Search papers, labs, and topics across Lattice.
The paper introduces Adaptive Safety through Knowledge (ASK), a method that selectively queries a language model to assist reinforcement learning agents in out-of-distribution scenarios. ASK uses Monte Carlo Dropout to estimate uncertainty in the RL agent's actions and only consults the LM when this uncertainty exceeds a threshold, thus balancing performance and computational cost. Experiments in FrozenLake show that ASK achieves robust navigation in transfer tasks (reward of 0.95) without in-domain performance degradation, demonstrating the benefits of uncertainty-gated neuro-symbolic integration.
Unleashing smaller LMs only when RL agents are uncertain yields surprisingly robust out-of-distribution generalization without sacrificing in-domain performance.
Reinforcement learning (RL) agents often struggle with out-of-distribution (OOD) scenarios, leading to high uncertainty and random behavior. While language models (LMs) contain valuable world knowledge, larger ones incur high computational costs, hindering real-time use, and exhibit limitations in autonomous planning. We introduce Adaptive Safety through Knowledge (ASK), which combines smaller LMs with trained RL policies to enhance OOD generalization without retraining. ASK employs Monte Carlo Dropout to assess uncertainty and queries the LM for action suggestions only when uncertainty exceeds a set threshold. This selective use preserves the efficiency of existing policies while leveraging the language model's reasoning in uncertain situations. In experiments on the FrozenLake environment, ASK shows no improvement in-domain, but demonstrates robust navigation in transfer tasks, achieving a reward of 0.95. Our findings indicate that effective neuro-symbolic integration requires careful orchestration rather than simple combination, highlighting the need for sufficient model scale and effective hybridization mechanisms for successful OOD generalization.
