Research map / Vision-language and multimodal / Multimodal LLMs

Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoCA, an RL framework that externalizes VLM perception into interleaved recognition blocks, rewards perception via a text-only "blindfolded reasoner" proxy, and uses structured verbal verification for outcome rewards, routing credit to "bad seeing" or "bad thinking" failures.

Problem

Multimodal RL for VLMs gives only outcome supervision, so when an answer is wrong it is unclear whether perception or reasoning failed; this causes a "seesaw effect" where perception and reasoning gains trade off, and agentic tool-use approaches are costly.

Approach

Decompose generation into <recognition> (perception) and <think> (reasoning) blocks; reward perception by whether a text-only strong LLM can answer the question from the perception text alone without the image (Perception Verification); judge free-form answers by having an LLM execute a fixed verification protocol (Structured Verbal Verification); and in GRPO, dampen penalties on perception tokens

Key results

• MoCA (7B) averages 60.5 across nine benchmarks vs 51.9 for Qwen2.5-VL-7B base, with 86.6 on V* and 58.3 on SlideVQA.
• Ablation: removing PV (outcome reward only) drops V* by 6.9 points (79.7 vs 86.6); removing MoCA gating lowers perception and rich-modality scores by about 2-3 points.
• The PV oracle agrees with human majority vote at 86.31% accuracy (Cohen's kappa 0.707) on 979 samples; the VP verifier reaches 91.9% accuracy and 92.3% consistency on a 273-pair challenge set vs 78.6% for LLM prompting.

Surprising findings

• Merely prompting for the decomposed perception-reasoning format without RL drops performance below the base model (Instruction-Only: V* 68.3 vs 71.4 base).
• Plain LLM-judge rewards caused reward hacking and less stable RL training, performing worse than the rule-free w/o PV variant.

Limitations

• PV oracle has 9.19% false negatives relative to human judgment, so perception rewards can be noisy.
• Perception reward is binary and depends on a 14B text-only proxy; blindfolded test assumes perception text suffices for reasoning, which may not hold for visual patterns like mazes.
• Gains are reported on a single 7B base model and hyperparameters were tuned on 3B models; code and data are promised only upon acceptance.

Experimental setup

Datasets: ViRL39K VisualWebInstruct-Verified Pixel Reasoner dataset arXiv/Fin-Sheet/News-Layout/Info-Graphic/Manuals modality-rich collection VP-Challenge-Set

Benchmarks: V* HRBench InfoVQA DUDE SlideVQA MMLongBench-Doc MMMU EMMA MathVista

Metrics: accuracy accuracy vs human labels F1 consistency across 5 runs Cohen's kappa

Baselines: GPT-4o GPT-4o-mini Qwen2.5-VL-Instruct-7B Qwen2.5-VL-Instruct-72B LLaVA-OV VL-Rethinker R1-VL Pixel Reasoner DeepEyes Docopilot mPLUG-Owl3 MiMo-VL

Builds on: GRPO Qwen2.5-VL Pixel Reasoner VL-Rethinker information bottleneck

Closest papers by meaning

• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.93)
• PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning (CVPR 2026 · similarity 0.92)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.92)
• Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning (ICLR 2026 · similarity 0.91)
• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.91)
• Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.91)