Learning When to Look: A Disentangled Curriculum for Strategic Perception in Multimodal Reasoning

Multimodal Large Language Models (MLLMs) have shown remarkable potential in complex reasoning but remain notoriously brittle during long-chain visual tasks. A critical failure mode is "visual forgetting", where models progressively lose visual grounding as reasoning extends, a phenomenon aptly described as "think longer, see less". We posit this failure stems from current training paradigms prematurely entangling two distinct cognitive skills: (1) abstract logical reasoning ("how-to-think") and (2) strategic visual perception ("when-to-look"). This entanglement leads to a foundational cold-start deficiency and a strategic perception deficit, as models lack an autonomous policy for perception timing. In this paper, we propose a novel curriculum-based framework to disentangle these skills. First, we introduce a disentangled Supervised Fine-Tuning (SFT) curriculum that establishes a robust abstract reasoning backbone on text-only data before anchoring it to vision with a novel Perception-Grounded Chain-of-Thought (PG-CoT) paradigm. Second, we address the strategic perception deficit by formulating timing as a reinforcement learning problem. We design a Pivotal Perception Reward that teaches the model when to look by coupling perceptual actions to linguistic markers of cognitive uncertainty (e.g., "wait", "verify"), thereby cultivating an autonomous grounding policy. Our contributions include the formalization of these two deficiencies and a principled two-stage framework that transforms the model from a heuristic-driven observer into a strategic, grounded reasoner. Project: https://github.com/gaozilve- max/learning-when-to-look.