Learning from Situated and Embodied Interaction

How should interactive experience reshape the data, objectives, and evaluations used for language modeling? Despite remarkable progress from scaling over passive corpora, the projected exhaustion of high-quality text poses a growing bottleneck. Interaction with environments, humans, and other agents offers learning signals that are difficult to recover from passive data alone, including pragmatic grounding, internal world models, and social signals. This workshop studies how embodied multi-turn interaction with environments, humans, and other agents can serve as a learning signal for language models. We bring together researchers from language modeling, NLP, embodied AI, robotics, web agents, and multi-agent systems (communities that often study related problems but publish in separate venues). Our central question is how interaction can play a fundamental role in the problem of language modeling itself.