Abstract
We introduce Language Feedback Models (LFMs) that identify desirable behaviour — actions that help achieve tasks specified in the instruction — for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFM can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.
Learning a small and cost-effective Language Feedback Model from LLM feedback. We roll out an initial policy, then prompt an LLM to provide feedback on what actions the policy took during the rollout were productive in achieving the task outlined in the instruction. We then use this data to train a feedback model that predicts whether an action is productive given the instruction.
Policy improvement by imitating desirable behaviour identified by a learned feedback model. Given the instruction, we roll out a base policy, then identify productive actions that help achieve tasks specified in the instruction using the trained feedback model. Finally, we update the base policy by imitating the identified desirable behaviour.
Example of desirable behaviour identified in an example environment in ALFWorld, a kitchen instruction following benchmark.