Tsinghua's AAArena tests AI coding agents improving game bots from replays

Tsinghua researchers introduce AAArena, a game-agent benchmark built from 12 games in the university's yearly bot-building contest, with 1,920 archived human programs as rivals.

Coding agents rewrite their bots from match replays without changing model weights. Per Rohan Paul's report of the paper, detailed replays beat win/loss-only feedback in all three games tested, and a Pacman bot reached rank 1 with replays versus rank 11 without.

The paper reports that agents mostly stall on games with complex rules. Tripling the match budget did not lift any of four stuck bots to rank 1, according to the same report.