Search papers, labs, and topics across Lattice.
We introduce AdaStep, an adaptive step-credit weighting method for agentic reinforcement learning that addresses the challenge of sparse outcome rewards in long-horizon tasks by dynamically adjusting the influence of local advantages on trajectory-level signals. By formulating step-advantage estimation as a mean-squared-error problem and deriving an optimal shrinkage coefficient based on signal-to-total variance, AdaStep effectively disentangles action-dependent returns from downstream randomness. Experiments across ALFWorld, WebShop, and ScienceWorld demonstrate significant performance gains with minimal computational overhead.
Fine-grained credit assignment in LLM agents can be unreliable due to downstream randomness, but AdaStep's optimal shrinkage coefficient precisely balances local and global rewards.
Long-horizon LLM agents are typically trained with sparse outcome rewards, making trajectory-level objectives too coarse to distinguish the contribution of individual decisions. Step-level credit assignment provides finer-grained supervision, but its estimates can be unreliable because observed returns also depend on subsequent actions, environment transitions, and trajectory length. We propose AdaStep, an Adaptive Step-credit weighting method that controls how strongly each group-derived local advantage modifies the trajectory-level signal. We formulate this weighting as a mean-squared-error estimation problem for the latent step advantage and, under an explicit conditional sampling assumption, derive an optimal per-state shrinkage coefficient. The coefficient admits a signal-to-total-variance interpretation: it preserves local credit when return variation is attributable to the selected action and suppresses it when variation is dominated by downstream randomness. AdaStep requires only lightweight scalar computation, with no critic, additional rollouts, or extra model inference. Experiments with three model backbones on ALFWorld, WebShop, and ScienceWorld show consistent improvements over baselines at low computational cost.
