aiqbench analysis
This paper introduces SelectTTRL, a method that uses internal model representations to identify which test problems warrant expensive reinforcement learning updates during inference. By filtering out easily solvable instances before applying Test-Time Reinforcement Learning (TTRL), the approach significantly reduces computational costs while maintaining or improving performance across various reasoning benchmarks. The work advances efficient adaptation strategies for large language models by establishing a theoretical link between response representations and problem solvability.
SelectTTRL: Representation-Guided Data Selection for Efficient Test-Time Reinforcement Learning
arXiv:2610.16104v1 Announce Type: new Abstract: Test-Time Reinforcement Learning (TTRL) enables large language models to adapt directly on unlabeled test problems by constructing pseudo-labels from multiple sampled responses. However, existing TTRL methods typically apply the same expensive sampling and optimization procedure to all test problems, despite substantial differences in their training utility. In this paper, we study whether such problems can be identified before costly test-time training, without access to ground-truth answers. We uncover a stable relationship between internal response representations and problem-level empirical accuracy. Specifically, we define COS as the cosine similarity between a model's response representation and a correctness-related direction constructed from an independent question-answering dataset. COS consistently correlates with problem solvability across different models, datasets, and experimental settings, and we further provide a theoretical analysis explaining this relationship. Building on this finding, we propose SelectTTRL, a representation-guided data selection method that ranks test problems using COS and applies TTRL only to a selected percentile interval. SelectTTRL leaves the standard TTRL pipeline unchanged for the selected problems, while avoiding unnecessary multi-response sampling and reinforcement learning updates on the remaining ones. Experiments across multiple reasoning benchmarks demonstrate that SelectTTRL substantially reduces test-time computation while maintaining or improving the performance of standard TTRL. Code is available at https://github.com/heyiliyi/SelectTTRL-public.
Abstract:Test-Time Reinforcement Learning (TTRL) enables large language models to adapt directly on unlabeled test problems by constructing pseudo-labels from multiple sampled responses. However, existing TTRL methods typically apply the same expensive sampling and optimization procedure to all test problems, despite substantial differences in their training utility. In this paper, we study whether such problems can be identified before costly test-time training, without access to ground-truth answers. We uncover a stable relationship between internal response representations and problem-level empirical accuracy. Specifically, we define COS as the cosine similarity between a model's response representation and a correctness-related direction constructed from an independent question-answering dataset. COS consistently correlates with problem solvability across different models, datasets, and experimental settings, and we further provide a theoretical analysis explaining this relationship. Building on this finding, we propose SelectTTRL, a representation-guided data selection method that ranks test problems using COS and applies TTRL only to a selected percentile interval. SelectTTRL leaves the standard TTRL pipeline unchanged for the selected problems, while avoiding unnecessary multi-response sampling and reinforcement learning updates on the remaining ones. Experiments across multiple reasoning benchmarks demonstrate that SelectTTRL substantially reduces test-time computation while maintaining or improving the performance of standard TTRL. Code is available at this https URL.
Machine Learning (cs.LG)
arXiv:2610.16104 [cs.LG]
(or arXiv:2610.16104v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.16104 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 298 words
Related coverage
BRICS Draws a New Global Line: Full Text of New Delhi Declaration
slguardian.org · Sep 12, 2026
Links #5: 2026/07
LessWrong · Jul 31, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026
SQL Data Types: Reference and Best Practices
Databricks · Sep 8, 2026