AdaStep: Adaptive Step Credit Weighting for Agentic Reinforcement Learning

2026-10-02

AI Summary

In long-horizon LLM agent training, sparse rewards render step-level credit assignment unreliable. This work proposes AdaStep, a method that formulates weight construction as a mean squared error estimation problem and derives an optimal per-state shrinkage coefficient. By adaptively weighting local advantages, it controls their influence on trajectory-level signals. Notably, this approach requires only lightweight scalar computation, incurring no additional critic networks or model inference overhead. Evaluations on benchmarks such as ALFWorld demonstrate that AdaStep significantly improves the performance of three model baselines at low computational cost, effectively optimizing the credit assignment mechanism in reinforcement learning.

Abstract

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.

Problem

• credit assignment
• reinforcement learning
• LLM agents
• sparse rewards
• step-level advantage

Innovation

• Adaptive Step Credit Weighting
• Agentic Reinforcement Learning
• Shrinkage Coefficient
• Credit Assignment
• Signal-to-Variance