Tool Verification for Test-Time Reinforcement Learning
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911479630921728 |
|---|---|
| author | Liao, Ruotong Röhrich, Nikolai Wang, Xiaohan Zhang, Yuhui Samadzadeh, Yasaman Tresp, Volker Yeung-Levy, Serena |
| author_facet | Liao, Ruotong Röhrich, Nikolai Wang, Xiaohan Zhang, Yuhui Samadzadeh, Yasaman Tresp, Volker Yeung-Levy, Serena |
| contents | Test-time reinforcement learning (TTRL) has emerged as a promising paradigm for self-evolving large reasoning models (LRMs), enabling online adaptation on unlabeled test inputs via self-induced rewards through majority voting. However, a spurious yet high-frequency unverified consensus can become a biased and reinforced reward signal, leading to incorrect mode collapse. We address this failure mode with T^3RL (Tool-Verification for Test-Time Reinforcement Learning), which introduces test-time tool verification into reward estimation. Concretely, a verifier uses an external tool as evidence (e.g., from code execution) to upweight verified rollouts in a verification-aware voting, producing more reliable pseudo-labels for training. Across various math difficulties (MATH-500, AMC, and AIME 2024) and diverse backbone types, T^3RL significantly improves over TTRL, with larger gains on harder problems. More broadly, T^3RL can be viewed as verified online data synthesis, highlighting test-time tool verification as a key mechanism for stabilizing self-evolution. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_02203 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Tool Verification for Test-Time Reinforcement Learning Liao, Ruotong Röhrich, Nikolai Wang, Xiaohan Zhang, Yuhui Samadzadeh, Yasaman Tresp, Volker Yeung-Levy, Serena Artificial Intelligence Computation and Language Test-time reinforcement learning (TTRL) has emerged as a promising paradigm for self-evolving large reasoning models (LRMs), enabling online adaptation on unlabeled test inputs via self-induced rewards through majority voting. However, a spurious yet high-frequency unverified consensus can become a biased and reinforced reward signal, leading to incorrect mode collapse. We address this failure mode with T^3RL (Tool-Verification for Test-Time Reinforcement Learning), which introduces test-time tool verification into reward estimation. Concretely, a verifier uses an external tool as evidence (e.g., from code execution) to upweight verified rollouts in a verification-aware voting, producing more reliable pseudo-labels for training. Across various math difficulties (MATH-500, AMC, and AIME 2024) and diverse backbone types, T^3RL significantly improves over TTRL, with larger gains on harder problems. More broadly, T^3RL can be viewed as verified online data synthesis, highlighting test-time tool verification as a key mechanism for stabilizing self-evolution. |
| title | Tool Verification for Test-Time Reinforcement Learning |
| topic | Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2603.02203 |