Metraipolis › archive › dc_aee080c666b4

Are Verifier Errors Independent Within a GRPO Group? Evidence from Qwen2.5 Rollouts — arXiv:2609.06386

Source: arXiv (en)
Original: https://arxiv.org/abs/2609.06386
License (as archived): CC0-1.0
Archived: 2026-10-16T12:50:09.169Z
SHA-256 of the archived text: 4d8bb97fc0269791db02bb3041e0348300aaf57da8aadeeb6fa5c912e4fb6597
Stable URL of this page: https://metraipolis.com/documento/dc_aee080c666b4

How to cite

Are Verifier Errors Independent Within a GRPO Group? Evidence from Qwen2.5 Rollouts — arXiv:2609.06386 — arXiv (en), https://arxiv.org/abs/2609.06386 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_aee080c666b4 (Metraipolis archive dc_aee080c666b4, sha256 4d8bb97fc0269791db02bb3041e0348300aaf57da8aadeeb6fa5c912e4fb6597)

BibTeX

@misc{dcaee080c666b4,
  title = {Are Verifier Errors Independent Within a GRPO Group? Evidence from Qwen2.5 Rollouts — arXiv:2609.06386},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.06386},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_aee080c666b4; sha256 4d8bb97fc0269791db02bb3041e0348300aaf57da8aadeeb6fa5c912e4fb6597},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.06386. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Group-based reinforcement learning with verifiable rewards (RLVR) scoresmultiple completions per prompt using automatic verifiers. Analysesbased on independent verifier errors may overlook dependence associatedwith shared answer formats. We investigate this dependence in24,998 groups of eight completions generated by Qwen2.5-1.5B onMATH, GSM8K, and DeepMath-103K. We estimate a pooled within-groupverifier-error correlation of 0.530 (95% confidence interval:0.500--0.560). Under an exchangeable-error model, this correspondsto a design-effect-adjusted effective sample size of 1.70 for aneight-completion group. Dependence varies substantially across answerforms: fractions, radicals, symbolic expressions, and intervals exhibitstronger clustering than unit annotations and percent signs. Replayinggroup-relative advantages across four rule-based verifier configurationsidentifies at least one advantage-sign disagreement in up to 0.83% ofgroups. Because a group is repeated sampling for one prompt, thiswithin-group clustering may reflect shared prompt difficulty as well asshared answer form, and we do not attempt to separate the two this http URL studies of correlated judgments across multiple evaluators, ouranalysis examines dependence across completions scored by the sameverifier. These findings motivate prompt- and answer-form-aware analysesof verifier noise rather than characterizations based solely onaggregate error rates.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-09 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.06386 — 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 2026-10-16T12:30:25.259Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-09?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).