AGI Lab · Bench
Reversi Bench
Does more thinking make a stronger player?
Reversi matches between the same model at different reasoning efforts, measuring how thinking depth changes playing strength. The referee CLI holds the board and legal moves; players know only their color and an opaque game id — who the opponent is (which model, which effort) is written into the record only after the game ends. Alongside wins, stone margins, and move counts, reasoning-token usage is recorded, publishing the raw relationship between thinking cost and playing strength. Cross-model games share the same ledger.
Loading records…
Ratings come from a regularized Bradley-Terry model, so a seat is ranked by whom it beat rather than by raw win count. All scoring is mechanical — no LLM judging, and rankings move a lot while sample sizes are small.
Every game record, the referee CLI, and the methodology are public at github.com/tempi-tech/reversi-bench (METHOD.md, standings.json).