Masked IRL: LLM-Guided Reward Disambiguation from Demonstrations and Language

ArXi:2511.14565v2 Announce Type: replace-cross Robots can adapt to user preferences by learning reward functions from nstrations, but with limited data, reward models often overfit to spurious correlations and fail to generalize. This happens because nstrations show robots how to do a task but not what matters for that task, causing the model to focus on irrelevant state details. Natural language can directly specify what the robot should focus on, and, in principle, disambiguate between many reward functions consistent with the nstrations.