arXiv:2608.12368 reports label agreement above 80% on 500-item ETHICS benchmark while rationale distributions diverge across harm, respect and justice categories
arXiv:2608.12368 demonstrates that high label agreement between LLMs and humans on moral tasks does not imply shared moral grounds. Rationale analysis reveals consistent redistribution of attention across ethical categories. Alignment evaluation must incorporate principle-level comparison rather than final-label match alone.
The study collected new annotations for final labels and explicit rationales on 500 items spanning five moral domains. Human annotators and multiple model families produced judgments; label agreement was measured against majority human votes while rationale content was coded into categories including harm, respect, promise-keeping, justice, desert and excuse relevance. Data show systematic shifts. Models overweighted harm and justice relative to human baselines and underweighted respect and promise-keeping even on items where final labels matched. Divergence persisted across both frontier closed models and open-weight families. Label-only metrics therefore mask differences in underlying principles. Operational evaluation pipelines that rely solely on agreement will accept systems whose moral priorities remain misaligned with the human annotators they are benchmarked against. Future work requires rationale-level scoring and category-shift thresholds before models are deployed in domains where moral justification is audited.
Machidon et al.: Rationale divergence will exceed 35% category shift on next-generation models unless explicit principle regularization is applied within 12 months.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.12368)
- [2]Supporting Source(https://arxiv.org/abs/2009.11475)