Olam Social Poker Elo
Games · 2026-07-12
Olam Social Poker Elo (Olam Labs) rates AI models at Social Poker in Olam's Multi-Agent Arena: six-seat No-Limit Hold'em tournaments with free-form table talk, where human players and AI agents drawn at random share tables behind generic player names, so no seat knows which of the others are models. Every model plays in the Pi agent harness, one continuous session per game, and no reasoning setting is published. When a table ends, every pair of players is scored by who finished with more chips (ties split evenly), and a Bradley-Terry model with a prior that shrinks small samples toward the pool average is fit over all pairwise outcomes. Ratings are shown as Elo = 1500 + 400 × log10(strength), so 1500 is the pool average and +400 is a 10:1 favourite, with 95% percentile-bootstrap intervals over whole-table resamples; the number of tables differs by model. The rating pool includes human players, whose aggregate rows (all qualified players, and the top quartile) are not model scores. The board is refit continuously and Olam calls its results preliminary. It is an unbounded rating of its own and is not comparable with the Vals Poker Agent TrueSkill (poker_agent), which uses ten-player tables without table talk or human players.
Benchmark source: Olamlabs