Abstract:We introduce EvoLib, a test-time learning framework that enables large language models to accumulate, reuse, and evolve knowledge across problem instances without parameter updates or external supervision. Instead of adapting model parameters, our approach maintains a shared library of knowledge abstractions, including modular skills and reflective insights, automatically extracted from the model's own inference trajectories. To support continual improvement, we introduce a principled weighting and consolidation mechanism that jointly optimizes for immediate utility and long-term value. This allows simple, instance-specific abstractions to evolve into more general and reusable ones over time. Across challenging benchmarks in mathematical reasoning, code generation, and multi-turn agentic environments, EvoLib improves substantially over the top test-time scaling and learning methods without ground-truth feedback.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2605.14477 [cs.LG] |
| (or arXiv:2605.14477v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.14477 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Weijia Xu [view email]
[v1]
Thu, 14 May 2026 07:18:12 UTC (335 KB)