Does an AI's skill at Sokoban-like puzzles imply it was trained on them?
A user says an agent seems very good at grid-based, turn-based puzzles. The reply: “I have no clue,” followed by a guess that it probably trained on some puzzles, though probably not those.
TLDR
One participant says agents seem to generalize “a fair bit” in their experience. Another asks whether an agent’s apparent skill at playing and solving grid-based, turn-based Sokoban-like games means it was trained on such games. The reply explicitly acknowledges uncertainty—“I have no clue”—then guesses probably not those puzzles, but probably some puzzles. The same participant suggests training with controlled data would be useful for research.
Does an AI's skill at Sokoban-like puzzles imply it was trained on them?
A user says an agent seems very good at grid-based, turn-based puzzles. The reply: “I have no clue,” followed by a guess that it probably trained on some puzzles, though probably not those.
TLDR
One participant says agents seem to generalize “a fair bit” in their experience. Another asks whether an agent’s apparent skill at playing and solving grid-based, turn-based Sokoban-like games means it was trained on such games. The reply explicitly acknowledges uncertainty—“I have no clue”—then guesses probably not those puzzles, but probably some puzzles. The same participant suggests training with controlled data would be useful for research.
