When Prompts Become Pixels: Prompt-Region Grounding for Multimodal Reasoning

# Prompt-Region Grounding for Multimodal Reasoning
Source: [https://arxiv.org/html/2608.04726](https://arxiv.org/html/2608.04726)
## When Prompts Become Pixels: Prompt\-Region Grounding for Multimodal Reasoning

Yongxin Wang1, Ruizhe Zhou2, Yueling Tang2, Yingying Zhu3, Xuemin Zhao3, Xiaojun Chang1,4, Xiaodan Liang1,2

###### Abstract

Multimodal large language models increasingly reason over screenshots and documents where the task itself may be written in pixels\. Yet benchmarks usually place questions in text, leaving it unclear whether models use the same instruction equally well across channels\. We introduce Visualized Task Semantics \(VTS\), a controlled intervention that moves the question into the image while keeping the source problem and answer fixed\. Across six MLLMs and four benchmarks, accuracy drops in all 24 model\-task pairs, by 17\.8 points on average\. Models often transcribe the visual question correctly yet fail to use it, exposing a semantic channel gap beyond OCR\. To reduce this gap, we present prompt\-region grounding, whose core design aligns the question region with typed semantics and recovers its clean representation from a masked view\. At matched training cost, our method raises four\-benchmark VTS accuracy from 58\.0 to 66\.3 while preserving accuracy on the original interface, and requires no OCR or region metadata at inference\. Reading task\-bearing text and grounding it as an instruction for reasoning are distinct capabilities\.

## 1Introduction

Visual text has become a standard input to multimodal large language models \(MLLMs\)\. High\-resolution models can read screenshots, documents, and other text\-rich images with increasing accuracy\(Wang et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib25); Bai et al\.[2025b](https://arxiv.org/html/2608.04726#bib.bib4); Wu et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib31)\)\. Most work treats this text as evidence to extract: a label, value, or paragraph contributes content to the answer\. A visual question has a different function\. It determines how the evidence should be used and what answer is required\. Whether MLLMs preserve this task\-defining function when a question moves from prompt tokens into image pixels remains less understood\.

Standard reasoning protocols make this function difficult to study\. They place the question in the language channel and reserve the image for visual evidence\. Prior work reports weaker performance when instructions are visualized\(Li et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib15); An et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib1); Liu et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib18)\), but the measured gap can mix several effects: canvas expansion, image resizing, renderer errors, missing prompt content, and the change of semantic channel itself\. A model may also transcribe the visual question correctly without using it to control the answer\. An accuracy drop alone therefore cannot tell us whether the model failed to read the question or failed to use it\.

We study this problem with*Visualized Task Semantics*\(VTS\), a paired intervention that changes where the question is presented\. Given a benchmark item with questionqq, imagexx, and answeraa, the Original view presents\(q,x\)\(q,x\)\. VTS rendersqqabovexxand replaces the typed question with a fixed cue\. The source problem, visual evidence, and target answer remain paired within each item\. VTS then adds separate controls for canvas geometry, prompt duplication, and missing task content\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x1.png)Figure 1:The task stays fixed, but the answer changes when its question becomes pixels\. VTS renders the question inside the image and replaces the native prompt with a fixed cue\. The upper schematic omits this separate cue for space; the lower example displays it\. The source visual evidence and target answer are unchanged within each pair\.VTS reveals a systematic semantic channel gap\. Across six MLLMs and four benchmarks, accuracy falls in all 24 model\-task pairs, by 17\.8 points on average and by as much as 28\.0 points\. A blank\-canvas control remains within 1\.1 points of Original accuracy, and a duplicate\-question control remains within 0\.6 points\. Removing the question from both channels causes accuracy to collapse\. The loss therefore cannot be explained by the added canvas, image resizing, or renderer corruption alone\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x2.png)Figure 2:Every evaluated model is weaker when the question is in the image\. Bars show mean accuracy over MATH\-Vision, MathVista, ChartQA, and MMMU\. Both interfaces are shown because a small gap can also occur when performance is low\.We next test whether the remaining gap reduces to text recognition\. In a two\-task diagnostic, the base model transcribes 87\.6% of visual questions exactly but answers only 48\.4% correctly\. Reinserting the same model’s transcript into the text channel raises answer accuracy by 7\.7 points while leaving the composite unchanged\. A substantial part of the gap therefore appears after transcription: the model recovers the words, but those words exert less control as pixels than as prompt tokens\.

This finding motivates*prompt\-region grounding*\. Mixed\-interface replay presents questions in both channels, but it does not explicitly connect the visual question region to its typed counterpart\. Our method adds two region\-level objectives\. PVRD\-SG aligns the prompt region with a frozen representation of the typed question\. PRMLP masks part of that region and recovers the representation of its clean crop\. Component controls test both parts of this design: prompt\-panel readouts outperform full\-image and random\-region summaries, while matched targets outperform mismatched targets\. Region boxes and crops are used only during training\. At inference, the model receives one composite and answers directly, without OCR or localization metadata\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x3.png)Figure 3:Prompt\-region grounding\. Paired replay exposes the model to Original and VTS views with the same answer target\. PVRD\-SG connects the visual question region to its typed semantics, and PRMLP reconstructs a clean latent target from a masked prompt region\. GSPO then continues the grounded checkpoint using verifiable format and answer rewards\.At matched training cost on Qwen3\-VL\-4B, ordinary SFT reaches 58\.0 VTS accuracy and 69\.1 Original accuracy across four benchmarks\. Prompt\-region grounding raises these scores to 66\.3 and 70\.3, with a VTS gain on every task\. The same recipe improves a second backbone\. It also gains 4\.1 points on the independently constructed VISTA\-Bench protocol and 7\.9 points on 1,000 held\-out real\-world pages\.

VTS and prompt\-region grounding distinguish two capabilities that MLLM evaluations often treat as the same: reading text in an image and grounding that text as the task instruction\. VTS measures the distinction under paired problems, and the region\-level objectives train for task use without changing the inference interface\. The results show that the channel carrying a question can alter reasoning even when its wording and answer remain fixed\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x4.png)Figure 4:Training progressively closes the semantic channel gap\. Lines show four\-benchmark mean accuracy\. Vertical connectors mark the Original minus VTS gap at each stage: 15\.8, 8\.4, 4\.0, and 1\.5 points\.
## 2Related work

Multimodal models and reasoning\.High\-resolution MLLMs improve perception in images and documents\(Wang et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib25); Bai et al\.[2025b](https://arxiv.org/html/2608.04726#bib.bib4); Wu et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib31); Li et al\.[2024a](https://arxiv.org/html/2608.04726#bib.bib14)\)\. Other work improves multimodal reasoning through post\-training or by varying diagrams, problem instances, and image count\(Zhu et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib39); Kimi Team[2025](https://arxiv.org/html/2608.04726#bib.bib12); Zhang et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib34); Zou et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib40); Wang et al\.[2025a](https://arxiv.org/html/2608.04726#bib.bib26)\)\. Our study holds the source problem fixed and asks how training should handle a change in the interface carrying its question\.

Recent work trains multimodal reasoning with process rewards, contrastive reflection, and verifiable rewards\(Wang et al\.[2025b](https://arxiv.org/html/2608.04726#bib.bib27),[d](https://arxiv.org/html/2608.04726#bib.bib29); DeepSeek\-AI et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib7)\)\. We instead change where the task appears while keeping the answer fixed\. We apply GSPO after the supervised model\. The combination of PVRD\-SG and PRMLP is the supervised method contribution; GSPO is the final reinforcement\-learning stage\.

Visual prompting and questions\.VIM introduces visual\-modality instructions, measures the text\-to\-pixel gap across eight benchmarks, and trains on text and pixel interfaces\(Li et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib15)\)\. VoQA places the scene and question in one image and reconstructs the visual question before answering during supervised fine\-tuning\(An et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib1)\)\. VISTA\-Bench compares matched text and visualized\-text questions across renderers and OCR systems\(Liu et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib18)\)\. These works establish the problem; we do not claim that moving a question into pixels is new\.

VIM\-style mixed replay is therefore a central baseline rather than an omitted alternative\. VoQA is the closest reconstruction\-based method: it generates an intermediate question target before answering\. We instead test whether region\-level training can improve direct answering without an output\-side transcription stage\. VTS supplies source\-matched reasoning pairs for this comparison\. OCRBench v2 and MMLongBench\-Doc cover broader visual\-text localization and document layouts\(Fu et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib10); Ma et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib20)\)\. VTCBench studies a related modality gap for long text compressed into dense images\(Zhao et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib35)\)\.

Alignment and consistency learning\.RegionCLIP learns open\-vocabulary links between image regions and text\(Zhong et al\.[2022](https://arxiv.org/html/2608.04726#bib.bib38)\)\. Align\-KD matches modalities to compress MLLMs, while Align\-TI uses a teacher to transfer instruction\-relevant visual interactions\(Feng et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib9); Chen et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib5)\)\. These methods do not target printed questions, but they provide close precedents for cross\-modal distillation\. Masked representation learning provides a second precedent\. I\-JEPA predicts masked image regions\. BYOL updates a stop\-gradient target with an exponential moving average, whereas SimSiam stops the target gradient without that average\(Assran et al\.[2023](https://arxiv.org/html/2608.04726#bib.bib2); Grill et al\.[2020](https://arxiv.org/html/2608.04726#bib.bib11); Chen and He[2021](https://arxiv.org/html/2608.04726#bib.bib6)\)\. PRMLP instead predicts a detached clean\-crop representation from a composite whose prompt region has been masked\. The reported implementation uses an identity predictor\.

Paired replay teaches both interfaces but leaves their relationship implicit\. Prompt\-region grounding makes that relationship an explicit training signal\. It adds two losses on VTS and real\-world visual\-prompt examples:*prompt\-visual representation distillation with semantic grounding*\(PVRD\-SG\) connects the visual question region to a frozen representation of its typed semantics, and*prompt\-region masked latent prediction*\(PRMLP\) connects a masked question region to its clean visual representation\. Both losses share the original answer target and leave the inference interface unchanged\.

### Visualized task semantics

Let\(qi,xi,ai\)\(q\_\{i\},x\_\{i\},a\_\{i\}\)denote a question, its visual context, and its canonical answer\. The original benchmark view is

uiorig=concat⁡\(qi,xi\)\.u\_\{i\}^\{\\mathrm\{orig\}\}=\\operatorname\{concat\}\(q\_\{i\},x\_\{i\}\)\.\(1\)The model receivesqiq\_\{i\}as language tokens andxix\_\{i\}through the vision encoder\. VTS movesqiq\_\{i\}into the image\. A deterministic rendererrender\\operatorname\{render\}normalizes lightweight LaTeX, wraps the question to the image width, and draws it in a white panel above the source image:

x~i=render⁡\(qi,xi\),uivts=concat⁡\(qmin,x~i\)\.\\tilde\{x\}\_\{i\}=\\operatorname\{render\}\(q\_\{i\},x\_\{i\}\),\\qquad u\_\{i\}^\{\\mathrm\{vts\}\}=\\operatorname\{concat\}\(q\_\{\\min\},\\tilde\{x\}\_\{i\}\)\.\(2\)Hereqminq\_\{\\min\}is the fixed cue “Help me solve the problem”\. The renderer records the panel boxbib\_\{i\}and saves a clean cropcic\_\{i\}\. Both views use the same answeraia\_\{i\}\.

VTS changes both the question channel and the remaining text prompt\. We use three controls to interpret this change\. Canvas presentsconcat⁡\(qi,render⁡\(∅,xi\)\)\\operatorname\{concat\}\(q\_\{i\},\\operatorname\{render\}\(\\varnothing,x\_\{i\}\)\): the native question plus the same padded canvas with a blank panel\. Duplicate presentsconcat⁡\(qi,render⁡\(qi,xi\)\)\\operatorname\{concat\}\(q\_\{i\},\\operatorname\{render\}\(q\_\{i\},x\_\{i\}\)\), placing the question in both channels\. Image\-only presentsconcat⁡\(qmin,render⁡\(∅,xi\)\)\\operatorname\{concat\}\(q\_\{\\min\},\\operatorname\{render\}\(\\varnothing,x\_\{i\}\)\), so neither channel contains the task wording\. Renderer template, source\-image placement, decoding, and scorer otherwise remain fixed\. The controlled training subset uses only the Original and VTS views\.

During supervised training, the VTS renderer suppliesbib\_\{i\}andcic\_\{i\}; the GLM\-OCR pipeline supplies the corresponding region and crop for real\-world rows\(Duan et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib8)\)\. At VTS inference, the model receives only the image andqminq\_\{\\min\}\. No evaluation supplies a box, crop, OCR transcript, or other localization metadata\. The method uses known regions as supervision, but it is not an explicit task\-region detector\.

### Balanced replay

Following VIM’s cross\-interface mixture\(Li et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib15)\), the controlled subset requires one Original and one VTS row for every selected source item:

𝒟pair\\displaystyle\\mathcal\{D\}\_\{\\mathrm\{pair\}\}=\{\(uiorig,yi\)\}i∈ℐ∪\{\(uivts,yi\)\}i∈ℐ\.\\displaystyle=\\left\\\{\(u\_\{i\}^\{\\mathrm\{orig\}\},y\_\{i\}\)\\right\\\}\_\{i\\in\\mathcal\{I\}\}\\cup\\left\\\{\(u\_\{i\}^\{\\mathrm\{vts\}\},y\_\{i\}\)\\right\\\}\_\{i\\in\\mathcal\{I\}\}\.\(3\)The complete supervised pool is𝒟sft=𝒟pair∪𝒟rw\\mathcal\{D\}\_\{\\mathrm\{sft\}\}=\\mathcal\{D\}\_\{\\mathrm\{pair\}\}\\cup\\mathcal\{D\}\_\{\\mathrm\{rw\}\}, where𝒟rw\\mathcal\{D\}\_\{\\mathrm\{rw\}\}pairs held\-in real\-world images with their original prompts and extracted regions\. Letℐvis\\mathcal\{I\}\_\{\\mathrm\{vis\}\}index the VTS and real\-world visual\-prompt rows; each stores\(qi,bi,ci\)\(q\_\{i\},b\_\{i\},c\_\{i\}\)\. We verify the one\-to\-one construction within𝒟pair\\mathcal\{D\}\_\{\\mathrm\{pair\}\}using stable pair IDs and target and answer hashes\. Both views use the same assistant target:

yi=\{sans​\(ai\),answer\-only,strace​\(ri\)∥sans​\(ai\),trace\-bearing\.y\_\{i\}=\\begin\{cases\}s\_\{\\mathrm\{ans\}\}\(a\_\{i\}\),&\\text\{answer\-only\},\\\\ s\_\{\\mathrm\{trace\}\}\(r\_\{i\}\)\\mathbin\{\\\|\}s\_\{\\mathrm\{ans\}\}\(a\_\{i\}\),&\\text\{trace\-bearing\}\.\\end\{cases\}\(4\)Hererir\_\{i\}is a question\-scrubbed reasoning trace andaia\_\{i\}is the canonical answer\. The fixed serializers add the corresponding<think\>and<answer\>tags\. The target policy is frozen before training and shared by both views\. Trace\-bearing targets omit verbatim reconstructions ofqiq\_\{i\}; otherwise the run uses the answer\-only target\. The supervised loss is autoregressive negative log\-likelihood over assistant tokens:

ℒsft​\(θ\)=𝔼\(u,y\)∼𝒟sft​ℓNLL​\(θ;u,y\),\\mathcal\{L\}\_\{\\mathrm\{sft\}\}\(\\theta\)=\\mathbb\{E\}\_\{\(u,y\)\\sim\\mathcal\{D\}\_\{\\mathrm\{sft\}\}\}\\ell\_\{\\mathrm\{NLL\}\}\(\\theta;u,y\),\(5\)whereℓNLL\\ell\_\{\\mathrm\{NLL\}\}is the standard autoregressive loss on assistant\-target positions\.

### PVRD\-SG: Prompt\-region semantic grounding

Balanced replay exposes the model to both interfaces but does not explicitly tie the rendered question to its typed counterpart\. PVRD\-SG places this constraint on the prompt region rather than the whole image\. The distinction is important because a VTS composite contains two kinds of information: the question specifies the task, while the source image supplies the evidence\. A whole\-image summary mixes these roles and may be dominated by objects, charts, or diagrams outside the question panel\. Reading only from the recorded prompt box asks a narrower question: does the region that contains the instruction represent the same task as its typed counterpart? LetPitextP\_\{i\}^\{\\mathrm\{text\}\}be a fixed text\-only template forqiq\_\{i\}, and letQiQ\_\{i\}contain its question\-token positions\. Before training, we compute and freeze

ti=normalize⁡\(1\|Qi\|​∑j∈QiHθ0,jL​\(Pitext\)\)\.t\_\{i\}=\\operatorname\{normalize\}\\left\(\\frac\{1\}\{\|Q\_\{i\}\|\}\\sum\_\{j\\in Q\_\{i\}\}H\_\{\\theta\_\{0\},j\}^\{L\}\(P\_\{i\}^\{\\mathrm\{text\}\}\)\\right\)\.\(6\)For a visual\-prompt templatePivisP\_\{i\}^\{\\mathrm\{vis\}\}, the prompt boxbib\_\{i\}and the preprocessed image\-token grid define a setSiS\_\{i\}of visual\-token positions inside the prompt region\. We pool only those positions:

zivis=normalize⁡\(1\|Si\|​∑j∈SiHθ,jL​\(Pivis\)\)\.z\_\{i\}^\{\\mathrm\{vis\}\}=\\operatorname\{normalize\}\\left\(\\frac\{1\}\{\|S\_\{i\}\|\}\\sum\_\{j\\in S\_\{i\}\}H\_\{\\theta,j\}^\{L\}\(P\_\{i\}^\{\\mathrm\{vis\}\}\)\\right\)\.\(7\)PVRD\-SG minimizes cosine distance between the prompt\-region representation and the cached text target:

ℒsg​\(θ\)=𝔼i∈ℐvis​\[1−\(zivis\)⊤​sg⁡\(ti\)\]\.\\mathcal\{L\}\_\{\\mathrm\{sg\}\}\(\\theta\)=\\mathbb\{E\}\_\{i\\in\\mathcal\{I\}\_\{\\mathrm\{vis\}\}\}\\left\[1\-\(z\_\{i\}^\{\\mathrm\{vis\}\}\)^\{\\top\}\\operatorname\{sg\}\(t\_\{i\}\)\\right\]\.\(8\)By construction, the text target containsqiq\_\{i\}but neitheraia\_\{i\}norrir\_\{i\}\. Freezingθ0\\theta\_\{0\}, cachingtit\_\{i\}, and stopping its gradient keep the target fixed\. The readout excludes both the visual evidence outsidebib\_\{i\}and assistant target positions\. The loss is read from selected positions, but shared parameters can still propagate its effect beyond the box\. PVRD\-SG therefore does not teach the answer through its target\. It teaches the visual prompt region where the task semantics represented by the typed question should be available, while the ordinary autoregressive loss remains responsible for the reasoning trace and final answer\.

### PRMLP: Masked latent prediction

PVRD\-SG specifies a text\-side semantic target, but it does not directly require the visual representation to remain stable when parts of the rendered question are difficult to observe\. PRMLP supplies this image\-side constraint\. It is a latent consistency objective, not pixel reconstruction or transcript generation: the representation read from a partially masked prompt region must approach the representation of the corresponding clean crop\. On a scheduled PRMLP update,ℳρ,ω\\mathcal\{M\}\_\{\\rho,\\omega\}masks a fractionρ\\rhoof the prompt region under block\-sampling ruleω\\omega, while leaving the rest of the image unchanged\. Letxivisx\_\{i\}^\{\\mathrm\{vis\}\}be the VTS composite or real\-world image:

xiM=ℳρ,ω​\(xivis;bi\)\.x\_\{i\}^\{M\}=\\mathcal\{M\}\_\{\\rho,\\omega\}\(x\_\{i\}^\{\\mathrm\{vis\}\};b\_\{i\}\)\.\(9\)The online branch receivesxiMx\_\{i\}^\{M\}, and the target branch receives the clean prompt cropcic\_\{i\}:

zifull\\displaystyle z\_\{i\}^\{\\mathrm\{full\}\}=pool⁡\(Hθ​\(qmin,xiM\),Si\),\\displaystyle=\\operatorname\{pool\}\\left\(H\_\{\\theta\}\(q\_\{\\min\},x\_\{i\}^\{M\}\),S\_\{i\}\\right\),\(10\)zicrop\\displaystyle z\_\{i\}^\{\\mathrm\{crop\}\}=pool⁡\(Hθ​\(qmin,ci\),Sic\)\.\\displaystyle=\\operatorname\{pool\}\\left\(H\_\{\\theta\}\(q\_\{\\min\},c\_\{i\}\),S\_\{i\}^\{c\}\\right\)\.\(11\)HereSiS\_\{i\}selects the prompt\-region visual tokens in the masked composite, andSicS\_\{i\}^\{c\}contains the visual tokens in the crop view\. The token mask is mapped from the stored box after image preprocessing\. The loss is

ℒprmlp​\(θ\)=𝔼i∈ℐvis​\[1−cos⁡\(qϕ​\(zifull\),sg⁡\(zicrop\)\)\]\.\\mathcal\{L\}\_\{\\mathrm\{prmlp\}\}\(\\theta\)=\\mathbb\{E\}\_\{i\\in\\mathcal\{I\}\_\{\\mathrm\{vis\}\}\}\\left\[1\-\\cos\\left\(q\_\{\\phi\}\(z\_\{i\}^\{\\mathrm\{full\}\}\),\\operatorname\{sg\}\(z\_\{i\}^\{\\mathrm\{crop\}\}\)\\right\)\\right\]\.\(12\)Both branches use the current model, and the clean\-crop branch is detached\.qϕq\_\{\\phi\}is an optional predictor; the reported implementation sets it to the identity\. The target therefore changes asθ\\thetachanges; there is no exponential\-moving\-average encoder\. PRMLP is applied to visual\-prompt examples and can run everykksteps to limit its additional cost\. It uses no transcript supervision\. Because the target is visual, PRMLP can also be applied to real\-world prompts whose typography and layout differ from the VTS renderer, provided that the training example supplies a prompt region and its clean crop\.

The complete supervised recipe is

ℒtotal=ℒsft\+λsg​ℒsg\+λprmlp​ℒprmlp,\\mathcal\{L\}\_\{\\mathrm\{total\}\}=\\mathcal\{L\}\_\{\\mathrm\{sft\}\}\+\\lambda\_\{\\mathrm\{sg\}\}\\mathcal\{L\}\_\{\\mathrm\{sg\}\}\+\\lambda\_\{\\mathrm\{prmlp\}\}\\mathcal\{L\}\_\{\\mathrm\{prmlp\}\},\(13\)where both loss weights are fixed before training and recorded in the run manifest\.

The objectives supervise different relationships\. PVRD\-SG specifies which typed task semantics should be recoverable from the prompt region; PRMLP preserves the region’s visual representation under partial occlusion\. Their shared answer target and common prompt\-region readout keep both signals tied to the same task\-bearing pixels\. Neither objective changes the inference interface, and neither supplies an answer or a transcript at test time\.

### GSPO continuation

We continue the model trained with both PVRD\-SG and PRMLP using GSPO on a separate mixture of original and VTS examples\. The raw reward is

R​\(y^,a\)=0\.1​Rformat​\(y^\)\+0\.9​Rans​\(y^,a\)\.R\(\\hat\{y\},a\)=0\.1R\_\{\\mathrm\{format\}\}\(\\hat\{y\}\)\+0\.9R\_\{\\mathrm\{ans\}\}\(\\hat\{y\},a\)\.\(14\)Both terms are binary\.RformatR\_\{\\mathrm\{format\}\}checks the output grammar, andRansR\_\{\\mathrm\{ans\}\}applies a deterministic verifier to the parsed final answer\. The coefficients specify the raw reward, not each term’s share of the policy gradient, and the reward provides no separate prompt\-transcription supervision\. We refer to the resulting checkpoint as the GSPO continuation\.

## 4Experiments and analysis

The experiments ask two questions: how much does reasoning change when a question becomes pixels, and how much of that loss can prompt\-region grounding recover? The primary evaluation uses MATH\-Vision, MathVista, ChartQA, and MMMU\(Wang et al\.[2024a](https://arxiv.org/html/2608.04726#bib.bib24); Lu et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib19); Masry et al\.[2022](https://arxiv.org/html/2608.04726#bib.bib21); Yue et al\.[2024](https://arxiv.org/html/2608.04726#bib.bib33)\)\. Unless a table states otherwise, trained conditions report mean accuracy over completed runs\. “Mean” is the unweighted average of these four tasks\. Accuracy is reported in percent; gains and gaps are percentage points\.

### Experimental setup

The Original view keeps the question in the text channel\. VTS renders the same question into the image and supplies only a short fixed cue\. The visual evidence and answer remain paired\. We report Original accuracy, VTS accuracy, and their gap,Δ=Accorig−Accvts\\Delta=\\operatorname\{Acc\}^\{\\mathrm\{orig\}\}\-\\operatorname\{Acc\}^\{\\mathrm\{vts\}\}\.

Our main adaptation experiments use Qwen3\-VL\-4B\-Instruct\(Bai et al\.[2025a](https://arxiv.org/html/2608.04726#bib.bib3)\)\. The common supervised pool contains 24,761 controlled source examples, each exported once as Original and once as VTS, plus 50,389 real\-world visual\-prompt examples\. This produces 99,911 training views from 75,150 source examples\. A separate set of 1,000 real\-world examples is held out before export and is never used for SFT or GSPO\. Every adaptation condition uses the same training export, assistant\-target policy, initialization, optimizer, checkpoint rule, and evaluation protocol\.

Balanced Replay and SFT are distinct controls rather than consecutive stages\. Balanced Replay follows the standard supervised schedule and measures what the mixed Original/VTS exposure provides by itself\. SFT uses the same interface mixture and ordinary next\-token objective but continues until its measured training cost matches Ours\. The cost ledger counts all forward, target, and backward passes introduced by PVRD\-SG and PRMLP; SFT is stopped when its training FLOPs are within 5% of Ours\. GPU\-hours are recorded as a secondary system measure but are not the matching criterion\. Comparing Ours with Balanced Replay tests the gain over the standard mixed\-interface recipe, while comparing it with SFT tests whether ordinary additional optimization is sufficient\. We use the latter as the primary baseline for method gains\. Full data, hyperparameters, and evaluator details are in the supplement\.

### The semantic channel gap

Accuracy falls under VTS for every evaluated model and task\. We compare thinking and instruct variants of Qwen3\-VL\(Bai et al\.[2025a](https://arxiv.org/html/2608.04726#bib.bib3)\), InternVL3\.5\(Wang et al\.[2025c](https://arxiv.org/html/2608.04726#bib.bib28)\), and the reasoning model DeepEyes\(Zheng et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib37)\)\. The complete task\-level table is in the supplement\. For both Qwen3\-VL sizes, the thinking variants are stronger on typed questions but lose more when the question moves into the image\. On MATH\-Vision, Qwen3\-VL\-4B\-Thinking scores 60\.0 under Original, compared with 51\.6 for Qwen3\-VL\-4B\-Instruct, yet its gap is 27\.4 rather than 16\.8 points\.

A small gap is not sufficient when both accuracies are low\. DeepEyes\-7B has the smallest average gap but reaches only 26\.6 Original accuracy on MATH\-Vision\. Figure[2](https://arxiv.org/html/2608.04726#S1.F2)therefore keeps both interface accuracies visible rather than ranking models by gap alone\. Construction controls support this interpretation\. For Base, adding a blank panel stays within 1\.1 points of Original and duplicating the question across both channels stays within 0\.6, whereas removing the question causes accuracy to collapse\. Ours preserves this pattern and improves over Balanced Replay across the tested renderer variants \(Appendix Table[18](https://arxiv.org/html/2608.04726#A6.T18)and Figure[8](https://arxiv.org/html/2608.04726#A6.F8)\)\.

### Grounding visual task semantics

Table 1:Four\-task cost\-matched adaptation\. SFT follows the ordinary supervised objective and is trained to the same measured cost as Ours\. Benchmarks form column groups, with Original and VTS views shown beneath each benchmark\.At the same measured training cost, Ours improves VTS accuracy over SFT on all four tasks \(Table[1](https://arxiv.org/html/2608.04726#S4.T1)\)\. The gains are 7\.8 points on MATH\-Vision, 7\.0 on MathVista, 6\.2 on ChartQA, and 12\.4 on MMMU\. Original accuracy changes by 0\.6 to 1\.6 points\. Averaged across tasks, VTS accuracy rises from 58\.0 to 66\.3 and Original accuracy from 69\.1 to 70\.3, reducing the interface gap from 11\.2 to 4\.0 points\. The remaining gap is 0\.7 points on MATH\-Vision and 2\.4 on MathVista, compared with 6\.8 on ChartQA and 6\.1 on MMMU\.

The two baselines clarify where this improvement comes from\. Balanced Replay reaches 61\.4 mean VTS accuracy and 69\.8 Original accuracy, so Ours adds 4\.9 VTS points while retaining the typed interface within 0\.5 point\. Cost\-matched SFT reaches only 58\.0 VTS accuracy despite receiving the larger ordinary supervised budget\. Ours exceeds it by 8\.4 points on VTS and 1\.2 points on Original\. The result does not imply that longer SFT must always underperform a shorter schedule\. It shows that, under the matched data, optimizer, checkpoint, and cost protocol used here, additional next\-token training does not account for the gain from the region\-level objectives\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x5.png)Figure 5:Component contributions and specificity\. Marks show four\-task mean VTS accuracy\. The middle group compares spatial readouts at matched training cost; the bottom group contrasts matched and deranged targets\. Original accuracy ranges from 69\.8 to 70\.3 across the displayed conditions\.PVRD\-SG accounts for the larger single\-component gain, while PRMLP adds 1\.8 VTS points over PVRD\-SG when the objectives are combined \(Figure[5](https://arxiv.org/html/2608.04726#S4.F5)\)\. Original accuracy varies by at most 0\.5 points across these rows\. Reading from the known prompt panel outperforms full\-image and random\-region summaries at matched training cost\. Matching each target to its own example also beats a deranged target by 2\.6 points for PVRD\-SG and 1\.0 for PRMLP\. The benefit therefore depends on both where the representation is read and which semantic target it receives\. The representation probes follow the objective design: PVRD\-SG primarily improves typed\-question retrieval, while PRMLP primarily improves clean\-crop retrieval \(Appendix Table[11](https://arxiv.org/html/2608.04726#A6.T11)\)\.

The advantage also grows throughout training rather than appearing at one selected checkpoint\. At 25%, 50%, 75%, and 100% of the matched budget, Ours exceeds Balanced Replay by 1\.2, 2\.5, 4\.1, and 4\.9 VTS points, respectively, while Original accuracy stays within 0\.5 points \(Appendix Figure[8](https://arxiv.org/html/2608.04726#A6.F8)\)\.

### Recognition is not task execution

Table 2:Same\-composite recognition and use diagnostic\. Two\-task means on the frozen composite\. Self/GT reinsert transcripts; P\-crop/S\-crop add token\-matched prompt/scene crops; EM is exact match transcription\.All answer conditions retain the same frozen VTS composite\.*Direct*asks the model to answer that composite with the fixed cue used throughout VTS\.*Self*adds the same checkpoint’s cached transcription to the text channel, whereas*GT*adds the ground\-truth question\. Thus Self minus Direct measures how much answer accuracy is recovered when the model’s recognized words regain the native text\-channel role, and GT minus Self estimates the remaining effect of transcription errors\.*P\-crop*adds a clean crop of the prompt region as a second image\.*S\-crop*adds a scene crop sampled outside that region with matched dimensions, aspect ratio, image order, visual\-token count, and auxiliary cue\. Their difference tests whether isolating the task\-bearing region helps beyond supplying another image crop\. EM is exact match between the cached transcription and the ground\-truth question\.

The base model already transcribes the prompt well, yet reinserting its own transcript improves answer accuracy by 7\.7 points \(Table[2](https://arxiv.org/html/2608.04726#S4.T2)\)\. After full adaptation, Direct accuracy is 13\.7 points higher and the Self advantage falls to 0\.8; replacing Self with ground\-truth text adds only another 0\.5 point\. Thus recognition alone does not explain the base failure, while adaptation reduces the benefit of moving the recognized question back into the text channel\.

The crop interventions lead to the same interpretation from the visual side\. For Base, the prompt crop improves Direct accuracy by 6\.2 points, while the matched scene crop adds only 0\.9\. Ours answers the full composite at 62\.1 and reaches 63\.0 with the prompt crop and 62\.2 with the scene crop\. The prompt region is therefore useful when isolated for the base model, but after adaptation the full composite already provides nearly all of that benefit\.

Prompt\-swap and fixed\-OCR interventions show the same reduction in reliance on text\-channel reinsertion; their complete results are in Appendix Table[10](https://arxiv.org/html/2608.04726#A6.T10)and Figure[9](https://arxiv.org/html/2608.04726#A6.F9)\.

### RL refinement

Table 3:Cost\-matched GSPO continuation on Qwen3\-VL\-8B\-Instruct\. Values are four\-task mean accuracies\.At matched training cost, GSPO improves mean VTS accuracy by 1\.4 points over continued SFT without reducing Original accuracy \(Table[3](https://arxiv.org/html/2608.04726#S4.T3)\); format accuracy also rises from 98\.0 to 99\.3\. The cost\-matched InternVL3\.5\-8B transfer result is reported in Appendix Table[13](https://arxiv.org/html/2608.04726#A6.T13)\.

### Independent benchmark transfer

Table 4:External evaluations with distinct protocols\. VISTA reports official weighted VT/Text accuracy and VT gains over SFT; its Base row is Qwen3\-VL\-8B\-Instruct\. OCRBench v2 reports the official ZH and EN scores; avg gain is the unrounded mean of the language\-specific gains over SFT\.On the independently constructed VISTA\-Bench pairs, Ours raises weighted VT accuracy from 56\.2 to 60\.3 while Text accuracy changes from 59\.4 to 59\.6 \(Table[4](https://arxiv.org/html/2608.04726#S4.T4)\)\. With GSPO, VT reaches 60\.1 and the Text\-to\-VT gap is 1\.2 points\. On OCRBench v2, Ours scores 60\.5/65\.9 on ZH/EN, a 2\.3\-point average gain over SFT; Ours\+\+GSPO gains 2\.1 points\. OCRBench measures text\-rich capability rather than semantic\-channel equivalence, so it is not pooled with VISTA\. Category\-level VISTA results and other text\-rich benchmarks are in the supplement\.

### Real\-world task images

![Refer to caption](https://arxiv.org/html/2608.04726v1/x6.png)Figure 6:Real\-world natural\-page evaluation\. Accuracy under SFT and Ours across four categories\. Labels give the category scores and gains over SFT\.We evaluate 1,000 real\-world examples that are disjoint from training: 374 UI screenshots, 188 worksheets, 157 forms or slides, and 281 photographed pages\. Ours improves over SFT in every category, by 7\.1 to 8\.9 points \(Figure[6](https://arxiv.org/html/2608.04726#S4.F6)\)\. Example\-weighted accuracy rises from 58\.0 to 65\.8, a gain of 7\.9 points; the unweighted category macro rises from 57\.8 to 65\.6\. This set has no constructed typed counterpart, so it measures generalization to the evaluated real\-world pages rather than semantic\-channel equivalence\.

#### Training controls\.

Question\-scrubbed reasoning reaches 65\.8 VTS accuracy, only 0\.5 points below a full trace that may repeat the question \(Appendix Table[9](https://arxiv.org/html/2608.04726#A6.T9)\)\. The gain therefore does not depend on copying the question into the assistant target\. Ours also exceeds VoQA QRA\-SFT by 2\.1 VTS points while retaining higher Original accuracy\.

## 5Discussion and conclusion

The same task is not equally usable across input channels\. Moving a question from text tokens into image pixels lowers accuracy in all 24 evaluated model and task combinations, by 17\.8 points on average\. Blank\-canvas and duplicate question tests do not reproduce the loss, and the base model can transcribe most visual questions exactly while still answering them poorly\. The semantic channel gap is therefore not well described as an OCR failure alone\.

Prompt\-region grounding raises four\-benchmark VTS accuracy from 58\.0 to 66\.3 at matched cost while Original accuracy changes from 69\.1 to 70\.3\. Prompt\-panel and deranged\-target controls tie the gain to the intended spatial and semantic signals without identifying a unique internal mechanism\. Ours also gains 4\.1 points on the paired VISTA\-Bench interface and 7\.9 points on held\-out real\-world pages; the supplement reports a cost\-matched second\-backbone test\. Inference remains one direct model call without an OCR transcript or prompt box\.

## References

- An et al\. \(2025\)An, J\.; Jiang, L\.; Luo, J\.; Wu, W\.; and Huang, L\. 2025\.VoQA: Visual\-only Question Answering\.*arXiv preprint arXiv:2505\.14227*\.
- Assran et al\. \(2023\)Assran, M\.; Duval, Q\.; Misra, I\.; Bojanowski, P\.; Vincent, P\.; Rabbat, M\.; LeCun, Y\.; and Ballas, N\. 2023\.Self\-Supervised Learning from Images with a Joint\-Embedding Predictive Architecture\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 15619–15629\.
- Bai et al\. \(2025a\)Bai, S\.; Cai, Y\.; Chen, R\.; Chen, K\.; Chen, X\.; Cheng, Z\.; Deng, L\.; Ding, W\.; Gao, C\.; Ge, C\.; et al\. 2025a\.Qwen3\-VL Technical Report\.*arXiv preprint arXiv:2511\.21631*\.
- Bai et al\. \(2025b\)Bai, S\.; Chen, K\.; Liu, X\.; Wang, J\.; Ge, W\.; Song, S\.; Dang, K\.; Wang, P\.; Wang, S\.; Tang, J\.; et al\. 2025b\.Qwen2\.5\-VL Technical Report\.*arXiv preprint arXiv:2502\.13923*\.
- Chen et al\. \(2026\)Chen, L\.; Zhao, X\.; Ding, K\.; Feng, W\.; Miao, C\.; Wang, Z\.; Guo, W\.; Wang, Y\.; Zheng, K\.; Zhang, B\.; Li, Z\.; and Xiang, S\. 2026\.Beyond Next\-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions\.*arXiv preprint arXiv:2602\.09483*\.
- Chen and He \(2021\)Chen, X\.; and He, K\. 2021\.Exploring Simple Siamese Representation Learning\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*\.
- DeepSeek\-AI et al\. \(2025\)DeepSeek\-AI; Guo, D\.; Yang, D\.; Zhang, H\.; Song, J\.; Wang, P\.; Zhu, Q\.; Xu, R\.; et al\. 2025\.DeepSeek\-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning\.*Nature*, 645: 633–638\.
- Duan et al\. \(2026\)Duan, S\.; Xue, Y\.; Wang, W\.; Su, Z\.; Liu, H\.; Yang, S\.; Gan, G\.; Wang, G\.; Wang, Z\.; Yan, S\.; Jin, D\.; Zhang, Y\.; Wen, G\.; Wang, Y\.; Zhang, Y\.; Zhang, X\.; Hong, W\.; Cen, Y\.; Yin, D\.; Chen, B\.; Yu, W\.; Gu, X\.; and Tang, J\. 2026\.GLM\-OCR Technical Report\.*arXiv preprint arXiv:2603\.10910*\.
- Feng et al\. \(2025\)Feng, Q\.; Li, W\.; Lin, T\.; and Chen, X\. 2025\.Align\-KD: Distilling Cross\-Modal Alignment Knowledge for Mobile Vision\-Language Large Model Enhancement\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 4178–4188\.
- Fu et al\. \(2025\)Fu, L\.; Kuang, Z\.; Song, J\.; Huang, M\.; Yang, B\.; Li, Y\.; Zhu, L\.; Luo, Q\.; Wang, X\.; Lu, H\.; Li, Z\.; Tang, G\.; Shan, B\.; Lin, C\.; Liu, Q\.; Wu, B\.; Feng, H\.; Liu, H\.; Huang, C\.; Tang, J\.; Chen, W\.; Jin, L\.; Liu, Y\.; and Bai, X\. 2025\.OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning\.*arXiv preprint arXiv:2501\.00321*\.
- Grill et al\. \(2020\)Grill, J\.\-B\.; Strub, F\.; Altché, F\.; Tallec, C\.; Richemond, P\. H\.; Buchatskaya, E\.; Doersch, C\.; Avila Pires, B\.; Guo, Z\. D\.; Gheshlaghi Azar, M\.; Piot, B\.; Kavukcuoglu, K\.; Munos, R\.; and Valko, M\. 2020\.Bootstrap Your Own Latent: A New Approach to Self\-Supervised Learning\.In*Advances in Neural Information Processing Systems*\.
- Kimi Team \(2025\)Kimi Team\. 2025\.Kimi\-VL Technical Report\.*arXiv preprint arXiv:2504\.07491*\.
- Leng et al\. \(2025\)Leng, S\.; Wang, J\.; Li, J\.; Zhang, H\.; Hu, Z\.; Zhang, B\.; Jiang, Y\.; Zhang, H\.; Li, X\.; Bing, L\.; Zhao, D\.; Lu, W\.; Rong, Y\.; Sun, A\.; and Lu, S\. 2025\.MMR1: Enhancing Multimodal Reasoning with Variance\-Aware Sampling and Open Resources\.*arXiv preprint arXiv:2509\.21268*\.
- Li et al\. \(2024a\)Li, B\.; Zhang, Y\.; Guo, D\.; Zhang, R\.; Li, F\.; Zhang, H\.; Zhang, K\.; Zhang, P\.; Li, Y\.; Liu, Z\.; and Li, C\. 2024a\.LLaVA\-OneVision: Easy Visual Task Transfer\.*arXiv preprint arXiv:2408\.03326*\.
- Li et al\. \(2024b\)Li, X\.; Lu, Y\.; Gan, Z\.; Gao, J\.; Wang, W\. Y\.; and Choi, Y\. 2024b\.Text as Images: Can Multimodal Large Language Models Follow Printed Instructions in Pixels?arXiv:2311\.17647\.
- Lian et al\. \(2025\)Lian, S\.; Wu, C\.; Yang, L\. T\.; Yuan, H\.; Yu, B\.; Zhang, L\.; and Chen, K\. 2025\.Euclid’s Gift: Enhancing Spatial Perception and Reasoning in Vision\-Language Models via Geometric Surrogate Tasks\.*arXiv preprint arXiv:2509\.24473*\.
- Lin et al\. \(2026\)Lin, H\.; Liu, Z\.; Zhu, Y\.; Qin, C\.; Lin, J\.; Shang, X\.; He, C\.; Zhang, W\.; and Wu, L\. 2026\.MMFineReason: Closing the Multimodal Reasoning Gap via Open Data\-Centric Methods\.*arXiv preprint arXiv:2601\.21821*\.
- Liu et al\. \(2026\)Liu, Q\.; Feng, J\.; Wang, Y\.; Han, X\.; Cheng, Y\.; Zhu, Y\.; Diao, H\.; Zhuge, Y\.; and Lu, H\. 2026\.VISTA\-Bench: Do Vision\-Language Models Really Understand Visualized Text as Well as Pure Text?*arXiv preprint arXiv:2602\.04802*\.
- Lu et al\. \(2024\)Lu, P\.; Bansal, H\.; Xia, T\.; Liu, J\.; Li, C\.; Hajishirzi, H\.; Cheng, H\.; Chang, K\.\-W\.; Galley, M\.; and Gao, J\. 2024\.MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts\.In*International Conference on Learning Representations*\.
- Ma et al\. \(2024\)Ma, Y\.; Zang, Y\.; Chen, L\.; Chen, M\.; Jiao, Y\.; Li, X\.; Lu, X\.; Liu, Z\.; Ma, Y\.; Dong, X\.; Zhang, P\.; Pan, L\.; Jiang, Y\.\-G\.; Wang, J\.; Cao, Y\.; and Sun, A\. 2024\.MMLongBench\-Doc: Benchmarking Long\-context Document Understanding with Visualizations\.In*Advances in Neural Information Processing Systems*\.
- Masry et al\. \(2022\)Masry, A\.; Long, D\. X\.; Tan, J\. Q\.; Joty, S\.; and Hoque, E\. 2022\.ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning\.In*Findings of the Association for Computational Linguistics: ACL 2022*, 2263–2279\.
- Meng et al\. \(2025\)Meng, F\.; Du, L\.; Liu, Z\.; Zhou, Z\.; Lu, Q\.; Fu, D\.; Han, T\.; Shi, B\.; Wang, W\.; He, J\.; Zhang, K\.; Luo, P\.; Qiao, Y\.; Zhang, Q\.; and Shao, W\. 2025\.MM\-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule\-Based Reinforcement Learning\.*arXiv preprint arXiv:2503\.07365*\.
- Qiao et al\. \(2025\)Qiao, R\.; Tan, Q\.; Yang, P\.; Wang, Y\.; Wang, X\.; Wan, E\.; Zhou, S\.; Dong, G\.; Zeng, Y\.; Xu, Y\.; et al\. 2025\.We\-Math 2\.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning\.*arXiv preprint arXiv:2508\.10433*\.
- Wang et al\. \(2024a\)Wang, K\.; Pan, J\.; Shi, W\.; Lu, Z\.; Ren, H\.; Zhou, A\.; Zhan, M\.; and Li, H\. 2024a\.Measuring Multimodal Mathematical Reasoning with MATH\-Vision Dataset\.In*Advances in Neural Information Processing Systems*, volume 37\.
- Wang et al\. \(2024b\)Wang, P\.; Bai, S\.; Tan, S\.; Wang, S\.; Fan, Z\.; Bai, J\.; Chen, K\.; Liu, X\.; Wang, J\.; Ge, W\.; Fan, Y\.; Dang, K\.; Du, M\.; Ren, X\.; Men, R\.; Liu, D\.; Zhou, C\.; Zhou, J\.; and Lin, J\. 2024b\.Qwen2\-VL: Enhancing Vision\-Language Model’s Perception of the World at Any Resolution\.*arXiv preprint arXiv:2409\.12191*\.
- Wang et al\. \(2025a\)Wang, P\.; Li, Z\.\-Z\.; Yin, F\.; Ran, D\.; and Liu, C\.\-L\. 2025a\.MV\-MATH: Evaluating Multimodal Math Reasoning in Multi\-Visual Contexts\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 19541–19551\.
- Wang et al\. \(2025b\)Wang, W\.; Gao, Z\.; Chen, L\.; Chen, Z\.; Zhu, J\.; Zhao, X\.; Liu, Y\.; Cao, Y\.; Ye, S\.; Zhu, X\.; Lu, L\.; Duan, H\.; Qiao, Y\.; Dai, J\.; and Wang, W\. 2025b\.VisualPRM: An Effective Process Reward Model for Multimodal Reasoning\.*arXiv preprint arXiv:2503\.10291*\.
- Wang et al\. \(2025c\)Wang, W\.; Gao, Z\.; Gu, L\.; Pu, H\.; Cui, L\.; Wei, X\.; Liu, Z\.; Jing, L\.; Ye, S\.; Shao, J\.; et al\. 2025c\.InternVL3\.5: Advancing Open\-Source Multimodal Models in Versatility, Reasoning, and Efficiency\.*arXiv preprint arXiv:2508\.18265*\.
- Wang et al\. \(2025d\)Wang, Y\.; Yang, Z\.; Cao, M\.; Han, M\.; Lin, H\.; Zhu, Y\.; Chang, X\.; and Liang, X\. 2025d\.CARE What Fails: Contrastive Anchored\-REflection for Verifiable Multimodal Reasoning\.*arXiv preprint arXiv:2512\.19554*\.
- Wiedmann et al\. \(2025\)Wiedmann, L\.; Zohar, O\.; Mahla, A\.; Wang, X\.; Li, R\.; Frere, T\.; von Werra, L\.; Gosthipaty, A\. R\.; and Marafioti, A\. 2025\.FineVision: Open Data Is All You Need\.*arXiv preprint arXiv:2510\.17269*\.
- Wu et al\. \(2024\)Wu, Z\.; Chen, X\.; Pan, Z\.; Liu, X\.; Liu, W\.; Dai, D\.; Gao, H\.; Ma, Y\.; Wu, C\.; Wang, B\.; Xie, Z\.; Wu, Y\.; Hu, K\.; Wang, J\.; Sun, Y\.; Li, Y\.; Piao, Y\.; Guan, K\.; Liu, A\.; Xie, X\.; You, Y\.; Dong, K\.; Yu, X\.; Zhang, H\.; Zhao, L\.; Wang, Y\.; and Ruan, C\. 2024\.DeepSeek\-VL2: Mixture\-of\-Experts Vision\-Language Models for Advanced Multimodal Understanding\.*arXiv preprint arXiv:2412\.10302*\.
- Xi et al\. \(2025\)Xi, Z\.; Li, G\.; Fan, Y\.; Guo, H\.; Liu, Y\.; Fan, X\.; Liu, J\.; Ding, J\.; Zuo, W\.; Yin, Z\.; Bai, L\.; Ji, T\.; Gui, T\.; Zhang, Q\.; Torr, P\.; and Huang, X\. 2025\.BMMR: A Large\-Scale Bilingual Multimodal Multi\-Discipline Reasoning Dataset\.*arXiv preprint arXiv:2507\.03483*\.
- Yue et al\. \(2024\)Yue, X\.; Ni, Y\.; Zhang, K\.; Zheng, T\.; Liu, R\.; Zhang, G\.; Stevens, S\.; Jiang, D\.; Ren, W\.; Sun, Y\.; Wei, C\.; Yu, B\.; Yuan, R\.; Sun, R\.; Yin, M\.; Zheng, B\.; Yang, Z\.; Liu, Y\.; Huang, W\.; Sun, H\.; Su, Y\.; and Chen, W\. 2024\.MMMU: A Massive Multi\-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 9556–9567\.
- Zhang et al\. \(2024\)Zhang, R\.; Jiang, D\.; Zhang, Y\.; Lin, H\.; Guo, Z\.; Qiu, P\.; Zhou, A\.; Lu, P\.; Chang, K\.\-W\.; Qiao, Y\.; Gao, P\.; and Li, H\. 2024\.MathVerse: Does Your Multi\-modal LLM Truly See the Diagrams in Visual Math Problems?In*European Conference on Computer Vision*, 169–186\.
- Zhao et al\. \(2025\)Zhao, H\.; Wang, M\.; Zhu, F\.; Liu, W\.; Ni, B\.; Zeng, F\.; Meng, G\.; and Zhang, Z\. 2025\.VTCBench: Can Vision\-Language Models Understand Long Context with Vision\-Text Compression?*arXiv preprint arXiv:2512\.15649*\.
- Zheng et al\. \(2025\)Zheng, C\.; Liu, S\.; Li, M\.; Chen, X\.\-H\.; Yu, B\.; Gao, C\.; Dang, K\.; Liu, Y\.; Men, R\.; Yang, A\.; Zhou, J\.; and Lin, J\. 2025\.Group Sequence Policy Optimization\.*arXiv preprint arXiv:2507\.18071*\.
- Zheng et al\. \(2026\)Zheng, Z\.; Yang, M\.; Hong, J\.; Zhao, C\.; Xu, G\.; Yang, L\.; Shen, C\.; and Yu, X\. 2026\.DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning\.In*International Conference on Learning Representations*\.
- Zhong et al\. \(2022\)Zhong, Y\.; Yang, J\.; Zhang, P\.; Li, C\.; Codella, N\.; Li, L\. H\.; Zhou, L\.; Dai, X\.; Yuan, L\.; Li, Y\.; and Gao, J\. 2022\.RegionCLIP: Region\-based Language\-Image Pretraining\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 16793–16803\.
- Zhu et al\. \(2025\)Zhu, J\.; Wang, W\.; Chen, Z\.; Liu, Z\.; Ye, S\.; Gu, L\.; Tian, H\.; Duan, Y\.; et al\. 2025\.InternVL3: Exploring Advanced Training and Test\-Time Recipes for Open\-Source Multimodal Models\.*arXiv preprint arXiv:2504\.10479*\.
- Zou et al\. \(2025\)Zou, C\.; Guo, X\.; Yang, R\.; Zhang, J\.; Hu, B\.; and Zhang, H\. 2025\.DynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models\.In*International Conference on Learning Representations*\.

## Appendix ASupplementary overview

The supplement follows the paper’s main argument\. It first clarifies the relationship to prior visual\-question work and gives the implementation of prompt\-region grounding\. It then reports the full semantic channel gap, data construction, training details, additional evaluations, and complete qualitative rollouts\. The final section states the limits of the evidence and the release considerations for real\-world images\.

## Appendix BRelationship to prior visual\-question work

#### Visual questions and visualized text\.

VIM studies visual instructions and cross\-interface mixture training\(Li et al\.[2024b](https://arxiv.org/html/2608.04726#bib.bib15)\)\. VoQA places the scene and question in one image and reconstructs the visual question during supervised fine\-tuning\(An et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib1)\)\. VISTA\-Bench evaluates matched text and visualized\-text questions under multiple rendering and OCR conditions\(Liu et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib18)\)\. These studies establish that visualized questions can be difficult\. VTS complements them with a source\-fixed intervention, construction controls, and prompt\-region supervision\. We do not claim that moving a question into pixels is itself new\.

#### Local alignment and reasoning post\-training\.

Region–text alignment and teacher–student vision\-to\-text distillation predate our objectives\(Zhong et al\.[2022](https://arxiv.org/html/2608.04726#bib.bib38); Feng et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib9); Chen et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib5)\)\. Our method uses the known task\-bearing region of a rendered training example as the readout for two representation\-level targets\. For the final stage, GSPO clips importance ratios at the response\-sequence level\(Zheng et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib36)\); we apply it with verifiable answer and format rewards after supervised grounding\.

## Appendix CImplementation of prompt\-region grounding

Both auxiliary objectives use a prompt\-region box and clean crop\. The renderer stores them for controlled VTS rows, while the GLM\-OCR pipeline extracts them for real\-world rows\. After image preprocessing, the box is mapped to the corresponding visual\-token region and the selected states are pooled\. PVRD\-SG aligns this prompt\-region representation with a cached embedding of the typed question as a stop\-gradient target\. PRMLP masks random blocks inside the prompt region and predicts the detached representation of the clean crop\. The reported formal configuration uses an identity predictor\. Neither objective asks the model to transcribe the prompt\.

Capturing only the final normalized visual states avoids materializing every decoder layer\. PVRD\-SG reuses the visual\-prompt supervised forward pass\. PRMLP adds a masked\-composite view and a clean\-crop target view on its scheduled updates\. Prompt boxes, cached targets, and crop views are used only during training; evaluation uses one composite image and the fixed text cue\.

## Appendix DFull semantic channel gap results

Table 5:Original/VTS accuracy for all 24 model–task pairs\.Each benchmark separates Original and VTS accuracy\. T and I denote thinking and instruct variants; Mean gap is the unweighted four\-task mean of Original minus VTS\.Accuracy decreases in all 24 combinations\. The thinking variants obtain stronger Original scores than their instruct counterparts on several tasks, but they also have larger mean gaps\. DeepEyes has the smallest mean gap and the lowest Original MATH\-Vision accuracy, which is why we report both interface accuracies rather than gap alone\.

## Appendix EData, training, and evaluation

#### Supervised grounding data\.

The supervised pool contains 75,150 source examples\. Its controlled portion has 24,761 examples, each exported in paired Original and VTS views, yielding 49,522 training views\. VTS views store the full composite, prompt\-panel coordinates, typed\-question target representation, and clean prompt crop; Original views retain the native question and source image\. The other 50,389 source examples are held\-in real\-world visual prompts described below\. Each retains its original text prompt alongside the task\-bearing image\. The GLM\-OCR pipeline provides a prompt box and clean crop for each real\-world example\. The resulting training export has 99,911 views\. All conditions use this same export and assistant\-target policy\. Ours applies PVRD\-SG and PRMLP to the 24,761 VTS views and 50,389 real\-world examples; there is no separate real\-world SFT stage\.

The controlled examples are sampled from MMR1\(Leng et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib13)\), BMMR\(Xi et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib32)\), Euclid30K\(Lian et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib16)\), MMK12\(Meng et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib22)\), FineVision subsets\(Wiedmann et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib30)\), mmopenr1\-8k\(Lin et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib17)\), and WeMath2 subsets\(Qiao et al\.[2025](https://arxiv.org/html/2608.04726#bib.bib23)\)\.

Table 6:Source examples in the common supervised grounding pool\.Each controlled example produces one Original and one VTS training view, so 24,761 controlled examples yield 49,522 views\. Together with the real\-world split, the optimizer sees 99,911 views\. The 1,000\-example real\-world validation split is excluded from every training condition\.
### Real\-world data and annotation

#### Collection and split\.

We collected 51,389 real\-world image and question pairs\. Before training, we set aside 1,000 examples for the natural\-page evaluation and removed them from the SFT and GSPO exports\. The remaining 50,389 examples are merged with the 24,761 controlled source examples in Table[6](https://arxiv.org/html/2608.04726#A5.T6), yielding 75,150 source examples\. After the controlled examples are expanded into paired views, the supervised export contains 99,911 training views\. Each real\-world example retains its original text prompt as the semantic target and receives both region\-level objectives in addition to the autoregressive supervised loss\.

#### Annotation workflow\.

The GLM\-OCR pipeline\(Duan et al\.[2026](https://arxiv.org/html/2608.04726#bib.bib8)\)first extracts the task text, its bounding box, and the corresponding crop from each image\. GPT\-5\.4 then reads the source image and extracted prompt, annotates the question, and produces a reasoning trace with a final answer\. Gemini 2\.5 Pro performs the final quality check by testing whether the reasoning supports the answer\. Only examples that pass this check are exported for training\. Independent college\-student review agreed with the finalized annotations on 97\.4% of a 1,500\-example sample\.

#### GSPO data selection\.

We mine the GSPO split from the finalized SFT pool\. For each source problem, the supervised checkpoint produces eight independent completions, which are scored by the deterministic answer verifier\. Letci=∑j=18𝟏​\[a^i​j=ai\]c\_\{i\}=\\sum\_\{j=1\}^\{8\}\\mathbf\{1\}\[\\hat\{a\}\_\{ij\}=a\_\{i\}\]be the number of correct completions\. We retain examples with1≤ci≤51\\leq c\_\{i\}\\leq 5\. This pass@8\-style filter removes unsolved cases with no positive signal and nearly saturated cases with six to eight correct completions\. The retained 23,488 source problems are exported in both Original and VTS form, giving 46,976 training rows\. A separate 512\-problem development split gives 1,024 paired rows\. GSPO initializes from the Ours supervised checkpoint and uses0\.1​rformat\+0\.9​ranswer0\.1\\,r\_\{\\mathrm\{format\}\}\+0\.9\\,r\_\{\\mathrm\{answer\}\}\. The continued\-SFT control starts from the same checkpoint and uses matched training\-cost accounting\.

Supervised grounding

GSPO continuation

Table 7:Training hyperparameters\.Candidate\-mining completions construct the GSPO training set; five policy completions per prompt are sampled during optimization\.![Refer to caption](https://arxiv.org/html/2608.04726v1/figures/training_set_examples.png)Figure 7:Representative samples from the real\-world training collection\.The examples span photographed screens and textbook pages, worksheets and exam sheets, learning interfaces, slides, and other layouts in which the task\-defining text appears inside the visual input\. These samples come from the training split, not the held\-out 1,000\-example evaluation\.
#### Evaluation\.

MATH\-Vision, MathVista, ChartQA, and MMMU use their standard answer extraction and accuracy evaluators\. Every method row reports Original and VTS on the same task set\. VISTA\-Bench uses its official paired Text/VT examples and weighted overall score\. OCRBench v2, DocVQA, TextVQA, and ST\-VQA retain their own official metrics and are never averaged into a VTS headline\. The natural\-page evaluation contains 1,000 real\-world examples held out from training: 374 UI screenshots, 188 worksheets, 157 forms/slides, and 281 photographed pages\. It uses each original visual input and reports standard answer accuracy; no paired typed view is constructed\.

#### Release\.

We plan to release the finalized annotations, split manifests, processing scripts, and all images that pass the applicable consent, privacy, and license checks\.

## Appendix FFurther evidence

### Prompt cues and supervision

Table 8:Cue robustness\.Values are four\-task macro accuracies; cue variants keep the VTS image fixed\.Training conditionOriginalVTSVTS gainAssistant target; gain over answer onlyAnswer only69\.863\.2baselineQuestion\-scrubbed trace70\.165\.8\+2\.6\+2\.6Full trace70\.366\.3\+3\.1\+3\.1Prior recipes; gain over Balanced ReplayBalanced Replay69\.861\.4baselineQA\-SFT69\.763\.0\+1\.6\+1\.6VoQA QRA\-SFT69\.864\.2\+2\.8\+2\.8Ours70\.366\.3\+4\.9\+4\.9Table 9:Training\-target and closest\-baseline comparisons\. Values are four\-task mean accuracies\. Target rows match source items, steps, and training cost\. Balanced Replay is the VIM\-style mixed\-interface baseline\.Changing the VTS cue improves the base\-model accuracy from 53\.5 to at most 56\.2, whereas Ours reaches 66\.3–67\.4 under all three cues\. The channel gap therefore remains visible under stronger scaffolds\. For assistant targets, the question\-scrubbed trace reaches 65\.8 VTS accuracy, compared with 63\.2 for answer\-only and 66\.3 for the full trace\. Thus most of the trace benefit does not require repeating the question\. Ours also reaches 66\.3 VTS and 70\.3 Original accuracy, compared with 64\.2 and 69\.8 for VoQA QRA\-SFT\.

### Training dynamics and prompt following

The VTS gain grows from 1\.2 points at one quarter of the budget to 4\.9 points at the completed budget \(Figure[8](https://arxiv.org/html/2608.04726#A6.F8)\)\. Replay VTS accuracy is 57\.2/59\.0/60\.3/61\.4 across the four cost fractions; Ours reaches 58\.4/61\.5/64\.4/66\.3, giving gains of 1\.2/2\.5/4\.1/4\.9\. Original accuracy stays within 0\.5 points between the two recipes throughout the trajectory\.

Table 10:Prompt\-swap and channel\-conflict controls\.The upper panel changes the visual prompt; the lower panel gives the two channels different scorable targets\.Prompt\-follow accuracy rises from 31\.2 for Base to 58\.9 for Ours, while evidence\-follow accuracy falls from 52\.4 to 26\.0\. Under direct channel conflict, visual\-instruction following rises by 10\.2 points over balanced replay\. The latter result also motivates treating typed–visual instruction priority as a separate safety question\.

### Representation and PRMLP analyses

Table 11:Representation sanity checks\.Retrieval uses a fixed 512\-item pool; rank and similarity help screen for collapse\.PVRD\-SG primarily raises text retrieval accuracy, whereas PRMLP primarily raises crop retrieval accuracy\. Ours improves both\. Effective rank remains above 109 and off\-diagonal similarity remains at or below 0\.24 across the displayed recipes\.

Figure[9](https://arxiv.org/html/2608.04726#A6.F9)compares the PRMLP objective and schedule under the same total\-cost accounting\. The formal masked target gives the largest VTS gain over PVRD\-SG only\. The shorthand settings are: small \(λ=0\.001\\lambda=0\.001, every 4 steps\), formal \(0\.0030\.003, every 2\), frequent \(0\.0030\.003, every step\), and high weight \(0\.0100\.010, every 2\); the unmasked, learned\-predictor, and matched\-token variants use0\.0030\.003every 2 steps\. Using a learned predictor or a matched\-token crop remains close to the formal setting, while the unmasked target and high\-weight setting are weaker\. Original accuracy spans 69\.6–70\.3 across these configurations\.

Table 12:GSPO task\-level continuation\.Original and VTS accuracy are reported separately for every benchmark and as an unweighted four\-task mean\. Both updates start from the same Ours checkpoint and use matched training cost\. Means use full\-precision task scores before one\-decimal display rounding\.Table 13:InternVL3\.5\-8B task\-level transfer\.Original and VTS accuracy are reported separately under the same four evaluation protocols; Mean is the unweighted four\-task aggregate computed before one\-decimal display rounding\.
### Generalization beyond VTS

Table 14:External text\-rich benchmarks\.Baseline is cost\-matched SFT, and each benchmark retains its official score\.Table 15:Prior matched\-interface protocols\.Each protocol retains its own split, prompt, and evaluator; scores are not pooled\.Ours improves over SFT on DocVQA, TextVQA, and ST\-VQA, with gains from 0\.6 to 1\.3 points\. These benchmarks measure text\-rich capability rather than a paired change in the channel carrying the task\. On InternVL3\.5\-8B, Ours improves VTS accuracy by 4\.1 points over cost\-matched SFT while Original accuracy changes by 0\.2 \(Table[13](https://arxiv.org/html/2608.04726#A6.T13)\)\. On prior matched\-interface protocols, Ours is 3\.7 points above balanced replay on VIM Pixel and 5\.1 points above it on VoQA Visual\-only\. Relative to VoQA QRA\-SFT, the gains are 2\.5 and 1\.2 points, respectively\.

Table 16:Visual\-instruction safety check\.Columns report clean VTS accuracy, higher\-priority instruction accuracy, benign visual following, and visible/obfuscated attack success\.The full method raises benign visual\-follow accuracy while higher\-priority instruction accuracy remains near 94%\. Visible and obfuscated attack success do not increase across the displayed checkpoints\. This check characterizes the evaluated conflicts and is separate from the grounding analysis\.

Table 17:VISTA\-Bench results\.Every category separates the official VT and Text interfaces\. Base is Qwen3\-VL\-8B\-Instruct, and the overall accuracy weights categories by their sample counts\.Relative to SFT, Ours improves VT accuracy by 1\.4 points on multimodal perception, 4\.3 on multimodal reasoning, 5\.5 on multimodal knowledge, and 4\.4 on unimodal knowledge\. Its weighted VT gain is 4\.1 points, while weighted Text accuracy changes by 0\.2 points\.

### Renderer robustness and task use

Table 18:VTS construction controls\. Canvas keeps the typed question and adds a blank panel; Duplicate presents it in both channels; Image\-only removes it from both\.![Refer to caption](https://arxiv.org/html/2608.04726v1/x7.png)Figure 8:Training and renderer robustness\.Left: matched\-budget VTS learning curves at 25%, 50%, 75%, and 100% of the full training cost; the final Ours\-over\-Replay difference is 4\.9 points\. Right: VTS accuracy when only prompt\-panel placement or font scale changes; labels give Ours\-over\-Replay gains\.Canvas stays within 1\.1 points of Original for Base, and Duplicate stays within 0\.6\. Image\-only accuracy collapses because neither channel contains the question \(Table[18](https://arxiv.org/html/2608.04726#A6.T18)\)\. Ours improves over balanced replay under every renderer\. Smaller text and unfamiliar placement remain harder than the default and large\-font conditions, so the result shows robustness within the tested renderer family rather than invariance to layout \(Figure[8](https://arxiv.org/html/2608.04726#A6.F8)\)\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x8.png)Figure 9:Task\-use and PRMLP diagnostics\.Left: Direct and OCR\-transcript accuracy on the same VTS composite; the OCR gain contracts from 7\.6 to 3\.1 to 0\.7 points\. Right: four\-task VTS accuracy under PRMLP variants; the dashed line is PVRD\-SG only, and the formal masked target is highlighted\.For Base, supplying the OCR transcript on the unchanged composite adds 7\.6 points, while restoring the source image adds a further 1\.8 points\. For the full method, the corresponding gains are 0\.7 and 0\.2 points\. The small same\-composite gain at the final supervised stage is consistent with answering the composite directly \(Figure[9](https://arxiv.org/html/2608.04726#A6.F9)\)\. With OCR fixed, restoring the source adds 1\.8, 0\.5, and 0\.2 points for Base, replay, and Ours; ground\-truth text adds a further 2\.3, 1\.4, and 0\.6 points on the same composite\. On restored sources \(Base/replay/Ours\), fixed\-cue, OCR, and ground\-truth scores are 20\.7/21\.0/21\.2, 57\.8/60\.6/63\.0, and 59\.1/62\.0/63\.6\.

## Appendix GQualitative case studies

Aggregate accuracy hides how visual\-task failures propagate\. Each case below pairs a complete, verified rollout from Our Model with an incorrect Qwen3VL\-8B\-Thinking rollout\. Red text marks the first recognition, reasoning, or diagram error and every later quantity that depends on it\. All four images come from the held\-out real\-world split and were excluded from SFT and GSPO training\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/x9.jpg)

Bank reconciliation\.The task asks for the unadjusted cash\-book balance, starting from the GH¢12,345 credit balance on the bank statement\.

Our Model\.1\.Add items that make the cash book higher than the statement: the dishonoured cheque, bank charges, two standing orders, deposit in transit, post\-dated cheques recorded as banked, and the GH¢198 entry error\.

Δ\+=5,000\+120\+1,220\+2,600\+2,500\+23,000\+198=34,638\.\\Delta^\{\+\}=5\{,\}000\+120\+1\{,\}220\+2\{,\}600\+2\{,\}500\+23\{,\}000\+198=34\{,\}638\.2\.Subtract the direct deposit and the three unpresented cheques:

991\+1,371\+2,635=4,997,Δ−=3,900\+4,997=8,897\.991\+1\{,\}371\+2\{,\}635=4\{,\}997,\\qquad\\Delta^\{\-\}=3\{,\}900\+4\{,\}997=8\{,\}897\.3\.Reconcile the balances:

12,345\+34,638−8,897=GH¢​38,086​Dr\.12\{,\}345\+34\{,\}638\-8\{,\}897=\\boxed\{\\text\{GH\\textcent\}38\{,\}086\\ \\mathrm\{Dr\}\}\.This matches the cash\-book balance printed in the image\.

Qwen3VL\-8B\-Thinking\.1\.The rollout transcribes23,00023\{,\}000as2,300\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}2\{,\}300\}and4,3214\{,\}321as4,123\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}4\{,\}123\}\.2\.The second error removes the required correction:

Δentry=4,123−4,123=0\.\\Delta\_\{\\mathrm\{entry\}\}=\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}4\{,\}123\-4\{,\}123=0\}\.3\.It keeps the other signs unchanged and computes

Δ~\+=5,000\+120\+1,220\+2,600\+2,500\+2,300\+0=13,740\.\\widetilde\{\\Delta\}^\{\+\}=5\{,\}000\+120\+1\{,\}220\+2\{,\}600\+2\{,\}500\+\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}2\{,\}300\}\+\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}0\}=13\{,\}740\.4\.It therefore returns

12,345\+13,740−8,897=GH¢​17,188​Dr\.12\{,\}345\+13\{,\}740\-8\{,\}897=\\boxed\{\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{GH\\textcent\}17\{,\}188\\ \\mathrm\{Dr\}\}\}\.The answer is short by GH¢20,898, exactly the omitted GH¢20,700 and GH¢198\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/figures/case_loan_amortization.png)

Equal\-instalment loan\.The loan principal is Sh 3,000,000, the annual rate is 10%, and four equal payments are made, one at the end of each year\.

Our Model\.1\.Equate the principal to the present value of an ordinary annuity:

3,000,000=A​1−\(1\.10\)−40\.10\.3\{,\}000\{,\}000=A\\frac\{1\-\(1\.10\)^\{\-4\}\}\{0\.10\}\.2\.Solve for the annual instalment:

A=3,000,000​\(0\.10\)1−\(1\.10\)−4=Sh​946,412\.41\.A=\\frac\{3\{,\}000\{,\}000\(0\.10\)\}\{1\-\(1\.10\)^\{\-4\}\}=\\boxed\{\\text\{Sh \}946\{,\}412\.41\}\.3\.LetBtB\_\{t\}be the balance after paymenttt\. Since each year’s interest is 10% of its opening balance,

Bt=1\.10​Bt−1−A,B0=3,000,000\.B\_\{t\}=1\.10B\_\{t\-1\}\-A,\\qquad B\_\{0\}=3\{,\}000\{,\}000\.The schedule closes as

B1\\displaystyle B\_\{1\}=2,353,587\.59,\\displaystyle=2\{,\}53\{,\}8759,B2\\displaystyle B\_\{2\}=1,642,533\.94,\\displaystyle=1\{,\}42\{,\}3394,B3\\displaystyle B\_\{3\}=860,374\.92,\\displaystyle=60\{,\}7492,B4\\displaystyle B\_\{4\}=0\.00\.\\displaystyle=00\.The zero final balance verifies both the instalment and the amortization\.

Qwen3VL\-8B\-Thinking\.

1\.The rollout treats the loan as carryingflat interest on the original principal, so it charges

It=0\.10​\(3,000,000\)=300,000in every year\.I\_\{t\}=\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}0\.10\(3\{,\}000\{,\}000\)=300\{,\}000\}\\quad\\text\{in every year\}\.2\.It sets total interest to4​\(300,000\)=1,200,000\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}4\(300\{,\}000\)=1\{,\}200\{,\}000\}\.3\.It divides principal plus flat interest by four:

A~=3,000,000\+1,200,0004=Sh​1,050,000\.\\widetilde\{A\}=\\frac\{3\{,\}000\{,\}000\+\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}1\{,\}200\{,\}000\}\}\{4\}=\\boxed\{\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{Sh \}1\{,\}050\{,\}000\}\}\.4\.This cannot be the required equal instalment: its four discounted payments have present value

1,050,000​1−\(1\.10\)−40\.10=Sh​3,328,358\.72≠Sh​3,000,000\.1\{,\}050\{,\}000\\frac\{1\-\(1\.10\)^\{\-4\}\}\{0\.10\}=\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{Sh \}3\{,\}328\{,\}358\.72\}\\neq\\text\{Sh \}3\{,\}000\{,\}000\.

Figure 10:Finance rollouts\.Each example contrasts the verified derivation from Our Model with a Qwen3VL\-8B\-Thinking rollout containing a specific recognition or reasoning error\.![Refer to caption](https://arxiv.org/html/2608.04726v1/figures/case_circuit_experiment.jpg)

Two\-wire circuit experiment\.The procedure measures current and voltage for wires A and B, then plots voltage on the vertical axis against current on the horizontal axis\.

Our Model\.1\.The ammeter is in series and the voltmeter is across the test wire, so each pair\(IX,VX\)\(I\_\{X\},V\_\{X\}\)refers to the same wireXX\.2\.The requested axes give

SX=Δ​VXΔ​IX=RX\.S\_\{X\}=\\frac\{\\Delta V\_\{X\}\}\{\\Delta I\_\{X\}\}=R\_\{X\}\.3\.For wires of the same material and length,

RX=ρ​LAX=4​ρ​Lπ​dX2\.R\_\{X\}=\\frac\{\\rho L\}\{A\_\{X\}\}=\\frac\{4\\rho L\}\{\\pi d\_\{X\}^\{2\}\}\.4\.Therefore,

SASB=RARB=\(dBdA\)2\.\\boxed\{\\frac\{S\_\{A\}\}\{S\_\{B\}\}=\\frac\{R\_\{A\}\}\{R\_\{B\}\}=\\left\(\\frac\{d\_\{B\}\}\{d\_\{A\}\}\\right\)^\{2\}\}\.The thicker wire must have the smaller slope\. The photograph contains no micrometer readings or measured\(I,V\)\(I,V\)pairs, so it supports this symbolic answer but no numerical slope or diameter ratio\.

Qwen3VL\-8B\-Thinking\.1\.The rollout reverses the axes and defines

SX=Δ​IXΔ​VX=1RX\.\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}S\_\{X\}=\\frac\{\\Delta I\_\{X\}\}\{\\Delta V\_\{X\}\}=\\frac\{1\}\{R\_\{X\}\}\}\.2\.It treats the drawn line widths as measurements and setsdB=2​dA\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}d\_\{B\}=2d\_\{A\}\}\.3\.Using those two unsupported premises, it computes

SBSA=\(dBdA\)2=4\\frac\{S\_\{B\}\}\{S\_\{A\}\}=\\left\(\\frac\{\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}d\_\{B\}\}\}\{d\_\{A\}\}\\right\)^\{2\}=4and reports

SB=4​SA\.\\boxed\{\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}S\_\{B\}=4S\_\{A\}\}\}\.The result follows only after reversing the plotted axes and measuring a schematic as though it were drawn to scale\.

![Refer to caption](https://arxiv.org/html/2608.04726v1/figures/case_prism_net.png)

Net of a trapezoidal prism\.The two end faces are congruent trapezia\. Their side lengths, in cyclic order, are 2, 2\.5, 5, and 2\.5 cm\. Letℓ\\elldenote the common prism length, which is not numerically specified in the photographed prompt\.

Our Model\.1\.Identify the end facesA​B​C​DABCDandH​E​F​GHEFG\. They satisfy

A​D=H​E=2,A​B=D​C=2\.5,B​C=G​F=5​cm\.AD=HE=2,\\quad AB=DC=2\.5,\\quad BC=GF=5\\ \\text\{cm\}\.2\.A prism has one lateral rectangle for every side of its cross\-section\. The four rectangles therefore have dimensions

2×ℓ,2\.5×ℓ,5×ℓ,2\.5×ℓ\.2\\times\\ell,\\quad 2\.5\\times\\ell,\\quad 5\\times\\ell,\\quad 2\.5\\times\\ell\.3\.Place the rectangles in the same cyclic order as the four trapezium edges\. Attach one congruent trapezium to an outer edge of the strip and the second to the corresponding edge on the opposite side\.4\.The labelled net is therefore

two congruent trapezia and four lateral rectangles\.\\boxed\{\\text\{two congruent trapezia and four lateral rectangles\}\}\.Several planar arrangements are valid, but every valid net has this face inventory and preserves the vertex correspondences\.

Qwen3VL\-8B\-Thinking\.1\.The rollout treats the 5\-cm base label as the prism length and then assumesevery edge is​5​cm\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{every edge is \}5\\text\{ cm\}\}\.2\.It replaces the stated trapezium witha​5​cm×5​cm square\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{a \}5\\text\{ cm\}\\times 5\\text\{ cm square\}\}\.3\.It consequently classifies the solid asa cube\.4\.Its final net is

six​5​cm×5​cm squares\.\\boxed\{\{\\color\[rgb\]\{0\.70703125,0\.13671875,0\.09375\}\\text\{six \}5\\text\{ cm\}\\times 5\\text\{ cm squares\}\}\}\.This rollout discards the visible 2\-cm and 2\.5\-cm labels and omits both trapezoidal end faces\.

Figure 11:Science and geometry rollouts\.Our Model uses the stated axes, dimensions, and topology; Qwen3VL\-8B\-Thinking propagates a specific diagram\-reading error to the final answer\.

## Appendix HScope, limitations, and broader impact

VTS is a controlled rendering protocol\. The held\-out natural\-page evaluation tests generalization across four real\-world layout categories, but its unpaired design does not measure a natural\-page semantic\-channel gap\. It also does not cover all languages, handwriting, severe occlusion, dynamic interfaces, or unknown task regions\. VISTA\-Bench supplies an independent matched visualized\-text protocol, whereas OCRBench\-style and document benchmarks test broader text\-rich capability rather than semantic\-channel equivalence\.

Prompt\-region supervision assumes a known task region during training\. The deployed model needs neither a region nor OCR, but the method is not itself a general task\-region detector\. The second\-backbone result supports transfer to InternVL3\.5\-8B under the tested setup; broader architecture and scale studies remain necessary\.

Stronger visual\-instruction following creates an additional trust boundary\. Our bounded safety check does not show higher visible or obfuscated attack success, but it is not a substitute for adversarial evaluation\. Systems should preserve explicit instruction priority, distinguish trusted from untrusted image content, and support refusal when channels conflict\. Any release of user\-contributed screenshots or photographs also requires consent, de\-identification, source and license records, duplicate checks, and a removal process\.

Finally, attention heatmaps and representation retrievals are descriptive\. They can reject simple explanations such as complete representation collapse, but they do not identify a unique internal mechanism\. Our mechanistic wording is therefore limited to the measured spatial, semantic\-target, and behavioral contrasts\.