Iterative Deployment Improves Planning Skills in LLMs
This addresses AI safety concerns and offers an alternative training regime for LLMs, though it is incremental as it builds on existing fine-tuning and RL concepts.
The paper tackles the problem of improving planning skills in large language models by using iterative deployment with user-curated data, resulting in substantial improvements and emergent generalization to longer plans.
We show that iterative deployment of large language models (LLMs), each fine-tuned on data carefully curated by users from the previous models' deployment, can significantly change the properties of the resultant models. By testing this mechanism on various planning domains, we observe substantial improvements in planning skills, with later models displaying emergent generalization by discovering much longer plans than the initial models. We then provide theoretical analysis showing that iterative deployment effectively implements reinforcement learning (RL) training in the outer-loop (i.e. not as part of intentional model training), with an implicit reward function. The connection to RL has two important implications: first, for the field of AI safety, as the reward function entailed by repeated deployment is not defined explicitly, and could have unexpected implications to the properties of future model deployments. Second, the mechanism highlighted here can be viewed as an alternative training regime to explicit RL, relying on data curation rather than explicit rewards.