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

TL;DR

This work tackles the brittleness of multimodal LLMs in long-chain visual reasoning by separating abstract reasoning from visual grounding. It introduces a two-stage curriculum: a text-only disentangled SFT to build a robust reasoning backbone, followed by PG-CoT anchoring to vision, and an RL phase with a Pivotal Perception Reward to learn when to look. The approach yields state-of-the-art open-source performance on several 7B benchmarks and demonstrates strong improvements in perception-grounded reasoning, supported by cognitive-behavior analyses and comprehensive ablations. Overall, the framework provides a principled pathway to mitigate visual forgetting and develop a strategic, grounded multimodal reasoner with a dynamic perception policy.

Abstract

Multimodal Large Language Models (MLLMs) demonstrate significant potential but remain brittle in complex, long-chain visual reasoning 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 creates a foundational cold-start deficiency -- weakening abstract reasoning -- and a strategic perception deficit, as models lack a policy for when to perceive. 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 builds 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 resolve 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 learning an autonomous grounding policy. Our contributions include the formalization of these two deficiencies and the development of a principled, two-stage framework to address them, transforming the model from a heuristic-driven observer to a strategic, grounded reasoner. \textbf{Code}: \url{https://github.com/gaozilve-max/learning-when-to-look}.

Figures (6)

• Figure 4: Overview of our curriculum-based framework for strategic perception in multimodal reasoning. Stage 1 (left) employs a disentangled SFT curriculum: Phase 1 conducts text-only SFT to build abstract reasoning capabilities (cognitive warm-up), and Phase 2 anchors these behaviors to visual evidence via Perception-Grounded Chain-of-Thought (PG-CoT) data. Stage 2 (right) learns autonomous perception timing through RL using a composite reward that includes our novel Pivotal Perception Reward, which couples perception actions to linguistic markers of cognitive uncertainty (e.g., "wait", "verify", "first").
• Figure 5: Data generation pipeline for Perception-Grounded Chain-of-Thought (PG-CoT).Using a teacher MLLM (e.g., GPT-4o), we transform standard CoT reasoning traces by identifying logical breakpoints where visual verification is required (marked as "Recommend Perception Location") and inserting fine-grained observation segments.
• Figure 6: Cognitive behavior comparison.
• Figure :
• Figure :