LLMs can learn from their own experience, but the best learning method depends on the task—sometimes raw history works better than summaries, and parameter training can cause harmful negative transfer.
S³Gym is a benchmark that tests whether LLMs can improve themselves by testing their own behavior, judging the results, and learning from experience. The researchers tested three ways to incorporate learning—using conversation history, summarized memories, and parameter training—across seven text-based games.