Observed arrival · 2026-10-18
The Mafia Table Puts Language Models in the Seats
An experiment compares 16 language models playing ten-seat Mafia under three rule sets, with leaderboards, pair results, and an explanation of the methods.
- For
- Readers curious about language-model behavior
- Worth noticing
- Each rule set has 33–42 games; the page says score gaps below roughly 0.2 are noise.
Field notes
The game keeps a ten-seat roster across variants, while changing how players communicate and vote. The Sport Tournament rules restrict mafia coordination after night zero; Open table allows private mafia discussion each night, and the Imbalanced variant gives the mafia a nightly planning meeting against the town's single daily speech. Scores are adjusted against the average for each side under the same rules, and the page cautions that small leaderboard gaps may be noise.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue