Test-Time Learning with an Evolving Library

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Xu, Weijia, Sordoni, Alessandro, Singh, Chandan, Gero, Zelalem, Galley, Michel, Yuan, Xingdi, Gao, Jianfeng
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910219277172736
author Xu, Weijia
Sordoni, Alessandro
Singh, Chandan
Gero, Zelalem
Galley, Michel
Yuan, Xingdi
Gao, Jianfeng
author_facet Xu, Weijia
Sordoni, Alessandro
Singh, Chandan
Gero, Zelalem
Galley, Michel
Yuan, Xingdi
Gao, Jianfeng
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_14477
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Test-Time Learning with an Evolving Library
Xu, Weijia
Sordoni, Alessandro
Singh, Chandan
Gero, Zelalem
Galley, Michel
Yuan, Xingdi
Gao, Jianfeng
Machine Learning
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.
title Test-Time Learning with an Evolving Library
topic Machine Learning
url https://arxiv.org/abs/2605.14477