In a collaboration between UC Davis, Princeton University, Google, and Google DeepMind the paper "Reinforced Attention Learning", published on February 4, 2026, identifies a critical bottleneck in current Multimodal Large Language Models (MLLMs): while Reinforcement Learning (RL) has successfully scaled reasoning in text models, simply forcing multimodal models to generate verbose "chains of thought" often degrades their ability to perceive visual details. The authors argue that standard RL methods optimize for the result—the next token—rather than the process of finding information. To solve this, they introduce Reinforced Attention Learning (RAL), a framework that fundamentally shifts the post-training objective from maximizing token likelihood to directly optimizing internal attention distributions. Instead of just learning *what* to say, RAL treats the model's attention mechanism as a policy, explicitly teaching the model *where* to look within complex image or video inputs to derive the correct answer. The core technical innovation lies in how RAL formulates attention as a trainable policy using an advantage-weighted divergence objective. When the model produces a high-reward response, the algorithm minimizes the divergence between the current and past attention distributions, reinforcing the specific visual grounding patterns that led to success. Conversely, it penalizes attention patterns associated with low rewards. This method provides a more stable training signal than traditional token-level gradients, which often suffer from "reward hacking" where the model overfits to surface-level linguistic patterns rather than underlying logic. Additionally, the authors propose "On-Policy Attention Distillation," a novel distillation technique where a student model learns not just to mimic a teacher's output text, but to align its internal attention distribution with the teacher's, effectively inheriting the teacher's visual focus and reasoning structure. Empirically, RAL demonstrates consistent superiority over existing baselines like Group Relative Policy Optimization (GRPO) across diverse benchmarks, particularly in tasks requiring fine-grained visual search and long-video understanding. A striking discovery is the efficacy of "RAL-zero," a variant of the method where the explicit "thinking" process is removed entirely. Even without generating text-based rationales, RAL-zero achieves state-of-the-art performance on perception tasks by relying solely on optimized attention weights. This confirms the authors' hypothesis that directly supervising internal information allocation is a more principled and robust alternative to indirect supervision through textual outputs, paving the way for more grounded and perception-aware multimodal AI. Source: February 2026 Reinforced Attention Learning UC Davis, Princeton University, Google, Google DeepMind Bangzheng Li, Jianmo Ni, Chen Qu, Ian Miao, Liu Yang, Xingyu Fu, Muhao Chen, Derek Zhiyuan Cheng https://arxiv.org/pdf/2602.04884