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

We propose Plan, Align, Commit, Think (PACT), a hybrid architecture that unifies a fast, reactive Reinforcement Learning (RL) agent with a slow, deliberative Small Language Model (SLM) planner to mitigate the systematic degradation of reactive policies in novel environments. PACT invokes the SLM asynchronously to generate and verify candidate action plans; once a plan is confirmed to be safe, feasible, and complete via simulation, the RL policy is bypassed and the plan is executed directly, without retraining or modifying the underlying policy. Evaluated on three FrozenLake configurations of increasing difficulty, PACT outperforms all baselines while relying on a 2B-parameter SLM backbone. PACT demonstrates that deliberative planning and reactive execution are more powerful in concert than either is alone.