{"slug": "when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis", "title": "When Models Disagree: Rethinking LLM Evaluation for Public Comment Analysis", "summary": "Federal agencies using large language models to categorize public comments face a hidden risk: different models can produce fundamentally different categorizations of the same input, yet standard accuracy-based evaluations fail to detect this divergence. A study of 1,260 comments on a USDA docket found that inter-model thematic disagreement exceeded within-model prompt variation, and that an expert rubric suppressed deep interpretive conflict without resolving it. The researchers propose an Interpretive Audit Pipeline that treats multi-model disagreement as diagnostic of interpretive complexity, directing human review toward genuinely ambiguous input.", "body_md": "arXiv:2605.29025v1 Announce Type: new\nAbstract: Federal agencies are deploying large language models (LLMs) to categorize public comment corpora, where the model's organization of the record shapes what policymakers see and which arguments register. Standard evaluation, anchored on stance accuracy against a small validated set, cannot detect when different models produce materially different categorizations of the same public input. We propose an Interpretive Audit Pipeline that treats multi-model disagreement as diagnostic of interpretive complexity and directs human review toward genuinely ambiguous public input. Analyzing 1,260 public comments on a federal USDA docket across four LLMs, we find that inter-model thematic divergence exceeds within-model prompt variation, and that an expert rubric suppresses deep interpretive disagreement without resolving it. In a two-stage labeling study on a stratified 40-comment subsample, four LLMs and a human annotator labeled independently and then revised after seeing the others' labels. Revision behavior varied across labelers, and the human annotator's revisions frequently introduced framings absent from the ensemble's collective output. We argue disagreement-based evaluation is a necessary complement to accuracy metrics for LLM-assisted interpretive coding.", "url": "https://wpnews.pro/news/when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis", "canonical_source": "https://arxiv.org/abs/2605.29025", "published_at": "2026-05-29 04:00:00+00:00", "updated_at": "2026-05-29 04:21:16.011504+00:00", "lang": "en", "topics": ["large-language-models", "ai-policy", "ai-ethics", "ai-research", "natural-language-processing"], "entities": ["USDA", "LLMs"], "alternates": {"html": "https://wpnews.pro/news/when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis", "markdown": "https://wpnews.pro/news/when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis.md", "text": "https://wpnews.pro/news/when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis.txt", "jsonld": "https://wpnews.pro/news/when-models-disagree-rethinking-llm-evaluation-for-public-comment-analysis.jsonld"}}