From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models

Conference: ICML 2026
arXiv: 2605.20177
Code: https://ucsc-vlaa.github.io/VLM-CapCurriculum/ (Project Page)
Area: Multi-modal VLM
Keywords: Visual Perception, Staged Post-training, Capability-dimension Curriculum, RLVR, Visual Mathematical Reasoning

TL;DR

This paper identifies that current VLM post-training overemphasizes "long-chain reasoning" while neglecting perception bottlenecks. It explicitly decouples post-training into three independent stages: "Visual Perception → Textual Reasoning → Visual Reasoning," and utilizes RLVR (instead of caption SFT) to specifically refine perception. This approach enables Qwen3-VL-8B to achieve relative improvements of approximately +5.9% and +1.2% on visual math and perception benchmarks, respectively, while shortening reasoning traces by 20.8%.

Background & Motivation

Background: Following DeepSeek-R1, the mainstream paradigm for VLM post-training involves "long-chain CoT + RLVR," where tasks such as visual QA, geometry, chart understanding, and visual math are jointly optimized in a single stage (e.g., the Mixed Reward in LLaVA-CoT and VLAA-Thinker), aiming to improve accuracy by encouraging the model to "think longer."

Limitations of Prior Work: The authors performed an attribution analysis of errors made by Qwen3-VL-8B on three visual math datasets using Claude-Haiku-4.5. They found that 86.9% of incorrect answers originated from perception errors in the first step, rather than subsequent reasoning failures. Once perception fails, extended chain-of-thought only "self-justifies" based on incorrect premises, even when repeatedly re-examining the image.

Key Challenge: Merged training (mixing perception and reasoning data) assumes all capabilities can be optimized simultaneously by the same reward. However, visual perception is a more fundamental capability than "reasoning" and requires specific objectives and data. If reasoning is trained prematurely, the model adopts a "long-chain + self-persuasion" behavioral pattern without properly aligning visual features, leading to a "perception tax" where MMStar scores drop by 1.6% after reasoning-only training.

Goal: (1) Demonstrate that perception must be trained separately with specialized data; (2) Identify the optimal sequence of stages; (3) Compare whether caption-SFT or RLVR is more suitable for learning perception; (4) Integrate "stage-by-capability dimension" and traditional "incremental difficulty curriculum learning" into a unified framework.

Key Insight: Post-training is viewed as the sequential shaping of three capabilities: visual perception, textual reasoning, and visual reasoning. Since perception serves as the "scaffolding" for subsequent reasoning, it should be stabilized before visual reasoning is attempted; otherwise, early reasoning training may contaminate subsequent perception signals.

Core Idea: Post-training is organized using a "capability-dimension curriculum" \(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\), employing RLVR instead of caption SFT during the perception stage.

Method

Overall Architecture

The method consists of two main components: (1) Perception Data Synthesis—reverse-generating "must-see-to-answer" QA from 15K DOCCI image-text pairs and filtering samples using a "image vs. caption" differential; (2) Staged GRPO Training—sequentially training on three types of data (\(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\)) for the same number of epochs. Each stage uses the same GRPO hyperparameters with the visual encoder enabled throughout. The result is a VLM that is stronger in both visual math and perception with shorter reasoning traces.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    subgraph SYN["Perception Data Synthesis + Differential Filtering (Design 1)"]
        direction TB
        A["DOCCI 15K Image-Text Pairs"] --> B["Qwen2.5-72B generates perception QA<br/>from captions"]
        B --> C["Differential Filtering: Keep only cases where<br/>Image fails ∧ Caption succeeds<br/>(Intersection of Qwen2.5-VL 7B / 32B)"]
    end
    C --> D["Perception Data D_perc"]
    E["Textual Reasoning D_text<br/>(ORZ-Math-13k)"]
    F["Visual Reasoning D_vis<br/>(CLEVR-Math / GeoQA, etc.)"]
    subgraph STAGE["Capability-dimension Staged GRPO (Design 2)"]
        direction TB
        G["Stage 1: Perception<br/>Using RLVR instead of SFT (Design 3)"] --> H["Stage 2: Textual Reasoning"]
        H --> I["Stage 3: Visual Reasoning"]
    end
    D --> G
    E --> H
    F --> I
    I --> J["Post-trained VLM: Perception↑ / Reasoning↑<br/>Reasoning trace shortened by 20.8%"]

Key Designs

1. Perception Data Synthesis + Dual-model Differential Filtering: Creating "must-see-to-answer" QA to block language priors.

To train perception in isolation, a set of samples with pure signals that cannot be guessed via language is required. The authors first use Qwen2.5-72B to generate perception-focused QA pairs \((Q,A)=f_{\text{gen}}(C)\) from fine-grained DOCCI captions. Then, two paths are run for each candidate: \(\hat{A}_{\text{img}}=f_\theta(I,Q)\) (image only) and \(\hat{A}_{\text{cap}}=f_\theta(C,Q)\) (caption only). Only samples satisfying \(\mathbb{I}[\hat{A}_{\text{img}}\neq A]\land\mathbb{I}[\hat{A}_{\text{cap}}=A]\) are retained. Finally, an intersection is taken after filtering with Qwen2.5-VL-7B and 32B base models. This reverse engineering ensures that samples solvable by caption alone (testing language) are removed, leaving only those that expose perception deficits.

2. Capability-dimension Staged GRPO: Establishing "seeing clearly" before adding "reasoning."

Merged training assumes all capabilities optimize via a single reward, but perception is more fundamental. Training reasoning first causes "self-persuasion" without visual alignment. Using the same GRPO loss:

\[\mathcal{J}_{\text{GRPO}}(\theta)=\mathbb{E}_{x,y}\Big[\tfrac{1}{G}\sum_i \min(\rho_i A_i, \text{clip}(\rho_i,1-\epsilon,1+\epsilon)A_i)\Big] - \beta\, \text{KL}(\pi_\theta\|\pi_{\text{ref}}),\]

training proceeds in the order \(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\). The reward \(R(x,y_i)=r_{\text{acc}}+r_{\text{format}}\) is consistent across stages. Steps are allocated as 90 / 375 / 465 to match the 930 steps of the merged baseline. The visual encoder remains unfrozen to prevent the drift of visual representations. This "capability-dimension curriculum" is orthogonal to traditional difficulty curricula.

3. Perception Stage using RLVR instead of caption-based SFT: Preventing low-quality captions from polluting perception signals.

Traditional SFT using full captions as targets often introduces off-policy supervision from captions that are lower quality than pre-training data, which can degrade existing capabilities. The authors note that perception QA answers are naturally short (colors, spatial relations, attributes). Thus, exact match can be used as \(r_{\text{acc}}\), allowing seamless integration with GRPO's on-policy sampling. Comparative experiments show that replacing RL with SFT in the perception stage results in a drop of 8.1% on WeMath for Qwen2.5-VL-7B and 1.6% for Qwen3-VL-8B.

Loss & Training

GRPO group size \(G=5\), max response length 2048. Three-stage steps: 90 / 375 / 465. The visual encoder is frozen in the merged baseline but fully open in the staged approach. Training used 8×H200 with the EasyR1 framework and Claude-Haiku-4.5 as the judge.

Key Experimental Results

Main Results

Comparison of Qwen3-VL-8B with reasoning-VLMs of similar scale (Acc %):

Model	MathVista	WeMath	RWQA	MMStar	Overall AVG
Qwen3-VL-8B (Base)	72.40	50.86	70.85	70.00	62.19
OneThinker-8B	75.10	54.57	71.50	70.20	64.87
Qwen3-VL-8B (Staged, Ours)	75.90	56.10	74.51	73.07	65.77

Relative to the base model: WeMath +5.24, RWQA +3.66, MMStar +3.07. Relative to OneThinker: WeMath +1.53, RWQA +3.01, MMStar +2.87.

Ablation Study: Training Paradigm + Stage Order

Configuration (Qwen3-VL-8B)	Vis Math AVG	Perception AVG	Overall	Note
Base	45.17	79.21	62.19	No post-training
Merged	49.64	79.71	64.67	Standard baseline
Staged 1→2→3 (Perc→Text→Vis)	51.10	80.44	65.77	Default
Order 3→2→1 (Reverse)	37.70 (7B)	74.17 (7B)	55.93 (7B)	-4.6 pts vs normal order
Perception Stage as SFT	-1.6 on WeMath	—	—	8B; -8.1% on 7B

Key Findings

• Perception errors are the real bottleneck: 86.9% of failures are due to perception. Reasoning-only training triggers a "perception tax" (MMStar -1.6% for 7B), while adding perception data improves RWQA by 3.0%.
• Stage order is critical: Reversing the order (3→2→1) drops visual math performance below the merged baseline, suggesting perception gradients are overwritten by established long-chain behaviors if trained last.
• Shorter reasoning ≠ Weaker reasoning: The staged model's responses are 20.8% shorter than the merged model (445 vs 562 tokens), yet it achieves higher accuracy (+1.46 pts in visual math), with perception errors dropping from 805 to 781.
• Curriculum dimensions are stackable: Combining the capability-dimension curriculum with a difficulty curriculum yields a +4.43% improvement over the merged baseline.

Highlights & Insights

• De-mystifying "Long-chain CoT Universalism": A simple error attribution experiment (86.9% perception errors) challenges the assumption that "longer reasoning is always better," offering a "diagnosis before prescription" paradigm.
• Differential Filtering for Data Synthesis: The "fails on image ∧ succeeds on caption" criterion cleverly identifies perception blind spots without manual labeling.
• New Design Freedom: Capability serves as an orthogonal dimension to difficulty in curriculum design, offering a unified perspective for multi-stage post-training.

Limitations & Future Work

• Fixed step ratios between stages may not be optimal; adaptive stopping was not explored.
• Perception data relies on natural image captions, offering limited coverage for charts, tables, or dense OCR.
• Evaluations utilized Qwen and InternVL series only; scaling laws for 70B+ models remain unverified.
• The judge model (Claude-Haiku-4.5) has a perception-error annotation accuracy of 82.5%, potentially introducing systematic bias in the 86.9% attribution figure.

Related Work & Insights

• vs. Reasoning-only RL (OneThinker/WeThink): These assume perception is a pre-training byproduct. This work shows explicitly modeling perception is more efficient.
• vs. Difficulty Curriculum (Curr-ReFT/PC-GRPO): This work adds a "capability" dimension that is orthogonal and stackable.
• vs. Perception Benchmarks (VisOnlyQA): While prior work "diagnoses" the bottleneck, this work provides the "treatment" through specific data synthesis and training protocols.

Rating

• Novelty: ⭐⭐⭐⭐
• Experimental Thoroughness: ⭐⭐⭐⭐⭐
• Writing Quality: ⭐⭐⭐⭐
• Value: ⭐⭐⭐⭐

Related Papers

• [ICML 2026] Bad Seeing or Bad Thinking? Rewarding Perception for Vision-Language Reasoning
• [ICML 2026] Med-Scout: Curing MLLMs' Geometric Blindness in Medical Perception via Geometry-Aware RL Post-Training
• [ICML 2026] Efficient Reasoning with Hidden Thinking
• [ICML 2026] Focusing Where Vision Matters: Selective Training for Large Vision Language Models via Visual Information Gain
• [ICML 2026] Learn to Think: Improving Multimodal Reasoning through Vision-Aware Self-Improvement Training