Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.32572

Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR

Shashi Kumar·Yanis Labrak·Sergio Burdisso·Esaú Villatoro-Tello·Petr Motlicek·Kadri Hacioglu·Andreas Stolcke
View original source ↗Read source document ↗Export metadata ↓

Abstract

Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a target, the student inherits the teacher's errors. We show that using the same pseudo-labels as a reward reference for Group Relative Policy Optimization (GRPO) is more robust to label noise. In the banking domain, GRPO yields 29.3% to 56.2% relative WER improvements over SFT on Whisper-generated labels, and the advantage holds across four Whisper sizes, four teacher families, three domains and four languages. While SFT reproduces almost 100% of teacher errors verbatim, GRPO passes through only 27%. Rollout statistics explain why, erroneous pseudo-label words rarely appear in on-policy rollouts, so group-relative advantages stay aligned with the clean transcript, whereas SFT forces them into the policy's support and collapses the reward variance GRPO relies on.

Explore related subjects

cs.CLeess.AS

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Shashi Kumar, Yanis Labrak, Sergio Burdisso, Esaú Villatoro-Tello, Petr Motlicek, Kadri Hacioglu, Andreas Stolcke. 2026-10-23. Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR. https://arxiv.org/abs/2610.32572

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

A survey of diversity quantification in natural language processing: The why, what, where and how

The concept of diversity received increasing attention in natural language processing (NLP) in recent years. It became an advocated property of datasets and systems, with many measures used to quantify it, but also many cross-paper inconsistencies. To address this fragmentation, we take inspiration from other scientific fields where diversity has been more thoroughly conceptualized. We build upon (Sarkar, 2010) and (Stirling, 2007), two unified cross-domain frameworks inspired from ecology, economy and other domains. The first one advocates explicit justification of diversity endorsement. The latter distinguishes three dimensions of diversity: variety, balance, and disparity. Surveying over 300 recent diversity-related papers from ACL Anthology, we propose an NLP-specific framework with 4 perspectives: why diversity is important, what diversity is measured on, where it is measured, and how. We find that in NLP richness (the simplest variety) and distance averaging (the simplest disparity) are the most prevalent measure families, and that pure balance measures are nearly absent. Moreover, half of the measures used in NLP do not conform to Stirling's framework. Our analysis increases comparability of approaches to diversity in NLP, allowing us to formulate recommendations for the field. We recommend that the field systematically use a unified terminology, and that it engage in debates about the desirable properties of diversity measures as well as about the normative assumptions behind their use.

cs.CL↗
arXiv2026

What MLLMs Learn about When they Learn about Multimodal Reasoning

Evaluation of multimodal reasoning models is typically reduced to a single accuracy score, implicitly treating reasoning as a unitary capability. We introduce MathLens, a benchmark of textbook-style geometry problems that exposes this assumption by operationally decomposing performance into perception, reasoning, and multimodal-specific components. Each problem is derived from a symbolic specification and accompanied by visual diagrams, text-only variants, multimodal questions, and targeted perceptual probes, enabling controlled measurement of each component. Using this decomposition, we show that common training strategies are associated with systematically different capability profiles that are invisible under aggregate accuracy. Across the models we study, reinforcement learning primarily improves perceptual grounding and robustness to diagram variation, while our analysis suggests that textual SFT yields gains through reflective reasoning. In contrast, as perception and reasoning improve, a growing fraction of remaining errors fall outside these components and are categorized as multimodal-specific. These results suggest that apparent progress in multimodal reasoning reflects shifting balances among subskills rather than uniform advancement, motivating evaluation beyond scalar accuracy.

cs.CL↗
arXiv2026

EvalCards: A Framework for Standardized Evaluation Reporting

Evaluation reports guide the use of general-purpose AI (GPAI) models, yet inconsistent reporting often leaves essential questions unanswered. Through a review of prior work, we identify three central issues in evaluation reporting practices today, reproducibility, accessibility, and accountability, and examine how existing reporting artifacts fall short in addressing them. We introduce the EvalCards framework, outlining its key dimensions and presenting a preliminary format for concise evaluation summaries. We conclude by discussing how the framework can guide the future development of evaluation reporting standards.

cs.CL↗