Research map / Vision-language and multimodal / Multimodal LLMs

PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning

CVPR 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 PDCR, which splits V-L reasoning steps into visual-perception and textual-reasoning clusters via a white-image Visual Dependence Score and Otsu thresholding, then normalizes confidence-gain advantages within each cluster, improving GRPO/PACR on Qwen2.5-VL.

Problem

Naively applying a globally normalized ground-truth confidence-growth process reward to vision-language reasoning distorts the signal for sparse visual perception steps, since the global statistics are dominated by dense textual reasoning steps.

Approach

Decompose reasoning steps without external labels using a Visual Dependence Score (log-likelihood ratio of a step under the real image vs. a white image), cluster steps with Otsu's method, and compute min-max normalized confidence-gain advantages separately within each cluster.

Key results

• On Qwen2.5-VL-7B, PDCR averages 52.9 across seven benchmarks vs. 52.2 for PACR, 52.0 for DAPO and 51.5 for GRPO.
• On Qwen2.5-VL-3B, PDCR averages 45.2 vs. 44.4 for PACR and 43.6 for GRPO.
• Random decomposition on 7B drops average to 52.3, showing the gain depends on the Visual Dependence Score clustering; Otsu clustering reaches 76.2% agreement with GPT-5 labels vs. 67.5% for best Top-K.

Surprising findings

• Rotating the image by 90 degrees gives poor skill separation (65.9%), suggesting VLMs are fairly rotation-invariant for this signal.
• Random decomposition with the same decomposed advantage barely improves over PACR, so the gain depends on correct clustering.

Limitations

• Higher training compute per step than GRPO due to extra forward passes for confidence gains and visual dependence scores.
• Binary hard clustering may misassign hybrid perception-reasoning steps.
• Skill labels for validation were produced by GPT-5 and checked on only 100 steps by humans.

Experimental setup

Datasets: Vision-SR1-47K GPT-5-annotated skill labels on 10,000 Qwen2.5-VL-7B trajectories

Benchmarks: MMMU-Pro MMMU RealWorldQA VisNumBench MathVerse MathVision HallusionBench

Metrics: accuracy average accuracy skill classification accuracy Cohen's kappa

Baselines: GRPO DAPO PACR Visionary-R1 Perception-R1 Vision-R1 Random Decomposition

Builds on: GRPO PACR DAPO Otsu's method EasyR1

Closest papers by meaning

• Improving Vision-language Models with Perception-centric Process Reward Models (CVPR 2026 · similarity 0.92)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.92)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.91)
• SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.90)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.90)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.90)