NormGym benchmark shows agents learn from own norm violations but not from others
xuanalogue · x · 2026-11-20
Researchers released NormGym, a benchmark that evaluates agents' normative competence — the human-like ability to infer a new community's social norms by watching what others do and what gets corrected, rather than being told.
Key findings:
- Agents can sometimes learn and adjust from their own norm violations;
- They struggle to learn norms from the behavior of people around them.
The authors argue this exposes a core gap: current agents pursue their own goals rather than thinking and acting as members of a community — a key challenge for building socially capable agents. Paper and project page are public.
More from coding & agent
- MAC Bench: 8-agent teams beat solo agents via headcount, not coordination — gregd_nlp · 2026-11-20
- Every adopted 'compound engineering' company-wide so its codebase gets easier to maintain — every · 2026-11-20
- Start custom LLMs at 30B: runs on a single GPU and works with LoRA — ziv_ravid · 2026-11-20
- Not every agent task needs a general LLM — specialized small models by task type — ziv_ravid · 2026-11-20
- Before fine-tuning, distill traces into skills in-context to test if training is worth it — ziv_ravid · 2026-11-20
- Fine-tuning is mostly iteration: cleaning data beats tuning hyperparameters — ziv_ravid · 2026-11-20