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: Multimodal VLM
Keywords: Visual Perception, Staged Post-training, Capability-dimension Curriculum, RLVR, Visual Mathematical Reasoning

TL;DR

This paper argues that current VLM post-training overemphasizes "long-chain reasoning" while neglecting perception bottlenecks. It explicitly decouples post-training into three independent stages: "Visual Perception \(\rightarrow\) Textual Reasoning \(\rightarrow\) Visual Reasoning," using RLVR (instead of caption SFT) to specifically refine perception. This approach improves Qwen3-VL-8B by 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," co-optimizing tasks like VQA, geometry, chart understanding, and visual math in a single stage (e.g., Mixed Reward in LLaVA-CoT or VLAA-Thinker), hoping the model improves accuracy by "thinking longer."

Limitations of Prior Work: The authors performed an attribution analysis of Qwen3-VL-8B errors across three visual math datasets using Claude-Haiku-4.5. They found that 86.9% of incorrect answers originated from errors in the first step of image perception, rather than subsequent reasoning failures. Once perception fails, elongated chain-of-thought only "rationalizes" incorrect premises, often repeating the same erroneous reading even after re-examining the image.

Key Challenge: Merged training (mixing perception and reasoning data) assumes all capabilities can be optimized via the same reward signal. However, visual perception is a more fundamental capability than reasoning and requires specialized objectives and data. If reasoning is prioritized, the model develops a "long-chain + self-persuasion" behavioral pattern without aligning visual features, leading to a "perception tax" where reasoning-only training drops MMStar performance by 1.6%.

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) Unify "capability-dimension stages" and traditional "difficulty-increasing curriculum learning" into a single framework.

Key Insight: Post-training should be viewed as sequentially shaping three capabilities: visual perception, textual reasoning, and visual reasoning. If perception is the "scaffolding" for subsequent reasoning, it should be solidified before visual reasoning; otherwise, learning reasoning first may contaminate the learning signals for perception.

Core Idea: Organize post-training using a "capability-dimension curriculum" \(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\), utilizing 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 "perception-essential" QA from 15K DOCCI image-text pairs and filtering for samples that truly test perception via "Image vs. Caption" contrast; (2) Staged GRPO Training—sequentially running the same number of epochs for the three data categories in the order \(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\). Each stage uses identical GRPO hyperparameters with the vision encoder unfrozen 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 (Image Wrong ∧ Caption Right)<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 "Perception-Essential" QA to prevent language prior leakage

To train perception independently, the authors need perception questions that cannot be guessed via language priors. They use Qwen2.5-72B to generate perception-focused QA pairs \((Q,A)=f_{\text{gen}}(C)\) from fine-grained DOCCI captions. Each candidate is tested via two paths: providing only the image \(\hat{A}_{\text{img}}=f_\theta(I,Q)\) and providing only the caption \(\hat{A}_{\text{cap}}=f_\theta(C,Q)\). Samples are retained only if \(\mathbb{I}[\hat{A}_{\text{img}}\neq A]\land\mathbb{I}[\hat{A}_{\text{cap}}=A]\). This "reverse engineering" ensures that if a sample can be answered via caption, it tests language rather than vision; only samples where "the caption works but the image fails" expose true perception deficits.

2. Capability-Dimension Staged GRPO: Establishing "Seeing" before "Reasoning"

While merged training assumes a single reward can optimize all skills, perception is more foundational. Training reasoning first causes the model to rationalize mistakes. The authors use the 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 in the sequence \(\mathcal{D}_{\text{perc}} \rightarrow \mathcal{D}_{\text{text}} \rightarrow \mathcal{D}_{\text{vis}}\). The rewards \(R(x,y_i)=r_{\text{acc}}+r_{\text{format}}\) are consistent across stages. Total steps (930) are divided by epoch into 90 / 375 / 465 to match the merged baseline. The vision encoder remains unfrozen to prevent the refined visual representations from drifting in later stages. This "capability curriculum" is orthogonal to traditional difficulty-based curricula.

3. RLVR for Perception instead of Caption-based SFT: Avoiding low-quality caption contamination

Traditional SFT using full captions as targets often injects off-policy supervision that harms existing capabilities due to lower caption quality compared to pre-training data. Since perception QA naturally requires short answers (colors, spatial relations), the authors use exact match as \(r_{\text{acc}}\) for on-policy GRPO sampling. This keeps the model near its current policy. Experiments show that replacing RL with SFT in the perception stage results in performance drops 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. Steps for the three stages are 90 / 375 / 465 (totaling 930 to match the merged baseline). The vision encoder is frozen in the merged baseline but unfrozen in the staged approach. Training utilized 8×H200 GPUs with the EasyR1 framework and Claude-Haiku-4.5 as the judge.

Key Experimental Results

Main Results

Using Qwen3-VL-8B as the backbone compared to other reasoning VLMs (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

Compared to base: WeMath +5.24, RWQA +3.66, MMStar +3.07. Compared to the strong baseline OneThinker: WeMath +1.53, RWQA +3.01, MMStar +2.87.

Ablation Study: Training Paradigm and 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
Stage order 3→2→1 (Reasoning back to Perc)	37.70 (7B)	74.17 (7B)	55.93 (7B)	Drops 4.6% vs. 1→2→3
Perception via Caption SFT	-1.6 on WeMath	—	—	8B; -8.1% on 7B

Key Findings

• Perception errors are the true bottleneck: 86.9% of Qwen3-VL-8B errors stem from perception. Reasoning-only post-training triggers a "perception tax" (MMStar drops 1.6% on Qwen2.5-VL-7B), while adding perception data increases RWQA by 3.0%.
• Stage order is critical: Placing perception last (3→2→1) drops visual math from 42.3% to 37.7% on Qwen2.5-VL-7B, lower than the merged baseline. Early long-chain behaviors override the low-noise gradients of the perception stage.
• Shorter reasoning \(\neq\) Weaker reasoning: The staged model response length is 20.8% shorter than the merged version (445 vs. 562 tokens), yet visual math accuracy is 1.46% higher. Improved perception naturally streamlines the chain-of-thought.
• Curriculum dimensions are additive: Capability-dimension curriculum (staged) is orthogonal to difficulty-dimension curriculum. Combining both yields a +4.43% gain over the merged baseline.

Highlights & Insights

• Revisiting "Long-Chain CoT Universalism": A simple error attribution experiment (86.9% perception errors) challenges the assumption that "longer reasoning is always better." This "diagnose before prescribing" research paradigm is applicable across various model scaling scenarios.
• Differential Filtering for Perception Data: Using "Image Wrong ∧ Caption Right" as a selection criterion is a clever reverse-engineering method. It identifies perception blind spots without manual labeling, a logic applicable to audio, video, or 3D modalities.
• Capability-Dimension as a New Design Degree of Freedom: Traditional curricula focus on difficulty; this work introduces "capability" as an orthogonal dimension, opening a unified perspective for multi-stage post-training (e.g., decoupling RLHF, DPO, and RLVR).

Limitations & Future Work

• The fixed epoch/step ratio across stages may not be optimal; adaptive stopping or dynamic weights were not explored.
• Perception data is based on DOCCI natural image captions, offering limited coverage for charts, tables, or dense OCR (DocVQA gains were smaller).
• Validation was limited to Qwen and InternVL series (up to 32B); scaling laws for 70B+ models or non-Transformer-Decoder architectures remain untested.
• The judge model (Claude-Haiku-4.5) has a 82.5% perception-error labeling accuracy; the 86.9% bottleneck figure may be slightly over- or underestimated.

Related Work & Insights

• vs. Reasoning-only RL (OneThinker / WeThink): These treat perception as a pre-training byproduct. This paper proves that "explicit perception modeling" is more efficient than simply increasing reasoning RL data.
• vs. Difficulty Curriculum (Curr-ReFT / PC-GRPO): Traditional curricula order data by difficulty; this method orders by capability (perception \(\rightarrow\) reasoning). The two are complementary.
• vs. Perception Benchmarks (VisOnlyQA / NoReGeo): While those works "diagnose" the bottleneck, this paper provides the "prescription" through a specific data synthesis and training protocol.

Rating

• Novelty: ⭐⭐⭐⭐ Defines capability-dimension as a new orthogonal curriculum dimension.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Cross-backbone validation, 8 benchmarks, stage-order analysis, and SFT/RL comparisons.
• Writing Quality: ⭐⭐⭐⭐ Clear narrative flow (Diagnosis \(\rightarrow\) Decoupling \(\rightarrow\) Sequencing \(\rightarrow\) RLVR).
• Value: ⭐⭐⭐⭐ A low-cost, high-reward correction to current VLM post-training pipelines.

Related Papers

• [ICML 2026] Bad Seeing or Bad Thinking? Rewarding Perception for Vision-Language Reasoning
• [ICML 2026] From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning
• [CVPR 2026] Improving Vision-language Models with Perception-centric Process Reward Models
• [CVPR 2026] Seeing What Matters: A Training-Free Self-Guided Framework for Multimodal Detail Perception and Reasoning
• [CVPR 2026] Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design