Back to academics

NormGym

2025 - Present | Python | Reinforcement Learning | LLMs

Why I built it

Norms are usually described in natural language, but testing them in dynamic environments is difficult. I built NormGym to stress-test norm-following and norm-violating behavior under controlled settings.

What it does

  • Defines agents with normative constraints and role-specific goals.
  • Simulates interactions in configurable social environments.
  • Tracks rewards, violations, and recovery trajectories.

What I learned

Designing a concise event schema made downstream analysis much easier than collecting unstructured logs.