Title: Scalable Inductive Generalization in RL from Specifications via Decoupled Behavioral Cloning

Abstract: Inductive generalization offers a framework for reinforcement learning (RL) wherein task instances that are inductively related are associated with inductively related policies. While previous studies have modeled this structure through a higher-order policy-evolution function trained directly via RL, this approach is hindered by limited scalability. Specifically, as the number of training tasks increases, the aggregation of reward signals tends to become noisy and contradictory, which destabilizes the training process and diminishes generalization performance. To address these challenges, we introduce DIBS, a method based on decoupled behavioral cloning that disentangles the learning of task-specific policies from the learning of the evolution function. Our approach first employs standard RL to train individual teacher policies for each task. Subsequently, it fits the evolution function using behavioral cloning on state-action pairs labeled by these teachers. This strategy substitutes the unstable, noisy reward aggregation with dense and robust supervision. Consequently, DIBS delivers marked enhancements in both training stability and zero-shot generalization capabilities compared to established RL and meta-RL algorithms.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC