Reinforcement learning has proven effective at sharpening the reasoning abilities of language models, yet its benefits are less consistent when applied to visual tasks. A new preprint on arXiv examines this gap, arguing that the problem lies in how failures are diagnosed.

The authors note that a model's poor performance on a visual task can stem from different underlying causes—such as a failure to perceive the relevant visual information or a failure to reason correctly about it. Without distinguishing between these, standard reinforcement learning may target the wrong component, limiting improvement.

The paper, titled "Perception or Reasoning? Diagnosing and Tackling Reinforcement Learning Bottlenecks in VLMs," proposes methods to identify which bottleneck is at play and then address it accordingly. The abstract is brief, but the central claim is clear: treating all visual-task errors as reasoning errors is a mistake, and better diagnosis is key to making reinforcement learning work for vision-language models.