BS-Bench is a benchmark and paper for studying strategic misrepresentation in multi-agent LLM play. Four agents play Bullshit with private cards, legal bluffing, objective lie labels, and explicit peer challenge opportunities.
What It Proves
- Seeded TypeScript game engine for hidden-information multi-agent play.
- Hosted-model evaluation pipeline, frozen 600-game pilot, CSV exports, figures, and arXiv-ready paper packaging.
- Claim discipline: every headline result maps to a cohort, metric, CSV, figure, and limitation.
Main Result
In the frozen pilot, a plain-language honesty mandate sharply reduces optional lying when truthful play is available, but it also changes table dynamics: challenge frequency falls and lie success rises. The paper keeps this narrow: one benchmark family, one hosted-model cohort, and descriptive behavioral claims rather than broad claims about deception in general.
Artifact Status
The arXiv PDF builds, the source bundle is packaged, and the extracted upload bundle compiles cleanly.