arXiv:2609.05437 releases NormReact with 450 annotated scenarios for metanorm evaluation
arXiv:2609.05437 introduces NormReact to measure second-order metanorm reasoning. Six LLMs overpredict punishment relative to human annotations, with error increasing at greater social distance. The distortion affects any deployment that models observer responses to violations.
The paper introduces a two-axis framework for metanorm reasoning: emotional appraisal and behavioral response. It defines tasks for predicting self-regulation by violators and other-regulation by observers. NormReact annotates each scenario for violator gender and observer social closeness. Six models were tested against human baselines. Alignment drops as social distance increases.
Data indicate LLMs default to harsher enforcement patterns. They select punishment categories at rates 20-35 points above human judgments in distant-observer conditions. Inaction predictions fall below human levels across all closeness tiers. Gender annotations show smaller but consistent divergence.
Operational risk follows directly. Systems deployed in mediation or policy simulation will overstate sanction frequency and understate tolerance. This mismatch scales with relational distance, the exact variable that governs real-world norm enforcement. No prior first-order alignment benchmark captured this second-order distortion.
Subsequent work must expand NormReact coverage and retest frontier models on the same splits.
Anthropic: Claude 3.5 successor alignment gap on NormReact distant-observer subset remains above 18 points through 2027 release
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.05437)
- [2]Supporting Source(https://arxiv.org/abs/2305.10601)