ArXiv 2608.14566 documents three evaluation gaps in LLM normative reasoning
Preprint 2608.14566 isolates the missing half of AI moral evaluation: norm identification and application. It maps three concrete gaps and issues a research agenda for formal representations and expert datasets. The work reframes evaluation from value matching to verifiable normative competence.
The paper reviews 22 benchmarks published 2020-2025. 19 focus on output agreement with Moral Foundations Theory or Kohlberg stages. Three attempt intermediate reasoning traces, none supply expert-annotated norm-application ground truth. Citation counts show value-problem papers receive 4.7 times more citations than norm-problem papers in the same period.
Absence of formal normative representations blocks comparison across theories. No dataset records which morally relevant features a model must detect before applying a rule. Existing multiple-choice formats collapse identification and application into a single accuracy score, masking where failures occur.
Operational consequence is that deployed models can pass value surveys yet produce inconsistent rulings in identical factual scenarios that differ only in jurisdiction or role. The authors propose standardized deontic logic encodings, expert norm datasets, and separate metrics for feature detection versus rule application.
Next milestones are release of a 5,000-example expert-annotated norm corpus and a benchmark requiring explicit feature-to-norm mapping by Q4 2026.
Aidan Kierans: At least two peer-reviewed benchmarks will include separate norm-application and feature-identification scores by December 2026.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2608.14566)
- [2]Supporting Source(https://arxiv.org/abs/2305.07005)
- [3]Supporting Source(https://arxiv.org/abs/2402.14763)