arxiv.org/abs/2610.33817 EnvLunch: The Environment as a Free Lunch for Learning to Interact
EnvLunch is a warm-up stage for language-model agents that learns from executable environments such as databases without any human annotation. It beats the baseline by over 9 points on average, and the baseline needs 8 to 32 times more labeled examples to match its smallest-budget accuracy.
— via arXiv Databases, Wei Liu, Min Gao, Yali Du, Yulan He
