AI

NormGym: Evaluating Normative Competence through Social Interaction

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/11/2026, 05:03:24 UTC.

Briefings

The NormGym benchmark evaluates whether LLM agents can acquire unknown social norms through interaction rather than just recognizing existing ones. Results show the strongest model achieved a 55.3% joint task-and-compliance success rate, yet it struggled to infer unseen norms from others' behavior and sanctions. Fine-tuning was found insufficient to improve this capability.

Key points

• NormGym is an interactive benchmark where agents must infer hidden, arbitrary norms from social interaction to succeed in multi-agent tasks.
• The strongest evaluated LLM model achieved a 55.3% joint task-and-compliance success rate in NormGym.

Context

Normative competence is essential for agents operating in unfamiliar human communities, but existing evaluations often test recognition of known norms rather than acquisition of unknown ones. This benchmark addresses that gap by requiring inference of unseen norms from social interaction.

Limitations

Results are from a preprint (arXiv:2611.08613v1) with limited scope. The study covers diverse norms and environments but does not include all possible agent types or community structures. Findings are subject to further validation and may vary across different model architectures and interaction protocols.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/11/2026, 06:56:27 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeic6slii7r6fhjkzhi33om7m3sv5bwefqmnhor5zbbuu6ezeam23iq