If a language model can rewrite the software it uses to play a game, how can it tell whether an edit was worth keeping? Jens Lehmann and Sahar Vahdati let each model build a program, play five practice games and revise that program twice. They then compared all three versions on twenty games whose results had been withheld during development. Across 27 independent sequences, the practice scores rose by 8.42 points on average. On the reserved games, the gain shrank to 2.80 points, and twelve sequences finished below their starting score. Cost tells a similarly mixed story. Eighteen final programs were cheaper to run, but only nine of those also played better. By preserving the original programs and their descendants, the study reveals something a final leaderboard cannot show on its own: when an agent already plays well, how much of that ability did it acquire through revision, and how much was present from the start?
aiweekly.co.in
“Where Machines Make Headlines.”