Research map / Large language models / AI and society

Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces NormReact, a 450-scenario human-annotated benchmark for evaluating LLM metanorm reasoning (violator self-regulation, observer other-regulation, and norm enforcement across social distance), finding LLMs overpredict punitive sanctions versus humans.

Problem

Existing social norm benchmarks do not test second-order (metanorm) reasoning: how violators and observers are expected to feel and react to norm violations, and how these reactions vary with relational closeness between observer and violator.

Approach

Operationalizes metanorm reasoning along emotional appraisal (Haidt moral emotions) and behavioral responses (descriptive vs injunctive sanctions), with three binary tasks (self-regulation, other-regulation, norm enforcement) and social distance (strong tie, weak tie, stranger) plus violator gender manipulation, comparing LLM predictions to 871 human raters.

Key results

• Average behavioral F1 for descriptive norms drops from about .8 (strong ties) to .7 (weak ties) to .48 (strangers).
• Closed-source models approximate human shame (~13-17% vs 19%) reasonably, while open-source models overestimate shame (~24-28%) and guilt (~28-32%).
• Across models, LLMs overpredict sanctions and gossip (up to 66% overprediction for weak ties in open-source models) where humans prefer inaction; humans report 60% do-nothing for strangers.

Surprising findings

• Claude underestimates both norm appropriateness and wrongness, unlike other closed-source models.
• LLMs shift from gossip to verbal confrontation when asked for ideal responses (injunctive framing), whereas humans keep a preference for restraint.

Limitations

• Scope limited to metanorms in American society.
• Only instruction-tuned models evaluated; unclear whether bias originates in pretraining or post-training.
• Did not test prompt wording variations; name-swap manipulation may be insufficient to activate gendered expectations.

Experimental setup

Datasets: Social-Chem-101

Benchmarks: NormReact

Metrics: F1 precision recall Spearman correlation chi-square entropy bootstrap confidence intervals

Baselines: human ratings (leave-one-out ceiling)

Builds on: Social-Chem-101 Haidt moral emotions framework Bicchieri social norm theory

Closest papers by meaning

• MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)
• Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities (AAAI 2026 · similarity 0.88)
• Language Models Represent Beliefs of Self and Others (ICML 2024 · similarity 0.86)
• Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.86)
• Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.86)
• Position: Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback (ICML 2024 · similarity 0.86)