Research

Paper

AI LLM February 26, 2026

Same Words, Different Judgments: Modality Effects on Preference Alignment

Authors

Aaron Broukhim, Nadir Weibel, Eshin Jolly

Abstract

Preference-based reinforcement learning (PbRL) is the dominant framework for aligning AI systems to human preferences, but its application to speech remains underexplored. We present a controlled cross-modal study of human and synthetic preference annotations, comparing text and audio evaluations of identical semantic content across 100 prompts. Audio preferences prove as reliable as text, with inter-rater agreement reaching good levels (ICC(2,k) $\approx$ .80) at $\sim$9 raters -- the first ICC-based reliability characterization in the preference annotation literature for either modality. However, modality reshapes how people judge: audio raters exhibit narrower decision thresholds, reduced length bias, and more user-oriented evaluation criteria, with near-chance cross-modality agreement. Synthetic ratings further align with human judgments and predict inter-rater agreement, supporting their use both for triaging ambiguous pairs and as full replacements for human annotations.

Metadata

arXiv ID: 2602.22710
Provider: ARXIV
Primary Category: cs.SD
Published: 2026-02-26
Fetched: 2026-02-27 04:35

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.22710v1</id>\n    <title>Same Words, Different Judgments: Modality Effects on Preference Alignment</title>\n    <updated>2026-02-26T07:34:15Z</updated>\n    <link href='https://arxiv.org/abs/2602.22710v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.22710v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Preference-based reinforcement learning (PbRL) is the dominant framework for aligning AI systems to human preferences, but its application to speech remains underexplored. We present a controlled cross-modal study of human and synthetic preference annotations, comparing text and audio evaluations of identical semantic content across 100 prompts. Audio preferences prove as reliable as text, with inter-rater agreement reaching good levels (ICC(2,k) $\\approx$ .80) at $\\sim$9 raters -- the first ICC-based reliability characterization in the preference annotation literature for either modality. However, modality reshapes how people judge: audio raters exhibit narrower decision thresholds, reduced length bias, and more user-oriented evaluation criteria, with near-chance cross-modality agreement. Synthetic ratings further align with human judgments and predict inter-rater agreement, supporting their use both for triaging ambiguous pairs and as full replacements for human annotations.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.SD'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.HC'/>\n    <published>2026-02-26T07:34:15Z</published>\n    <arxiv:comment>Submitted to Interspeech 2026 for review</arxiv:comment>\n    <arxiv:primary_category term='cs.SD'/>\n    <author>\n      <name>Aaron Broukhim</name>\n    </author>\n    <author>\n      <name>Nadir Weibel</name>\n    </author>\n    <author>\n      <name>Eshin Jolly</name>\n    </author>\n  </entry>"
}