME with the underlying search using fixed model weights but no TTT or TTME. Within each framework, the comparison holds the model, task initialization, evaluator, primary candidate budget and search rules constant.
TTME scores better in the evaluated settings, supporting the interpretation that adding memory benefits those fixed-weight searches.
- Efficiency in a subset
For Qwen3-8B on AC1, AC2 and the Erdős task within TTT-Discover, the authors also compare computation with TTT. TTME reduces runtime by 47.7–58.4% when evaluation time is excluded, and reduces floating-point operations by 82.8–84.6%.
These task-specific ranges, alongside the score results, support an efficiency advantage in this subset; they do not establish it across all evaluated settings.
Deep
- Original paper authors
- Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song
- Submitted to arXiv
- License
- CC BY-SA 4.0