Resolving Action Bottleneck: Agentic Reinforcement Learning Informed by Token-Level Energy

TL;DR

This paper identifies the action bottleneck in agentic reinforcement learning, showing that current methods misallocate training signals, and proposes ActFocus, a token reweighting approach that improves performance across multiple environments.

Contribution

The paper introduces ActFocus, a simple token reweighting method informed by energy-based modeling, to better allocate training signals in agentic reinforcement learning.

Findings

ActFocus outperforms PPO and GRPO across four environments.

It yields up to 65.2 and 63.7 percentage point improvements.

Token-level training signals concentrate on action tokens, revealing the Action Bottleneck.

Abstract

Agentic reinforcement learning trains large language models using multi-turn trajectories that interleave long reasoning traces with short environment-facing actions. Common policy-gradient methods, such as PPO and GRPO, treat each token in a trajectory equally, leading to uniform credit assignment. In this paper, we critically demonstrate that such uniform credit assignment largely misallocates token-level training signals. From an energy-based modeling perspective, we show that token-level training signals, quantified by their correlations with reward variance of different rollouts sampled from a given prompt, concentrate sharply on action tokens rather than reasoning tokens, even though action tokens account for only a small fraction of the trajectory. We refer to this phenomenon as the Action Bottleneck. Motivated by this observation, we propose an embarrassingly simple token reweighting approach, ActFocus, that downweights gradients on reasoning tokens, along with an additional energy-based redistribution mechanism that further increases the weights on action tokens with higher uncertainty. Across four environments and different model sizes, ActFocus consistently outperforms PPO and GRPO, yielding final-step gains of up to 65.2 and 63.7 percentage points, respectively, without any additional runtime or memory cost.