When in Doubt, Plan It Out: Committed Small Language Model Deliberation for Reactive Reinforcement Learning

The article introduces the PACT architecture, which integrates a reactive reinforcement learning (RL) policy with a deliberative Small Language Model (SLM) planner. This hybrid system utilizes a 2B-parameter SLM to asynchronously generate and validate action plans, demonstrating superior performance over traditional RL methods in various FrozenLake environments. The approach highlights the effectiveness of combining deliberative planning with reactive execution, offering insights for practitioners aiming to enhance RL robustness in complex scenarios.