AI Science Tracker
EN

NormGym finds current language-model agents struggle to infer new social norms

Affiliation not stated in the reviewed arXiv metadata · 5 Nov 2026 · computer science

Discovery

NormGym finds current language-model agents struggle to infer new social norms

Source check: Source checked. Not independent scientific confirmation.

Finding: NormGym evaluates agents that must infer hidden norms through interaction while pursuing a task. The strongest evaluated LLM achieved 55.3% joint task-and-compliance success. Models sometimes changed behavior after sanctions directed at themselves but struggled to infer norms from other agents’ behavior and sanctions; fine-tuning was not sufficient in the reported tests.

Technology: NormGym interactive benchmark with large-language-model agents

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
LLM agents interact with simulated communities, observe behavior and sanctions, and attempt to infer and apply hidden norms.
Test and result
LLM agents were evaluated on hidden-norm tasks across benchmark environments; the strongest model’s joint task-and-compliance success was 55.3%, and peer-behavior cues remained difficult to use.
Limitation / still unknown
The findings concern the evaluated agents and benchmark norms/environments; transfer to real communities is not established.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2611.08613v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.