Metraipolis › archive › dc_2ccd2a64565c
Multilingual Verifier Bias in RLVR: Benchmark, Rollout Diagnosis, and the Cross-Lingual Selection Bottleneck — arXiv:2608.20362
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.20362
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-18T10:18:09.508Z
- SHA-256 of the archived text
90024d1dbfd78bc1e7d442d724583c4159bd4561352fd4416a2eba1d287f1e12- Stable URL of this page
https://metraipolis.com/documento/dc_2ccd2a64565c
How to cite
Multilingual Verifier Bias in RLVR: Benchmark, Rollout Diagnosis, and the Cross-Lingual Selection Bottleneck — arXiv:2608.20362 — arXiv (en), https://arxiv.org/abs/2608.20362 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_2ccd2a64565c (Metraipolis archive dc_2ccd2a64565c, sha256 90024d1dbfd78bc1e7d442d724583c4159bd4561352fd4416a2eba1d287f1e12)
BibTeX
@misc{dc2ccd2a64565c,
title = {Multilingual Verifier Bias in RLVR: Benchmark, Rollout Diagnosis, and the Cross-Lingual Selection Bottleneck — arXiv:2608.20362},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.20362},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_2ccd2a64565c; sha256 90024d1dbfd78bc1e7d442d724583c4159bd4561352fd4416a2eba1d287f1e12},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.20362. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Reinforcement learning with verifiable rewards (RLVR) is a standard recipe for training large language models on mathematical reasoning, where an answer verifier serves as a language-neutral reward function. We show that this assumption fails in multilingual settings: an exact-match verifier turns format and script variation into language-dependent false-negative reward noise. We introduce a reusable protocol for auditing multilingual RLVR rewards: a verifier-robustness suite, a rollout-diagnosis procedure, and language-conditioned reward-error metrics for Japanese, English, and Chinese answers. On MGSM rollouts with k=8, the exact-match proxy rejects trusted-correct answers at sharply different rates by language across Qwen3-4B, Qwen3-8B, and Llama-3.1-8B-Instruct; for Qwen3-8B, the false-negative rate reaches 0.642 on JP against 0.122 on EN and 0.073 on CN. A plain-numeric probe localizes the mechanism to the final-answer interface: an interface model drives reward-error VLB to zero while the residual accuracy gap is unchanged. We then expose a cross-lingual selection bottleneck: on MGSM250 rollouts, a target-local aggregation rule using no trusted labels closes 55-78% of the average selection gap, and over 95% of repairs require genuine cross-lingual support. The bottleneck replicates on a 483-problem MATH-500 set. A controlled training audit shows that rule-GRPO raises trusted accuracy while the reward-error VLB stays high. The unifying message is operational: multilingual RLVR rewards should be audited by language and by answer interface before they are optimized. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-24 in cs.CL). Categorie arXiv: Computation and Language (cs.CL) ; Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2608.20362 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-18T10:17:37.552Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-24?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).