Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning

Achieving robust perception-reasoning synergy is a central goal for advanced Vision-Language Models (VLMs). Recent advancements have pursued this goal via architectural designs or agentic workflows. However, these approaches are often limited by static textual reasoning or complicated by the significant compute and engineering burden of external agentic complexity. Worse, this heavy investment does not yield proportional gains, often witnessing a "seesaw effect" on perception and reasoning. This motivates a fundamental rethinking of the true bottleneck. In this paper, we argue that the root cause of this trade-off is an ambiguity in modality credit assignment: when a VLM fails, is it due to flawed perception ("bad seeing") or flawed logic ("bad thinking")? To resolve this, we introduce a reinforcement learning framework that improves perception-reasoning synergy by reliably rewarding the perception fidelity. We explicitly decompose the generation process into interleaved perception and reasoning steps. This decoupling enables targeted supervision on perception. Crucially, we introduce Perception Verification (PV), leveraging a "blindfolded reasoning" proxy to reward perceptual fidelity independently of reasoning outcomes. Furthermore, to scale training across free-form VL tasks, we propose Structured Verbal Verification, which replaces high-variance LLM judging with structured algorithmic execution. These techniques are integrated into a Modality-Aware Credit Assignment (MoCA) mechanism, which routes rewards to the specific source of error -- either bad seeing or bad thinking -- enabling a single VLM to achieve simultaneous performance gains across a wide task spectrum.

When a vision-language AI model gets a question wrong about an image, we don't know if it saw the wrong thing or thought about it incorrectly — much like a doctor misreading an X-ray versus misinterpreting a correctly-read one. This "bad seeing vs. bad thinking" ambiguity means current training methods cannot target the actual source of error, causing a frustrating tradeoff: improving visual skills degrades reasoning, and vice versa. We solve this by splitting the model's response into explicit perception steps (what it sees) and reasoning steps (what it concludes). To evaluate perception independently, we introduce a "blindfolded reasoner" test: we feed the model's visual descriptions — without the image — to a text-only AI. If it can answer correctly from those descriptions alone, the perception was accurate; if not, the model saw poorly. We combine this with a new structured verification method for evaluating final answers, creating a credit assignment system that precisely blames either "bad seeing" or "bad thinking." Our approach, MoCA, is the first to break the perception-reasoning tradeoff. A single 7B-parameter model simultaneously improves on perception-heavy, reasoning-heavy, and document understanding tasks — even surpassing GPT-4o on several benchmarks — without the computational cost of multi-turn agentic systems.