Reinforced Attention Learning

TL;DR

Reinforced Attention Learning (RAL) is a novel policy-gradient framework that optimizes internal attention distributions in multimodal models, leading to improved reasoning and perception across image and video benchmarks.

Contribution

RAL introduces a new approach to post-training in multimodal models by directly optimizing attention policies instead of output tokens, enhancing cross-modal understanding.

Findings

RAL achieves consistent performance improvements over baselines.

On-Policy Attention Distillation enhances cross-modal alignment.

Experiments demonstrate effectiveness across diverse benchmarks.

Abstract

Post-training with Reinforcement Learning (RL) has substantially improved reasoning in Large Language Models (LLMs) via test-time scaling. However, extending this paradigm to Multimodal LLMs (MLLMs) through verbose rationales yields limited gains for perception and can even degrade performance. We propose Reinforced Attention Learning (RAL), a policy-gradient framework that directly optimizes internal attention distributions rather than output token sequences. By shifting optimization from what to generate to where to attend, RAL promotes effective information allocation and improved grounding in complex multimodal inputs. Experiments across diverse image and video benchmarks show consistent gains over GRPO and other baselines. We further introduce On-Policy Attention Distillation, demonstrating that transferring latent attention behaviors yields stronger cross-modal alignment than standard knowledge distillation. Our results position attention policies as a principled and general alternative for multimodal post-training.