WorksComputer Science
ResVision: Spectrum-Matched Resizing and Context-Preserving Inspection for Visual Reasoning
Abstract
Agentic visual reasoning improves vision-language models by selectively enlarging evidence without adding new scene content, but the benefit depends on whether fine structure remains accessible and whether local observations stay grounded in the full scene. We identify three recurring bottlenecks: high-frequency detail can be attenuated during visual-token formation; spatially distributed evidence can remain encoded yet difficult for the language model to retrieve; and isolated crops can weaken whole-detail correspondence through scale and positional mismatch. ResVision addresses these limitations with two complementary components. Spectrum-Matched Resizing (SMR) estimates image complexity and selects an adaptive global sampling scale, avoiding unnecessary enlargement on simple inputs while allocating more visual tokens to detail-rich images. Context-Preserving Inspection (CPI) returns each inspected crop together with the complete working view, preserving local detail and global context across tool turns. We evaluate Direct and Agentic variants on four frontier VLMs and 29 benchmarks spanning six visual-reasoning families. Complete ResVision (Agentic Base plus SMR plus CPI) is best for every model, improving by 3.36 to 4.94 points over Direct Base and 1.87 to 2.94 points over Agentic Base. These results show that adaptive visual sampling and context-preserving inspection address distinct, complementary constraints in multimodal reasoning.
Project Page: https://pevek.ai
Resources: GitHub · HuggingFace · ModelBlog
Figure 1. Accuracy across six benchmark families and four VLMs. ResVision denotes SMR in Direct mode and SMR+CPI in Agentic mode; values are three-run means.
Introduction
Agentic visual reasoning often improves the accuracy of vision-language models (VLMs) over direct inference through iterative visual inspection [3, 29, 46, 42, 18]. The model selects a region, enlarges it, examines the result, and decides where to inspect next. Yet these operations introduce no new scene content and merely change how existing pixels are presented to the model. This raises a fundamental question: how can changes in visual presentation substantially alter reasoning performance?
We investigate this puzzle through a systematic analysis of visual tokenization and its effects on downstream reasoning [13, 28, 12, 4, 36, 35, 5], identifying three mechanisms. Figure 2 illustrates the first two across Stages 1–3 of visual-token processing. When a full image is encoded, fine details may be attenuated within individual visual tokens, reflecting an Intra-Token Bandwidth Limit. By contrast, evidence distributed across neighboring tokens may remain encoded but difficult to retrieve during reasoning, resulting in a Contiguous-Token Retrieval Breakdown. Cropping and enlargement can alleviate both limitations by reallocating visual tokens to selected regions. However, image dimensions alone cannot determine the appropriate scale: a smooth photograph, a dense diagram, and a fine texture can have similar dimensions yet require different sampling densities [25, 44, 19].

Figure 2. Three mechanisms underlying the benefits and failures of agentic inspection. Fine details may be attenuated within visual tokens; distributed evidence may be difficult to retrieve; and isolated crops may weaken whole–detail correspondence. SMR targets evidence accessibility by adapting scale before tokenization, while CPI targets correspondence by retaining the complete working view with each inspected detail.
The third mechanism arises when a selected crop is re-encoded as an independent image at Stage 4. The crop has a new origin and scale, while surrounding objects, competing candidates, and spatial landmarks are excluded from the local view. This creates a potential Whole–Detail Reference Gap: the detail becomes easier to see but harder to locate within the full image, compare with alternatives, or interpret in context [32, 47]. A clearer crop can therefore lead to a less accurate answer than the complete image alone. Reliable agentic reasoning must make fine details accessible while preserving their correspondence to the full scene.
We introduce ResVision, a training-free framework that addresses these challenges through two complementary techniques. Spectrum-Matched Resizing (SMR) estimates each image’s fine-structure demand and selects the minimum global scale that places representative spatial frequencies within the visual-token passband under a bounded visual-token budget. Using the resulting working view, the VLM selects regions for closer inspection. Context-Preserving Inspection (CPI) returns each enlarged crop alongside the complete working view as separate, labeled image inputs. The crop provides local detail, while the complete view retains the spatial context and alternative evidence needed for interpretation. SMR adapts global sampling, the VLM selects where to inspect, and CPI maintains the full-image reference.
We evaluate these mechanisms through controlled diagnostics and downstream reasoning tasks. Mechanistic probes trace representations from patch projection to the positioned image blocks provided to the language model, revealing information attenuation, retrieval failure, and weakened whole–detail correspondence. Across four VLM families and 29 benchmarks, complete ResVision (Agentic Base+SMR+CPI) reaches a four-model mean Macro-29 of 75.93, improving by 3.88 points over Direct Base and 2.28 points over Agentic Base (VISTA). The gains hold for every model: 3.36–4.94 points over Direct Base and 1.87–2.94 points over Agentic Base. Focused component ablations on five diagnostic benchmarks further demonstrate the complementary contributions of SMR and CPI.
Our contributions are threefold. (1) We identify three mechanisms underlying agentic visual reasoning: Intra-Token Bandwidth Limit, Contiguous-Token Retrieval Breakdown, and Whole–Detail Reference Gap, explaining both the benefits and failure modes of local inspection. (2) We introduce SMR, which replaces fixed visual-token allocation with an image-adaptive rule that selects the minimum global scale needed to place representative spatial frequencies within the visual-token passband. (3) We introduce CPI, which pairs each selected crop with the complete working view, preserving a full-image reference for locating, comparing, and interpreting local evidence throughout multi-turn reasoning.
Diagnosing Visual-Token Bottlenecks
VLMs encode images into finite sequences of visual tokens before reasoning. We use Qwen 3.8-Flash-Next (Qwen 3.8) to trace visual evidence through tokenization and downstream reasoning, diagnosing three mechanisms along this path (Figure 2). Fine structure can be attenuated within individual tokens; evidence can remain encoded across neighboring tokens yet be difficult to retrieve; and an independently encoded crop can become harder to relate to the complete image. The first two mechanisms concern accessibility, explaining why changes in presentation can help without introducing new scene content. The third concerns correspondence, explaining why clearer local evidence can still yield a worse answer.
Intra-Token Bandwidth Limit
The first diagnostic examines whether the visual encoder preserves fine cues in a usable form. A narrow gap or small symbol can be present in the source image yet attenuated when compressed with surrounding texture during tokenization. This distinguishes information present in the image from information accessible to the model. Let denote the question, the complete input image, the answer, isotropic resizing by a factor , and mutual information. Deterministic resizing satisfies the data-processing bound
Different presentations of the same scene content can nevertheless yield different visual representations. Resizing changes how existing structure is sampled and allocated to visual tokens. For enlargement with , a spatial frequency is presented at . If the encoder response is non-increasing over , then
This motivates selecting a scale that moves fine structure into a frequency range the encoder represents more faithfully. We compare the Stage-1 patch projection with the Stage-3 final visual tokens using three probes (Figure 3). Amplitude retention. At Stage 3, decreases from 0.995 at DC to 0.447 at eight cycles per token, while the Stage-1 response remains substantially stronger. Linear recoverability. The recoverability cutoff of shifts from above 16 cycles per token at Stage 1 to 0.77 at Stage 3. Spatial-phase sensitivity. Opposite-phase gratings remain well separated at Stage 1 at the tested frequencies above two cycles per token, whereas their Stage-3 distance is nearly zero. Together, these probes characterize an effective visual-token bandwidth: fine structure present in the input becomes attenuated, less linearly recoverable, and harder to distinguish in the final visual representation.

Figure 3. Fine-structure accessibility across visual encoding stages. Probes on Qwen 3.8 compare Stage-1 patch projections with Stage-3 final visual tokens. (a) measures amplitude retention. (b) measures linear recoverability. (c) Opposite-phase distance measures sensitivity to spatial arrangement. Together, the probes show reduced amplitude retention, linear recoverability, and phase sensitivity at high frequencies after full visual encoding.
Contiguous-Token Retrieval Breakdown
Evidence can remain encoded across neighboring visual tokens yet be difficult to assemble during reasoning. We examine how a coherent pattern is represented when it crosses a token boundary, and whether the model can use information distributed across the complete token grid. Figure 4 presents the boundary and counting probes.

Figure 4. Boundary sensitivity and count retrieval across visual tokens. (a) A fixed pattern moves from a token center toward its boundary. The position-normalized response is mirrored about the boundary, showing sensitivity to token grid alignment. (b) A trained linear readout recovers all evaluated counts from the frozen Stage-3 token grid, while the language model’s counting accuracy drops at and above one object per token.
Token-boundary response. We move the same pattern from a token center toward its boundary while holding appearance and scale fixed. Figure 4(a) plots the position-normalized response and mirrors the sweep about the boundary. At Stage 1, similarity to the token-centered pattern drops sharply as the pattern is divided across neighboring tokens. Stage-3 contextualization retains more of the cue, but the representation remains sensitive to grid alignment. The same structure is therefore represented differently when it lies within a token region and when it straddles a boundary. Resizing changes where structures fall on the token grid, linking presentation scale to this boundary sensitivity.
Encoded count versus retrieved count. We place objects across horizontal token positions. For each , we generate 60 strips, using 45 to train a diagnostic readout and 15 for evaluation. The VLM remains frozen. Its complete Stage-3 token grid is concatenated in spatial order, and a ridge-regularized linear classifier is trained to predict . We present the same evaluation images to the language model and measure exact-count accuracy and the probability assigned to the correct integer. The results are shown in Figure 4(b).
On the evaluation set, the trained readout recovers every count from the frozen Stage-3 grid. The language model’s exact-count accuracy is 93.3–100.0% for , falls to 26.7% at , and reaches 0% above that density. This readout–generation gap distinguishes encoded count information from the model’s ability to retrieve, organize, and enumerate it during generation. Together, the probes characterize a contiguous-token retrieval bottleneck and motivate adjusting presentation scale to improve how evidence is distributed over the token grid.
Whole–Detail Reference Gap
The third diagnostic examines whether an enlarged detail remains correctly associated with its source scene. Crop-and-zoom can alleviate the first two bottlenecks, but Qwen 3.8 processes the crop as a separate image at Stage 4, with its own scale and 2D rotary position frame. Surrounding landmarks and competing evidence are excluded from this local view. The resulting representation therefore changes both the visibility of the detail and its relationship to the complete image.
To describe the coordinate change, let crop have top-left source coordinate , let be a point’s offset within , and let be the linear scale factor. In continuous image coordinates, the same source point has whole-image and crop-local coordinates
The crop resets the image origin and changes the source-pixel stride to per output pixel. Position IDs specify locations within each image, while whole–detail correspondence requires relating the crop frame to the source frame. The model must establish this relation across separate image blocks and interaction turns.
We evaluate this correspondence on 50 BabyVision images. For each image, a selected crop is presented at scales from to , and the model is asked to ground its detail in the complete scene. Figure 5 compares three conditions: an isolated crop under scale mismatch, an isolated crop after positional reset, and the crop presented jointly with the complete image. At each scale, the local crop content is held fixed across conditions. Joint presentation maintains 72% grounding accuracy across all tested scales, whereas isolated
Figure 5. Whole–detail grounding with and without the complete image. On 50 BabyVision images, pairing the crop with the complete image maintains 72% grounding accuracy across the tested scales from 1× to 4×. Isolated crops achieve 40–58% accuracy at 1.5×–4×.
crops achieve 40–58% at 1.5×–4×. The complete view supplies landmarks, alternative candidates, and a spatial reference for interpreting the enlarged detail. A complementary representation-level diagnostic in the supplementary material shows that identical Stage-1 patch features diverge after contextual encoding and crop-local position assignment. These results motivate presenting every selected detail alongside the complete working view to maintain whole–detail correspondence.

Figure 6. The ResVision reasoning loop. Spectrum-Matched Resizing selects an image-adaptive global scale and constructs a fixed working view. The VLM selects a region for closer inspection, and Context-Preserving Inspection returns its enlarged crop alongside the complete working view as separate, labeled image inputs. The VLM uses both views to answer or select the next region.
Method
The diagnosis motivates two design requirements: adapt visual sampling to image structure and preserve the scene reference for each local observation. ResVision implements these requirements in a training-free framework. Before inference, Spectrum-Matched Resizing (SMR) constructs a fixed, image-adaptive working view. The VLM reasons over this view and selects regions for closer inspection. Context-Preserving Inspection (CPI) returns each enlarged crop alongside the same complete working view. The components act at distinct points along the evidence path: SMR controls how scene structure is sampled before tokenization; agentic inspection concentrates evidence in selected regions; and CPI keeps each local observation jointly available with its source-scene landmarks.
Spectrum-Matched Resizing
SMR is designed to bring representative image frequencies into the encoder passband at minimum visual-token cost, before Stage-1 sampling. For an input image of width and height , let be the effective pixel stride per Stage-3 visual token and a spatial frequency in cycles per source pixel. After isotropic resizing by a factor , the same pattern has token-relative frequency
in cycles per visual token. Its visual-token cost is
Increasing lowers the token-relative frequency but increases token cost quadratically. SMR uses these relations to select a scale from image structure under a visual-token budget.
Active-region spectral estimate. We compute the grayscale gradient magnitude, partition it into tiles, and retain the fraction with the highest gradient energy. These active tiles emphasize structured regions even when they occupy a small part of the image. After subtracting each retained tile’s mean, we aggregate its non-DC Fourier power into a radial spectrum , where is radial spatial frequency in cycles per source pixel. For nonzero total power and an energy-quantile level , we define
Smooth content yields a lower , while dense diagrams and fine textures yield a higher value. The energy quantile summarizes the dominant structural frequencies and reduces sensitivity to isolated high-frequency components from noise or compression artifacts. All images use the same estimation procedure.
Minimum-cost frequency matching. Let denote the target frequency limit in cycles per visual token. The frequency-matching objective is
whose solution is . For an admissible token interval with , the scale bounds are
We clip the target scale to :
The clipping step constrains the target scale to the admissible range. The lower bound maintains a minimum sampling resolution, while the upper bound controls token use. Smooth content remains compact, whereas dense diagrams, text, and fine textures receive more sampling capacity. Images with zero non-DC power use .
A fixed working image. The selected scale defines the working image , with dimensions aligned to the backend grid and aspect ratio preserved up to rounding. The initial input, every crop, pixel measurement, and repeated complete view are derived from . Tool coordinates use this same working-image frame throughout the interaction. For multiple input images, their order is preserved under a shared total visual-token budget. When SMR is disabled, and the backend applies its standard preprocessing.
Context-Preserving Inspection
CPI returns the fixed working view with every newly rendered crop at Stage 4, placing the source scene alongside the local observation. At inspection step , the VLM selects a region in . The shared detail renderer produces
where denotes the crop of image inside box , and is the local enlargement operator.
For an image , let denote its backend-encoded image-token block. CPI supplies the complete working view and the enlarged crop as separate, labeled image inputs, giving the visual reply
where denotes ordered concatenation of labeled image blocks. The backend assigns each image block its standard image positions. Joint presentation makes the enlarged detail and its grounding landmarks available within the same input. For multi-image questions, denotes the complete ordered set of working-image blocks, including images not selected for inspection.
The same local appearance can support different answers in different scenes. Let be the correct answer to question for image . Two scenes can satisfy
The complete view supplies the scene information that distinguishes these cases and associates the visible cue with its source. In CPI, both views originate from the same fixed : the complete view provides the working coordinates and scene layout, while the crop exposes the selected region at greater detail.
Let denote the visual input history after crop inspections. The initial view and successive replies accumulate as
Each new detail thus arrives with a current scene reference. Earlier crops remain available for comparison, while the repeated complete view provides the context needed to establish the detail’s identity, position, scale, and meaning. Pixel-reading requests return textual color values sampled from .
Experiments
We first evaluate two inference modes on a broad 29-benchmark suite. Direct mode compares Direct with Direct+ResVision, where ResVision contributes its SMR front end without tools. Agentic mode compares Agentic with Agentic+ResVision, where ResVision contributes the complete SMR+CPI system. Intermediate Agentic variants are retained as component ablations rather than main methods. We then use the original five-benchmark diagnostic slice for focused allocation, robustness, and latency analyses.
Experimental Setup
The main evaluation combines 26 MM400-aligned benchmarks with BabyVision [7], VLMsAreBiased [40], and BLINK [14]. The resulting 29 benchmarks span mathematical and logical reasoning, chart and OCR understanding, hallucination and counting, embodied and spatial reasoning, medical and world knowledge, and general visual understanding. We report Macro-29, the equal-weight mean of the 29 benchmark accuracies, so large datasets do not dominate the aggregate.
We evaluate four frontier VLMs: Kimi-K3 [30], Qwen3.8-Flash-Next [33], GLM-5.3-Flash [48], and DeepSeek-V4.1-Flash-1017 [11]. Direct uses no visual tools, while Direct+ResVision adds SMR. Agentic is VISTA [18]; Agentic+ResVision combines the same inspection loop with SMR and CPI. Both Agentic configurations share the same four-call tool ceiling. Every configuration uses temperature 0, the model’s maximum reasoning-effort setting, a 65,536-token generation ceiling, and three independent runs.
Main Results on 29 Benchmarks
Agentic+ResVision is best for every model. It reaches 76.90 on Qwen3.8, 79.02 on Kimi-K3, 71.97 on DeepSeek, and 75.83 on GLM (Table 1). Relative to Direct, the respective gains are 4.94, 3.36, 3.57, and 3.67 points; relative to Agentic, they are 2.94, 2.42, 1.87, and 1.88 points. Averaged equally across models, the complete system scores 75.93, compared with 72.05 for Direct and 73.65 for Agentic—gains of 3.88 and 2.28 points, respectively.
| Model | Direct | Direct+ResVision | Agentic | Agentic+ResVision |
| Qwen3.8-Flash-Next | 71.96 | 74.72 | 73.96 | 76.90 |
| Kimi-K3 | 75.66 | 77.30 | 76.60 | 79.02 |
| DeepSeek-V4.1-Flash | 68.40 | 70.02 | 70.10 | 71.97 |
| GLM-5.3-Flash | 72.16 | 73.92 | 73.95 | 75.83 |
| Four-model mean | 72.05 | 73.99 | 73.65 | 75.93 |
Table 1. Macro-29 accuracy under the four main conditions. Direct+ResVision enables SMR; Agentic+ResVision enables the complete SMR+CPI system. Each entry is the equal-weight mean over 29 benchmarks and three runs.
Both presentation and inspection contribute consistently. Direct+ResVision raises the four-model mean from 72.05 to 73.99, with per-model gains of 1.62–2.76 points, demonstrating that adaptive global presentation helps even without tool use. Agentic Base itself averages 73.65, 1.61 points above Direct, while complete ResVision adds another 2.28 points over Agentic Base. Thus neither a better initial presentation nor inspection alone explains the final result. The focused component analysis below separates the intermediate Agentic+SMR and CPI-only variants.
The two components remain complementary. Without SMR, CPI improves Agentic Base by 1.10–1.97 points (Table 2); Agentic+SMR and CPI-only average 75.10 and 75.07, respectively. Complete ResVision reaches 75.93 and exceeds the stronger single-component variant by 0.40–0.94 points in every model. This supports the intended interpretation of ResVision as Agentic Base+SMR+CPI rather than either component in isolation.
| Model | Agentic Base | +SMR | +CPI | +SMR+CPI |
| Qwen3.8-Flash-Next | 73.96 | 75.96 | 75.06 | 76.90 |
| Kimi-K3 | 76.60 | 78.27 | 78.57 | 79.02 |
| DeepSeek-V4.1-Flash | 70.10 | 70.87 | 71.57 | 71.97 |
| GLM-5.3-Flash | 73.95 | 75.30 | 75.09 | 75.83 |
Table 2. CPI-only component ablation. Macro-29; complete ResVision enables SMR+CPI.
The gains are broad but not uniform across tasks. Complete ResVision is the unique best main condition on 21 of 29 benchmarks for Qwen3.8, 16 for Kimi-K3, and 15 each for DeepSeek and GLM. Relative to Agentic Base, its largest family-level gains occur in Math & Logic (4.59–8.34 points across models) and Hallucination & Count (2.08–6.19 points). Medical & World changes by 0.42–2.19 points, while General Visual is comparatively flat (−0.30 to +1.07). Macro-29 therefore provides a conservative test: it includes many tasks with weak sensitivity to visual-token allocation, yet the aggregate gain remains positive for every model.
Benchmark-by-Benchmark Scorecard
Table 3 expands the four main presentation conditions into all 29 benchmark accuracies. Each panel compares Direct, Direct+ResVision, Agentic, and Agentic+ResVision; the Agentic+SMR intermediate remains in the component summary above. Pale-blue cells mark the best main condition for each benchmark within a model. The scorecard makes the breadth of the evaluation visible while also showing that no single method dominates every individual task.
| Qwen3.8-Flash-Next | Kimi-K3 | ||||||||
| Benchmark | Direct Base | Direct +ResVision | Agentic Base | Agentic +ResVision | Benchmark | Direct Base | Direct +ResVision | Agentic Base | Agentic +ResVision |
| Math & Logic | Math & Logic | ||||||||
| BabyVision | 47.16 | 70.88 | 56.61 | 77.75 | BabyVision | 55.07 | 74.05 | 59.37 | 80.76 |
| VisuLogic | 39.42 | 52.33 | 48.67 | 58.17 | VisuLogic | 47.08 | 60.08 | 50.33 | 65.67 |
| Kangaroo2025 | 86.75 | 90.92 | 91.67 | 94.08 | Kangaroo2025 | 94.00 | 94.92 | 95.58 | 96.25 |
| MathVision | 92.50 | 93.67 | 92.83 | 95.25 | MathVision | 95.33 | 95.67 | 94.00 | 96.00 |
| Rebus | 50.45 | 51.25 | 58.06 | 55.96 | Rebus | 81.98 | 84.58 | 83.78 | 86.09 |
| Chart & OCR | Chart & OCR | ||||||||
| OCR-Reasoning | 76.83 | 77.08 | 77.25 | 80.08 | OCR-Reasoning | 79.50 | 81.75 | 79.83 | 80.75 |
| CharXiv-RQ | 82.00 | 85.83 | 84.92 | 88.67 | CharXiv-RQ | 85.50 | 87.83 | 89.83 | 90.42 |
| ChartMuseum | 82.42 | 82.92 | 83.75 | 86.92 | ChartMuseum | 83.42 | 85.75 | 85.42 | 89.50 |
| CharXiv-Des | 95.58 | 96.25 | 95.33 | 96.75 | CharXiv-Des | 95.58 | 95.58 | 96.00 | 95.75 |
| Hallucination & Count | Hallucination & Count | ||||||||
| VLMsAreBiased | 62.17 | 75.00 | 71.58 | 88.75 | VLMsAreBiased | 46.17 | 48.75 | 53.58 | 60.50 |
| HallusionBench | 85.75 | 88.08 | 87.00 | 89.00 | HallusionBench | 84.33 | 83.17 | 85.58 | 84.58 |
| CountQA | 77.33 | 75.67 | 74.08 | 73.50 | CountQA | 69.92 | 68.92 | 71.83 | 72.17 |
| Embodied & Spatial | Embodied & Spatial | ||||||||
| BLINK | 81.00 | 85.75 | 83.33 | 86.00 | BLINK | 84.83 | 88.00 | 87.33 | 89.25 |
| MMSI | 51.08 | 53.58 | 52.83 | 53.50 | MMSI | 57.00 | 55.83 | 55.25 | 55.58 |
| ERQA | 69.08 | 70.83 | 69.33 | 74.67 | ERQA | 69.25 | 71.17 | 69.67 | 73.17 |
| MindCube | 81.83 | 83.25 | 84.58 | 86.25 | MindCube | 89.33 | 88.67 | 87.75 | 88.17 |
| CV-Bench | 91.58 | 92.58 | 91.25 | 92.92 | CV-Bench | 92.92 | 92.33 | 93.25 | 91.67 |
| OmniSpatial | 76.75 | 78.50 | 78.17 | 79.17 | OmniSpatial | 75.92 | 74.67 | 77.08 | 76.00 |
| Medical & World | Medical & World | ||||||||
| MuirBench | 79.17 | 80.83 | 78.75 | 79.25 | MuirBench | 77.42 | 79.08 | 77.25 | 77.83 |
| MedXpertQA-MM | 69.33 | 68.92 | 71.25 | 71.08 | MedXpertQA-MM | 75.17 | 74.92 | 75.17 | 76.42 |
| WorldVQA | 27.67 | 28.50 | 30.75 | 32.08 | WorldVQA | 50.17 | 47.25 | 46.92 | 51.83 |
| RealWorldQA | 87.50 | 87.92 | 88.17 | 88.17 | RealWorldQA | 83.25 | 83.17 | 83.17 | 85.17 |
| General Visual | General Visual | ||||||||
| SimpleVQA | 57.92 | 59.42 | 57.75 | 61.25 | SimpleVQA | 73.75 | 71.08 | 73.17 | 70.67 |
| MMMU | 84.50 | 84.83 | 84.83 | 85.50 | MMMU | 86.42 | 87.00 | 85.75 | 87.08 |
| MMMU-Pro | 78.67 | 79.58 | 78.58 | 80.00 | MMMU-Pro | 82.00 | 83.92 | 82.50 | 83.42 |
| MMStar | 82.08 | 83.92 | 82.92 | 84.33 | MMStar | 80.08 | 80.33 | 81.67 | 81.67 |
| MMBench-CN | 93.92 | 93.25 | 94.17 | 92.25 | MMBench-CN | 93.42 | 93.75 | 93.08 | 93.67 |
| CCBench | 84.92 | 84.33 | 86.75 | 86.17 | CCBench | 91.92 | 91.17 | 91.92 | 91.17 |
| ZeroBench | 11.33 | 11.00 | 9.67 | 12.67 | ZeroBench | 13.33 | 18.33 | 15.33 | 20.33 |
| Macro-29 | 71.96 | 74.72 | 73.96 | 76.90 | Macro-29 | 75.66 | 77.30 | 76.60 | 79.02 |
Table 3. Benchmark-by-benchmark Macro-29 scorecard across four frontier VLMs. Values are mean accuracy (%) over three independent runs. Direct+ResVision uses SMR, while Agentic+ResVision uses SMR+CPI; the best main condition in each benchmark row is highlighted.
| Benchmark | DeepSeek-V4.1-Flash Direct Base | DeepSeek-V4.1-Flash Direct +ResVision | DeepSeek-V4.1-Flash Agentic Base | DeepSeek-V4.1-Flash Agentic +ResVision | Benchmark | GLM-5.3-Flash Direct Base | GLM-5.3-Flash Direct +ResVision | GLM-5.3-Flash Agentic Base | GLM-5.3-Flash Agentic +ResVision |
| Math & Logic | Math & Logic | ||||||||
| BabyVision | 33.51 | 48.45 | 42.61 | 59.36 | BabyVision | 40.46 | 60.05 | 51.63 | 68.21 |
| VisuLogic | 33.00 | 42.67 | 39.08 | 46.67 | VisuLogic | 39.67 | 45.75 | 48.33 | 54.92 |
| Kangaroo2025 | 81.83 | 84.00 | 87.92 | 88.25 | Kangaroo2025 | 88.50 | 87.67 | 91.50 | 91.42 |
| MathVision | 87.83 | 89.67 | 89.67 | 92.00 | MathVision | 84.92 | 90.08 | 90.67 | 93.92 |
| Rebus | 64.86 | 64.26 | 65.57 | 64.66 | Rebus | 73.57 | 72.47 | 69.77 | 66.37 |
| Chart & OCR | Chart & OCR | ||||||||
| OCR-Reasoning | 75.58 | 74.58 | 76.17 | 78.08 | OCR-Reasoning | 75.83 | 77.17 | 76.33 | 77.75 |
| CharXiv-RQ | 75.92 | 80.17 | 79.92 | 84.17 | CharXiv-RQ | 82.58 | 85.08 | 86.58 | 88.08 |
| ChartMuseum | 74.33 | 74.83 | 78.33 | 80.50 | ChartMuseum | 81.00 | 80.17 | 83.42 | 87.58 |
| CharXiv-Des | 95.58 | 96.08 | 95.83 | 96.25 | CharXiv-Des | 95.08 | 95.83 | 95.75 | 95.83 |
| Hallucination & Count | Hallucination & Count | ||||||||
| VLMsAreBiased | 43.17 | 50.17 | 51.58 | 60.17 | VLMsAreBiased | 48.33 | 56.58 | 63.42 | 69.92 |
| HallusionBench | 84.92 | 85.92 | 87.08 | 86.75 | HallusionBench | 84.67 | 85.42 | 86.00 | 86.25 |
| CountQA | 61.92 | 63.08 | 63.00 | 63.75 | CountQA | 60.17 | 61.08 | 64.00 | 66.25 |
| Embodied & Spatial | Embodied & Spatial | ||||||||
| BLINK | 79.25 | 81.83 | 79.00 | 81.50 | BLINK | 83.33 | 85.83 | 82.92 | 86.50 |
| MMSI | 48.33 | 49.08 | 50.17 | 48.83 | MMSI | 54.17 | 56.33 | 50.75 | 52.92 |
| ERQA | 64.33 | 66.33 | 65.33 | 70.08 | ERQA | 67.08 | 71.25 | 68.75 | 72.17 |
| MindCube | 63.17 | 65.33 | 60.92 | 62.50 | MindCube | 82.42 | 84.67 | 81.17 | 81.67 |
| CV-Bench | 90.50 | 90.33 | 91.00 | 89.92 | CV-Bench | 92.75 | 91.50 | 91.67 | 92.33 |
| OmniSpatial | 73.67 | 73.08 | 71.83 | 73.67 | OmniSpatial | 74.17 | 73.67 | 73.08 | 76.33 |
| Medical & World | Medical & World | ||||||||
| MuirBench | 75.17 | 76.58 | 74.67 | 76.08 | MuirBench | 78.92 | 78.67 | 80.25 | 78.42 |
| MedXpertQA-MM | 70.25 | 68.67 | 69.83 | 69.92 | MedXpertQA-MM | 73.25 | 72.08 | 72.75 | 74.67 |
| WorldVQA | 32.42 | 31.83 | 31.33 | 32.83 | WorldVQA | 44.75 | 41.83 | 41.25 | 43.83 |
| RealWorldQA | 83.25 | 81.33 | 84.08 | 85.25 | RealWorldQA | 86.58 | 85.58 | 85.58 | 86.50 |
| General Visual | General Visual | ||||||||
| SimpleVQA | 62.17 | 61.83 | 64.17 | 62.33 | SimpleVQA | 65.25 | 66.33 | 67.83 | 66.50 |
| MMMU | 84.25 | 85.25 | 85.42 | 85.25 | MMMU | 84.33 | 85.25 | 86.50 | 85.92 |
| MMMU-Pro | 77.25 | 77.42 | 79.33 | 78.00 | MMMU-Pro | 77.17 | 79.00 | 81.00 | 79.75 |
| MMStar | 79.75 | 79.75 | 80.08 | 80.58 | MMStar | 80.92 | 82.33 | 82.08 | 82.92 |
| MMBench-CN | 92.83 | 93.75 | 93.17 | 93.75 | MMBench-CN | 94.25 | 94.58 | 93.75 | 94.08 |
| CCBench | 86.50 | 84.58 | 87.17 | 85.00 | CCBench | 88.42 | 87.17 | 87.25 | 86.33 |
| ZeroBench | 8.00 | 9.67 | 8.67 | 11.00 | ZeroBench | 10.00 | 10.33 | 10.67 | 11.67 |
| Macro-29 | 68.40 | 70.02 | 70.10 | 71.97 | Macro-29 | 72.16 | 73.92 | 73.95 | 75.83 |
Table 3 (PDF p. 11). Benchmark-by-benchmark Macro-29 scorecard across four frontier VLMs (continued).
Focused Component Analysis
The following controlled analyses retain the original five-benchmark slice—BabyVision, VLMsAreBiased, BLINK, VStarBench [42], and MMVP [39]—because it provides dense three-run measurements for all ablations and exposes the visual bottlenecks studied in Sec. 2. Macro-5 is the equal-weight mean across these five benchmarks.
Table 4 isolates SMR and CPI on Kimi-K3 and Qwen3.8. SMR alone improves Direct by 5.26 and 8.88 points, respectively. Within agentic inference, CPI adds 4.15 and 5.39 points without SMR, and a further 2.21 and 1.85 points when SMR is enabled. The complete system exceeds Direct+SMR by 4.46 and 5.66 points, showing that adaptive presentation and context-preserving local inspection contribute at different stages.
| Configuration | Kimi-K3 | Qwen 3.8 |
| Direct | ||
| Base | ||
| +SMR | ||
| Agentic | ||
| Base | ||
| +CPI | ||
| +SMR | ||
| +SMR+CPI |
Table 4. Component ablations. Macro-5 mean SD over three runs.
Adaptive versus Fixed Visual-Token Allocation
We replace SMR with fixed targets of 1,024–5,120 visual tokens while retaining CPI and all other settings. Accuracy is non-monotonic in the fixed budget: the strongest fixed target is 2,048 tokens for both tested models, while adaptive SMR improves on it by 3.31 points on Kimi-K3 and 3.36 points on Qwen3.8 (Table 5). This supports allocating sampling density from image structure rather than using a uniform token target.
| Allocation | Kimi-K3 | Qwen 3.8 |
| Original | ||
| Fixed 1,024 | ||
| Fixed 2,048 | ||
| Fixed 3,072 | ||
| Fixed 4,096 | ||
| Fixed 5,120 | ||
| SMR (adaptive) |
Table 5. Adaptive versus fixed allocation. Macro-5 mean SD over three runs.
Inference-Configuration Sensitivity
Table 6 varies one inference setting at a time on BabyVision while retaining complete ResVision. Reasoning effort has the largest effect: moving from Low to Max raises accuracy by 22.68 points on Kimi-K3 and 8.68 points on Qwen3.8. Temperature changes produce a spread of at most 3.01 points, and increasing the generation ceiling beyond 32K does not consistently improve accuracy.
| Factor | Setting | Kimi-K3 | Qwen 3.8 |
| Temperature | 0 | ||
| 0.5 | |||
| 1.0 | |||
| Effort | Low | ||
| Medium | |||
| Max | |||
| Token cap | 32K | ||
| 65K | |||
| 131K |
Table 6. Configuration sensitivity on BabyVision (%). One setting varies at a time.
Inference-Time Cost
Table 7 reports mean end-to-end client latency over the focused five-benchmark slice. Complete ResVision takes 1.07–1.36× as long as Direct. Relative to VISTA, it is 2.8% faster on Qwen3.8 and 3.5–9.7% slower on the other three models. These measurements show that the accuracy gains do not require a large additional latency multiplier beyond agentic inspection itself.
| Model | Direct | VISTA | ResVision | |||
| Kimi-K3 | 248.5 | 264.8 | 274.2 | |||
| Qwen | 51.5 | 68.5 | 66.6 | |||
| GLM | 193.6 | 240.7 | 263.6 | |||
| DeepSeek | 100.4 | 97.7 | 107.2 | |||
| Mean ratio |
Table 7. End-to-end latency. Gray cells give seconds; blue cells give the ratio to Direct.
Related Work
Visual Scaling and Token Allocation
Modern VLMs inherit patch-based visual interfaces from ViTs and combine them with contrastive or instruction-tuned multimodal learning [34, 1, 24, 8, 10, 26, 2, 9]. A substantial line of work makes this interface more flexible. NaViT packs variable-resolution inputs, Pix2Struct preserves aspect ratio under a token budget, and FlexiViT supports multiple patch sizes [12, 23, 4]; TokenLearner, DynamicViT, and Token Merging instead adapt, prune, or merge representations within the encoder [36, 35, 5]. High-resolution VLMs divide native images into tiles or combine low- and high-resolution branches to expose dense detail [25, 44, 19]. These designs demonstrate that visual-token allocation strongly affects what a model can recover. However, they usually determine allocation from image geometry, a fixed resolution policy, or importance estimated after tokenization. The same pixel dimensions can contain very different amounts of fine structure, so geometry alone need not predict the sampling required by a particular image. ResVision addresses this complementary question before reasoning begins: SMR measures image structure and selects a budgeted working scale that matches it to the encoder’s usable frequency range, without retraining or modifying the visual backbone.
Agentic Visual Reasoning
Agentic visual reasoning extends active perception by allowing a model to gather evidence as its reasoning evolves [3, 29, 46]. Program- and tool-based systems compose specialist modules [17, 38, 45, 41, 27], while V* and VISTA let a general-purpose VLM actively search or retrieve visual observations [42, 18]. Related methods ground language in selected regions, introduce visual prompts, or use drawings and interleaved visual states to support reasoning [6, 32, 47, 37, 15, 20]. Web and desktop agents further show that perception and grounding remain limiting factors when visual observations drive actions [49, 22, 43], a capability directly examined by VisualToolBench [16]. Most of this literature focuses on choosing an action, invoking the right tool, or obtaining a sharper local observation. It gives less attention to how the returned observation should remain connected to the scene from which it was extracted. Visual Funnel is closely related in combining focal evidence with multiscale context [21], but relies on attention-guided views rather than maintaining a stable complete reference across an agentic trajectory. ResVision’s CPI targets this whole–detail relationship directly: each selected crop is presented together with the same complete working view, so subsequent reasoning can locate the detail, compare it with alternatives, and interpret it in context.
Conclusion
We identify three mechanisms that explain how local inspection can improve or impair visual reasoning by altering evidence accessibility and scene correspondence without introducing new scene content. ResVision translates these insights into a training-free framework: SMR improves visual accessibility by adapting the working scale to image structure, while CPI preserves whole–detail correspondence throughout local inspection. Across four VLM families and 29 benchmarks, complete ResVision (Agentic Base+SMR+CPI) is the strongest main configuration for every model, reaching a four-model mean Macro-29 of 75.93—3.88 points above Direct Base and 2.28 points above Agentic Base. The gains are largest on reasoning families that depend strongly on fine structure and distributed evidence, while flatter families expose the expected limits of additional visual sampling. Focused ablations nevertheless show that SMR and CPI each contribute and that their combination is consistently strongest. These findings highlight accessibility and correspondence as two fundamental requirements for agentic visual reasoning, paving the way for visual agents that can perceive fine details and interpret them within the broader scene.
LLM Usage Statement
GPT assisted with early-stage ideation, language editing, conceptual figures, statistical analysis. The human authors are responsible for the method design, experimental procedure, source verification, citations, scientific claims, and final manuscript.
Authors
Guojian Zhan, Jiahao Xu, Yahui Liu, Yao Hu, and Chuan Mu.
Evaluation Setup
Model Families and Endpoints
Table 8 records the four evaluated endpoints and their main inference settings. Max denotes the highest supported reasoning effort for each model.
| Model family | Evaluated endpoint identifier | Main effort | Temperature | Max tokens |
| Kimi | Kimi-K3 | Max | 0 | 65,536 |
| Qwen | Qwen3.8-Flash-Next | Max | 0 | 65,536 |
| GLM | GLM-5.3-Flash | Max | 0 | 65,536 |
| DeepSeek | DeepSeek-V4.1-Flash-1017 | Max | 0 | 65,536 |
Table 8. Model endpoints and main inference settings.
Shared Evaluation Controls
Direct answers from the original image without visual tools. VISTA and ResVision use the same backend, task input, prompts, tool implementation, temperature, reasoning effort, and generation ceiling. Both agentic methods have a hard ceiling of four executed calls. ResVision applies SMR before inference and uses CPI to return the complete working view alongside each crop. Main runs use temperature 0, Max effort, and a generation ceiling of 65,536 output tokens.
We evaluate each method in three independent runs per setting and report the mean and sample standard deviation across runs. For paired comparisons, we first average each method’s score on each question across the three runs, then compute 95% confidence intervals by resampling questions within each benchmark, using the same bootstrap samples for all methods. Macro-5 is the unweighted mean of the five benchmark accuracies.
Benchmark Profiles
The five benchmarks span structured diagrams, counterfactual patterns, cross-view correspondence, high-resolution search, and deceptively similar natural images. Table 9 gives the evaluated snapshot sizes, totaling 1,679 questions. Figure 7 illustrates their contrasting visual demands.
| Snapshot | Questions | Primary capability covered in this evaluation |
| BabyVision | 388 | Foundational discrimination, tracking, spatial, and pattern reasoning |
| VLMsAreBiased | 400 | Visual evidence under conflict with learned semantic priors |
| BLINK | 400 | Core visual perception recast from classical computer-vision tasks |
| VStarBench | 191 | Search and grounding in high-resolution, visually crowded scenes |
| MMVP | 300 | Basic visual patterns that expose representation-level shortcomings |
| Total | 1,679 |
Table 9. Evaluation benchmarks and snapshot sizes.
Figure 7. Figure 7: Five benchmarks, five visual demands. Real image–question examples from our evaluation snapshots, with reference answers. Task text is excerpted to omit answer-format instructions and, for BLINK, annotation scaffolding.
Complete Results and Run-Level Stability
Table 10 reports benchmark-level means and sample standard deviations; Gain is the Macro-5 improvement over Direct in percentage points. Table 11 gives the three individual runs behind each aggregate.
| Method | Baby | Bias | BLINK | VStar | MMVP | Macro-5 | Gain |
| Kimi-K3 | |||||||
| Direct | 0.00 | ||||||
| VISTA | +3.33 | ||||||
| ResVision | +9.73 | ||||||
| Qwen3.8-Flash-Next | |||||||
| Direct | 0.00 | ||||||
| VISTA | +5.52 | ||||||
| ResVision | +14.54 | ||||||
| GLM-5.3-Flash | |||||||
| Direct | 0.00 | ||||||
| VISTA | +5.76 | ||||||
| ResVision | +11.91 | ||||||
| DeepSeek-V4.1-Flash-1017 | |||||||
| Direct | 0.00 | ||||||
| VISTA | +5.64 | ||||||
| ResVision | +11.89 |
Table 10. Five-benchmark results across four VLMs. Each entry is the mean sample standard deviation across three independent runs. ResVision uses the complete SMR+CPI configuration.
| Method | Run | BabyVision | VLMsAreBiased | BLINK | VStarBench | MMVP | Macro-5 | Correct |
| Kimi-K3 | ||||||||
| Direct | 1 | 54.90 | 45.75 | 84.25 | 89.53 | 85.00 | 71.89 | 1159 |
| 2 | 55.15 | 46.25 | 84.75 | 90.58 | 85.00 | 72.35 | 1166 | |
| 3 | 55.15 | 46.50 | 85.50 | 90.58 | 85.67 | 72.68 | 1172 | |
| VISTA | 1 | 57.99 | 52.75 | 86.75 | 91.62 | 85.67 | 74.96 | 1215 |
| 2 | 59.54 | 53.75 | 86.75 | 91.62 | 86.00 | 75.53 | 1226 | |
| 3 | 60.57 | 54.25 | 88.50 | 92.15 | 86.67 | 76.43 | 1242 | |
| ResVision | 1 | 81.96 | 61.50 | 89.75 | 93.19 | 87.33 | 82.75 | 1363 |
| 2 | 80.41 | 60.75 | 89.00 | 93.19 | 87.00 | 82.07 | 1350 | |
| 3 | 79.90 | 59.25 | 89.00 | 91.62 | 86.67 | 81.29 | 1338 | |
| Qwen3.8-Flash-Next | ||||||||
| Direct | 1 | 46.91 | 61.75 | 80.50 | 91.62 | 85.00 | 73.16 | 1181 |
| 2 | 47.16 | 62.25 | 81.25 | 91.62 | 85.33 | 73.52 | 1188 | |
| 3 | 47.42 | 62.50 | 81.25 | 92.15 | 85.33 | 73.73 | 1191 | |
| VISTA | 1 | 56.44 | 70.50 | 82.50 | 94.76 | 86.67 | 78.17 | 1272 |
| 2 | 56.70 | 71.50 | 83.00 | 96.34 | 87.67 | 79.04 | 1285 | |
| 3 | 56.70 | 72.75 | 84.50 | 96.86 | 88.00 | 79.76 | 1298 | |
| ResVision | 1 | 78.87 | 89.50 | 86.75 | 96.34 | 91.67 | 88.63 | 1470 |
| 2 | 77.58 | 89.00 | 85.75 | 95.29 | 91.67 | 87.86 | 1457 | |
| 3 | 76.80 | 87.75 | 85.50 | 96.34 | 91.33 | 87.54 | 1449 | |
| GLM-5.3-Flash | ||||||||
| Direct | 1 | 39.69 | 47.75 | 83.00 | 89.53 | 83.67 | 68.73 | 1099 |
| 2 | 39.95 | 48.25 | 83.25 | 89.53 | 85.00 | 69.20 | 1107 | |
| 3 | 41.75 | 49.00 | 83.75 | 90.05 | 85.67 | 70.04 | 1122 | |
| VISTA | 1 | 50.26 | 62.75 | 82.25 | 91.10 | 85.00 | 74.27 | 1204 |
| 2 | 51.29 | 63.25 | 82.50 | 91.10 | 86.33 | 74.89 | 1215 | |
| 3 | 53.35 | 64.25 | 84.00 | 92.15 | 86.67 | 76.08 | 1236 | |
| ResVision | 1 | 66.75 | 70.75 | 87.25 | 92.15 | 89.33 | 81.25 | 1335 |
| 2 | 67.01 | 69.50 | 85.00 | 91.62 | 90.33 | 80.69 | 1324 | |
| 3 | 70.88 | 69.50 | 87.25 | 92.15 | 89.00 | 81.76 | 1345 | |
| DeepSeek-V4.1-Flash-1017 | ||||||||
| Direct | 1 | 32.47 | 41.75 | 78.75 | 77.49 | 89.00 | 63.89 | 1023 |
| 2 | 33.51 | 43.50 | 79.50 | 79.06 | 89.33 | 64.98 | 1041 | |
| 3 | 34.54 | 44.25 | 79.50 | 80.10 | 90.67 | 65.81 | 1054 | |
| VISTA | 1 | 41.24 | 50.75 | 78.75 | 89.01 | 89.67 | 69.88 | 1117 |
| 2 | 42.53 | 51.50 | 79.00 | 89.53 | 89.67 | 70.45 | 1127 | |
| 3 | 44.07 | 52.50 | 79.25 | 90.58 | 90.00 | 71.28 | 1141 | |
| ResVision | 1 | 59.54 | 61.25 | 82.00 | 94.76 | 87.33 | 76.98 | 1247 |
| 2 | 57.73 | 60.50 | 81.25 | 94.24 | 89.33 | 76.61 | 1239 | |
| 3 | 60.82 | 58.75 | 81.25 | 93.72 | 89.33 | 76.77 | 1243 |
Table 11. Individual-run results: benchmark accuracy (%), Macro-5, and total correct answers; R1–R3 include all three independent runs.
Individual Run Results
Each row in Table 11 reports benchmark accuracies, Macro-5, and total correct answers for one independent execution. All three runs contribute to the aggregate results.
Paired Effects and Statistical Reliability
For paired analyses, binary correctness is first averaged per question over the three runs. We compute 95% confidence intervals with 20,000 within-benchmark question-bootstrap replicates (seed 20261102), sharing each resample across methods. Macro-5 is the unweighted mean of the five benchmark effects. Exact McNemar tests use majority correctness across runs.
| Metric / benchmark | (pp) | 95% CI | Wins / Losses | |
| Kimi-K3 | ||||
| Macro-5 | +6.40 | [+5.17, +7.66] | 169/40 | |
| BabyVision | +21.39 | [+17.78, +25.09] | 97/7 | |
| VLMsAreBiased | +6.92 | [+4.50, +9.50] | 35/7 | |
| BLINK | +1.92 | [-0.08, +4.00] | 17/9 | 0.1686 |
| VStarBench | +0.87 | [-1.40, +3.32] | 4/2 | 0.6875 |
| MMVP | +0.89 | [-2.22, +4.11] | 16/15 | 1.0000 |
| Qwen3.8-Flash-Next | ||||
| Macro-5 | +9.02 | [+7.70, +10.35] | 224/44 | |
| BabyVision | +21.13 | [+17.44, +24.91] | 98/16 | |
| VLMsAreBiased | +17.17 | [+13.92, +20.58] | 73/6 | |
| BLINK | +2.67 | [+0.25, +5.17] | 27/12 | 0.0237 |
| VStarBench | +0.00 | [-1.92, +1.92] | 2/1 | 1.0000 |
| MMVP | +4.11 | [+1.11, +7.11] | 24/9 | 0.0135 |
| GLM-5.3-Flash | ||||
| Macro-5 | +6.15 | [+4.75, +7.55] | 181/72 | |
| BabyVision | +16.58 | [+12.89, +20.27] | 80/21 | |
| VLMsAreBiased | +6.50 | [+3.25, +9.75] | 50/23 | 0.0021 |
| BLINK | +3.58 | [+1.17, +6.08] | 25/12 | 0.0470 |
| VStarBench | +0.52 | [-2.62, +3.66] | 8/7 | 1.0000 |
| MMVP | +3.56 | [+0.44, +6.67] | 18/9 | 0.1221 |
| DeepSeek-V4.1-Flash-1017 | ||||
| Macro-5 | +6.25 | [+4.82, +7.69] | 207/88 | |
| BabyVision | +16.75 | [+12.71, +20.88] | 102/32 | |
| VLMsAreBiased | +8.58 | [+5.25, +12.00] | 59/21 | |
| BLINK | +2.50 | [-0.08, +5.08] | 28/20 | 0.3123 |
| VStarBench | +4.54 | [+1.40, +8.03] | 10/3 | 0.0923 |
| MMVP | -1.11 | [-3.67, +1.44] | 8/12 | 0.5034 |
Table 12. Paired accuracy gains over VISTA. Wins/losses use majority correctness; denotes the exact McNemar probability.
Breadth Across Official BabyVision Task Types
ResVision leads both baselines across BabyVision’s four official types (Figure 8). Its gains over VISTA are +18.14, +24.54, +17.31, and +8.84 points for pattern recognition, fine-grained discrimination, spatial perception, and visual tracking.
Figure 8. Breadth across official BabyVision types. Accuracy averages four models and all three independent runs. The four types contain 51, 163, 91, and 83 questions, respectively.
| BabyVision type | Direct | VISTA | ResVision | ||
| Fine-grained discrim. | 163 | 42.33 | 53.37 | 78.12 | +24.74 |
| Pattern recognition | 51 | 64.05 | 72.06 | 90.36 | +18.30 |
| Spatial perception | 91 | 39.65 | 49.91 | 68.13 | +18.22 |
| Visual tracking | 83 | 39.96 | 41.87 | 50.70 | +8.84 |
Table 13. BabyVision gains by official task type. Repeats and models are averaged within each question.
Consistency Across Repeated Runs
All correct counts questions answered correctly in every run; all wrong counts questions missed in every run. Mixed outcomes belong to neither group. Table 14 covers 6,716 model–question pairs.
| Model | Direct | VISTA | ResVision | |||
| AC | AW | AC | AW | AC | AW | |
| Kimi-K3 | 1,047 | 408 | 1,117 | 343 | 1,247 | 255 |
| Qwen | 1,035 | 340 | 1,117 | 238 | 1,336 | 139 |
| GLM | 954 | 405 | 1,016 | 278 | 1,189 | 219 |
| DeepSeek | 897 | 477 | 961 | 384 | 1,076 | 279 |
| Total | 3,933 | 1,630 | 4,211 | 1,243 | 4,848 | 892 |
Table 14. Consistent successes and errors across three runs. AC: correct in all three runs; AW: wrong in all three runs.
ResVision is correct in all three runs on 4,848 pairs, versus 4,211 for VISTA and 3,933 for Direct. Persistent errors fall from VISTA’s 1,243 to 892, a 28.2% reduction. This yields 637 additional repeatable successes over VISTA, alongside fewer persistent failures.
Figure 9. ResVision gains over VISTA. Mean accuracy differences in percentage points over all three independent runs. Each cell gives the signed difference on a scale centered at zero.

Figure 10. Three-run paired ResVision effects. ResVision minus VISTA accuracy with pointwise 95% intervals from a 20,000-replicate question bootstrap, resampling questions within each benchmark with shared samples across methods.
Ablations and Configuration Sensitivity
This section collects the controlled comparisons that separate ResVision’s two components, test adaptive visual-token allocation, and vary its inference configuration. The Direct+SMR control and fixed-allocation study use Kimi-K3 and Qwen3.8-Flash-Next; the agentic component design additionally covers GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017. All component and allocation results use the five-benchmark Macro-5, while the one-factor configuration sweeps use BabyVision. Every condition contains three independent runs.
Direct and Agentic Component Ablations
The Direct comparison isolates SMR before any tool use: Direct processes the original image, whereas Direct+SMR processes the spectrum-matched working view. The agentic comparison forms a complete design. Removing SMR retains CPI with the original image presentation; removing CPI keeps SMR but restores VISTA-style isolated-crop replies; removing both exactly recovers VISTA. Prompts, tools, inference settings, and call limits remain fixed within each comparison. Main-paper Table 4 reports the Direct controls and the full agentic design on Kimi-K3 and Qwen3.8-Flash-Next; Table 15 below extends the agentic design to all four model families.
| Model | VISTA | +SMR | +CPI | SMR+CPI |
| Kimi-K3 | ||||
| Qwen3.8-Flash-Next | ||||
| GLM-5.3-Flash | ||||
| DeepSeek-V4.1-Flash-1017 |
Table 15. Four-model agentic SMR CPI ablation. Entries are mean Macro-5 sample standard deviation over three independent runs. +SMR and +CPI enable one component; SMR+CPI is complete ResVision. Direct controls are reported in main-paper Table 4.
On the two models with Direct controls, SMR alone improves Direct by 5.26 points on Kimi-K3 and 8.88 on Qwen. Within agentic inference, each isolated component improves Macro-5 over VISTA for every model, and complete ResVision is best in all four cases. Enabling both adds 0.84–2.21 points over the stronger single-component variant, confirming complementary gains from adaptive scaling and context preservation.
Adaptive versus Fixed Visual-Token Allocation
The fixed-token control replaces SMR with a uniform target of 1,024, 2,048, 3,072, 4,096, or 5,120 visual tokens while retaining CPI. The strongest fixed target is 2,048 tokens for both tested models; adaptive SMR improves Macro-5 by 3.31 points on Kimi-K3 and 3.36 on Qwen3.8-Flash-Next (Table 5). The table’s Original row is the SMR-only, no-CPI component control, whereas SMR (adaptive) is complete ResVision. Larger fixed targets do not improve monotonically.
Inference Configuration on BabyVision
We next report one-factor sweeps for Kimi-K3 and Qwen3.8-Flash-Next. Each sweep changes one inference factor while holding the others at the main-experiment defaults, and all measurements use complete ResVision with SMR and CPI. Figure 11 provides the overview; the tables retain every setting and individual run.
Figure 11. BabyVision configuration sweeps for complete ResVision. Rows correspond to Kimi-K3 and Qwen3.8-Flash-Next; columns vary temperature, reasoning effort, and generation ceiling. Bars show means and sample standard deviations across all three independent runs in the complete ablation tables below.
Temperature. Kimi-K3 is effectively tied at temperature 0 and 0.5, while Qwen is strongest at 0.5. Across all three settings, the ranges are 3.01 points for Kimi-K3 and 2.40 for Qwen. Temperature therefore changes the optimum slightly without overturning the overall behavior of complete ResVision.
| Model | 0 (default) | 0.5 | 1.0 |
| Kimi-K3 | |||
| Qwen3.8-Flash-Next |
Table 16. Temperature ablation for ResVision on BabyVision (%). Mean sample standard deviation over three runs. Blue shading marks the best setting within each model.
Reasoning effort. Reasoning effort causes the largest shifts in the sweep. Kimi-K3 rises from 58.08 at Low to 80.76 at Max, and Qwen rises from 69.07 to 77.75. The monotonic recovery on both models shows that multi-view inspection depends on sufficient reasoning capacity to connect the acquired detail to its global reference.
| Model | Low | Medium | Max (default) |
| Kimi-K3 | |||
| Qwen3.8-Flash-Next |
Table 17. Reasoning-effort ablation for ResVision on BabyVision (%). Mean sample standard deviation over three runs. Blue shading marks the best setting within each model.
Generation ceiling. Kimi-K3 remains essentially flat from 32K to 131K (80.76–81.27), and Qwen varies by 2.32 points with its best result at 32K. These results separate the output allowance from the visual-token allocation studied in § 4.5: the former controls how long the model may reason and answer, whereas the latter controls how the image is presented.
Table 18: Generation-ceiling ablation for ResVision on BabyVision (%). Mean sample standard deviation over three runs. Blue shading marks the best setting within each model.
| Model | 32K | 65K (default) | 131K |
| Kimi-K3 | |||
| Qwen3.8-Flash-Next |
Table 18.
Individual Run Results
Table 19 lists all three independent run accuracies underlying every BabyVision configuration cell. Means and sample standard deviations use all three values. All three sweeps reuse the complete ResVision main-experiment control at temperature 0, 65,536 tokens, and the highest supported effort.
Table 19: Complete BabyVision configuration results (accuracy, %). R1–R3 are the three independent runs; all are included in the mean and sample standard deviation.
| Factor | Setting | R1 | R2 | R3 | Mean SD |
| Kimi-K3 | |||||
| Temperature | 0 (default) | 81.96 | 80.41 | 79.90 | |
| 0.5 | 80.15 | 80.67 | 80.67 | ||
| 1.0 | 77.58 | 78.35 | 77.32 | ||
| Effort | Low | 61.34 | 56.70 | 56.19 | |
| Medium | 75.52 | 72.16 | 76.55 | ||
| Max (default) | 81.96 | 80.41 | 79.90 | ||
| Max tokens | 32K | 80.67 | 81.96 | 80.67 | |
| 65K (default) | 81.96 | 80.41 | 79.90 | ||
| 131K | 81.19 | 81.70 | 80.93 | ||
| Qwen3.8-Flash-Next | |||||
| Temperature | 0 (default) | 78.87 | 77.58 | 76.80 | |
| 0.5 | 78.35 | 78.61 | 81.19 | ||
| 1.0 | 77.58 | 76.55 | 76.80 | ||
| Effort | Low | 69.85 | 69.33 | 68.04 | |
| Medium | 74.23 | 76.03 | 77.32 | ||
| Max (default) | 78.87 | 77.58 | 76.80 | ||
| Max tokens | 32K | 77.32 | 78.87 | 79.90 | |
| 65K (default) | 78.87 | 77.58 | 76.80 | ||
| 131K | 77.58 | 75.26 | 76.29 |
Table 19.
Additional Mechanistic Evidence
We supplement the main diagnostics with weight spectra, frequency reconstruction, phase sensitivity, and feature-size probes on Qwen3.8-Flash-Next. Its vision stack uses input patches, a 27-layer ViT of width 1,152, and a merger producing 2,048-dimensional visual tokens. Each merged token corresponds to a input region before contextual mixing.
Why Presentation Scale Changes Accessibility
Changing image scale redistributes existing structure across visual tokens. Under the effective low-pass view in § 2.1, enlargement shifts a pattern from frequency to , where the encoder response can be stronger [31]. The probes below measure how amplitude, recoverability, spatial phase, and feature size change through the visual stack. CPI complements this scale adaptation by preserving the scene reference needed to interpret each enlarged detail.
Weight Spectra and Effective Rank
Table 20 reports Shannon effective rank, participation ratio, and the number of singular values required to explain 90% of squared-singular-value energy for the patch projection and merger. The product summarizes the merger’s linear weights before its activation.
Table 20: Qwen3.8-Flash-Next visual-projection weight spectra. Effective rank summarizes spectral concentration; the final column is the number of singular values covering 90% of spectral energy.
| Matrix | Shape | Max. rank | Effective rank | Participation | 90% energy |
| Patch projection (gray) | 256 | 194.76 | 158.05 | 189 | |
| Patch projection (RGB) | 768 | 241.76 | 194.10 | 229 | |
| Merger | 4608 | 1449.69 | 863.28 | 1354 | |
| Merger | 2048 | 1163.61 | 848.85 | 1013 | |
| Merger | 2048 | 643.93 | 467.65 | 523 |
Table 20.
The RGB projection concentrates 90% of its spectral energy in 229 of 768 directions (29.8%); the merger weight product does so in 523 of 2,048 directions (25.5%). These measurements indicate anisotropic sensitivity. The following spatial and reconstruction probes characterize which frequencies remain accessible after encoding.
Frequency-Dependent Linear Reconstruction
Let be an input patch, its representation at a chosen readout, a fitted linear decoder, and the reconstructed patch. The decoder is fitted on 2,560 input patches of size and evaluated on 512 held-out patches. We compare the four concatenated patch projections (4,608 dimensions), the isolated merger output (2,048), and the complete ViT-plus-merger output (2,048). For radial Fourier band , we report
Here and are Fourier coefficients of the input and reconstruction. is the normalized spectral reconstruction score, and is the reconstructed-to-input amplitude ratio, providing an empirical modulation-transfer analogue for the fitted encoder–decoder path.
The full encoder retains a strong DC reconstruction while its high-frequency reconstruction is substantially weaker (Table 21). For example, falls from 0.989 at DC to 0.137 at four cycles per token-sized region. The merger-only row localizes part of this degradation.
| Representation | DC | ||||
| Patch projection | 1.000/1.000 | 0.999/0.999 | 0.998/0.998 | 0.985/0.986 | 0.922/0.928 |
| Merger only | 1.000/1.000 | 0.950/0.969 | 0.902/0.943 | 0.847/0.915 | 0.726/0.871 |
| Full encoder | 0.989/0.995 | 0.355/0.766 | 0.210/0.720 | 0.137/0.582 | /0.447 |
Table 21. Table 21: Frequency-dependent reconstruction on held-out patches. Each entry gives . Frequencies are cycles per 32 input pixels; DC denotes the zero-frequency component.
Phase Sensitivity and Small-Feature Detection
Sensitivity to spatial phase. A synthetic square-wave grating is shifted by half a period, preserving its frequency and contrast while reversing stripe positions. The probe measures between the two representations. At two and four cycles per token-sized region, the patch-projection distances are 1.423 and 1.634, whereas the final-token distances are 0.0687 and 0.0471 (Table 22). The full encoder therefore greatly reduces representational separation for the tested phase change.
| Representation | ||||
| Patch projection | 1.134 | 1.423 | 1.634 | — |
| Full encoder | — | 0.0687 | 0.0471 | 0.0341 |
Table 22. Cosine distance after a half-period grating shift. Frequencies are cycles per 32 input pixels. Dashes indicate frequencies without an individually tabulated value.
Sensitivity to feature size. A square feature of side is injected into a region of BabyVision task 0507. The probe compares features before and after injection. Table 23 lists the measured distances. The full-encoder response crosses the scene-specific reference of 0.3699 between and pixels, establishing a clear size-dependent response in this probe.
| Side (pixels) | Patch projection | Merger only | Full encoder | Above reference |
| 1 | 0.0036 | 0.0020 | 0.0038 | No |
| 2 | 0.0137 | 0.0069 | 0.2064 | No |
| 3 | 0.0296 | 0.0143 | 0.3292 | No |
| 4 | 0.0505 | 0.0255 | 0.4008 | Yes |
| 6 | 0.1015 | 0.0561 | 0.5732 | Yes |
| 10 | 0.2186 | 0.1097 | 0.6020 | Yes |
| 16 | 0.4052 | 0.0903 | 0.6267 | Yes |
Table 23. Cosine distance induced by a local square feature. The reference column compares the full-encoder response with within-scene variation of 0.3699.
Boundary and Trained-Readout Protocol
The boundary probe moves an unchanged pattern from the center of one merged-token region toward the intervening boundary. At each position, we compare the best-matching Stage-1 patch projection and Stage-3 final visual token with their respective centered-pattern references. The main-paper plot reflects this center-to-boundary sweep about the boundary, producing a position-normalized profile that removes absolute-position and reference-token effects while preserving the measured dependence on boundary distance. The released JSON retains the original unreflected measurements. This readout tests how token-grid alignment changes the representation without changing the pattern’s appearance or scale.
The counting probe uses a strip with horizontal token positions and evaluates . For each value of , we generate 60 independently arranged strips. We freeze the complete VLM and extract the Stage-3 token grid for every strip. The diagnostic feature is the spatially ordered concatenation
where is a final visual token and is the number of tokens in the complete grid. A ridge-linear classifier with coefficient 0.1 is trained on 45 strips per count and evaluated on the remaining 15 while the VLM remains frozen. Its held-out accuracy measures whether exact count is linearly decodable from the frozen representation.
For the generation comparison, the same evaluation images are paired with a direct integer-count prompt. We report greedy exact-match accuracy and the normalized probability assigned to the correct one- or two-digit integer. The contrast therefore holds the visual input and frozen VLM fixed while comparing a supervised linear readout with the model’s native generative readout.
Whole–Detail Representation Diagnostic
We complement the main-paper grounding experiment with a representation-level control on the same 50 BabyVision images. For each crop width, we match the same source patch between the complete image and the crop. The Stage-1 patch projection tests whether the local visual content itself changes; the Stage-3 final visual token measures drift after contextual visual processing; and the Stage-4 token additionally includes the crop-local positional frame. At 12.5% crop width, Stage-3 similarity falls to 0.389 and Stage-4 similarity to 0.305, while the Stage-1 projection remains unchanged. The dashed line marks the mean similarity between unrelated foreground tokens.
Figure 12. Representation drift for the same source patch. Local patch features remain identical at Stage 1, while contextual representations diverge as the crop narrows and diverge further after a new positional frame is assigned.
Auxiliary Retrieval and Grounding Diagnostics
Two complementary diagnostics show how retrieval and grounding failures unfold. During dense counting, the peak attended visual-token coordinate reaches the right edge before all objects have been enumerated. During crop grounding, resetting the crop origin produces localization error that grows with the target’s displacement from the source-image origin.

Figure 13. Auxiliary diagnostics for retrieval and separated-image grounding. (a) Dense counting drives the visual attention pointer to the grid boundary before enumeration is complete. (b) An isolated crop accumulates larger localization error as its source position moves away from the reset crop origin; retaining the complete view in the same turn reduces this error.
Connection to ResVision. Across the probes, high-frequency reconstruction weakens through the full encoder, phase distinctions contract, and the response to a local feature grows with its presented size. These measurements support the accessibility account in § 2: presentation scale changes which distinctions reach a useful representation. SMR operationalizes this observation by assigning scale from measured fine structure, while CPI addresses the complementary problem of preserving what enlarged evidence means in the complete scene. Appendix § 14 gives the exact resizing protocol used by the reported experiments.
Spectrum-Matched Resizing Protocol
This section records the working-image geometry, parameter settings, and implementation controls used in the SMR experiments.
Working-Image Geometry and Fixed-Token Control
Let be image , with native dimensions and backend stride . A target gives . The fixed-token controls evaluate ; the Qwen processor profile uses pixels per merged visual token. The target is a controlled allocation, whereas the stride is a processor property. Each backend therefore uses its verified processor geometry or token-count interface.
The scaled dimensions are aligned to the processor’s supported grid while preserving aspect ratio up to rounding and respecting the allocated token cap. Each run records the backend’s processed dimensions and image-token count.
For an ordered set of images, the fixed rule assigns the same target to each image, with total complete-view allowance . The resized set is supplied initially, used for cropping, and repeated in each CPI reply. Image order is preserved.
All tool coordinates and crops use the fixed working image . The local renderer, output-size allowance, and global interpolation rule are held fixed across conditions.
Spectrum-Matched Scale Selection
SMR follows the active-tile spectrum and minimum-cost frequency-matching rule in § 3.1. The image is converted to grayscale, partitioned into tiles, and ranked by gradient energy. Radial Fourier power from the retained tiles gives , which determines the target scale under the token bounds in Table 24. Images with zero non-DC power use the minimum-budget scale.
| Rule | Parameter | Evaluated setting |
| Fixed-token | Target count | 1024, 2048, 3072, 4096, 5120 |
| Spectrum | Active-tile fraction | 0.25 |
| Spectrum | Energy quantile | 0.90 |
| Spectrum | Target cycles per token | 0.77 |
| Spectrum | Per-image bounds | (512, 4608) |
| Backend property | Effective stride (Qwen profile) | 32 pixels |
Table 24. Resize parameters used in the reported study. SMR parameters are held fixed across benchmarks rather than tuned to a task label. Fixed-token resizing supplies the frequency-independent controls.
For multiple images under a shared cap, the adapter reserves the per-image minimum and proportionally reduces requested allocations above that minimum when necessary. Grid alignment remains within the resulting allocations. Scale choices are computed once per image and cached throughout the interaction.
The frequency target and token bounds remain fixed across benchmarks. Fixed-token resizing provides the frequency-independent control.
Implementation and Cost Accounting
Spectrum estimation runs once before inference. The selected scale and working image are cached for the interaction, and the reported end-to-end latency includes this preprocessing. Each run records the processed dimensions, realized visual-token allocation, and selected scale.
System-Level Error Taxonomy
The analysis covers the complete five-benchmark, three-run archive for VISTA and ResVision. Its unit is one judged prediction for one model, run, and question, giving predictions per method. A Kimi-K3 judge assigns every incorrect prediction to one mutually exclusive category using the task prompt, reference answer, final response, and inspection trace. Percentages use all 20,148 predictions as the denominator.
Six failure modes. Detail accessibility and detail interpretation are distinct: a cue may be difficult to distinguish at the full-image scale, or it may be visible but used incorrectly. We partition every incorrect prediction into six mutually exclusive classes.
F1—referent binding is an identity error: the model reads a local cue but attaches it to the wrong object, cell, image, or coordinate. In the D6 counting case (Appendix § 16.§ 1), the crop makes the vertical marks legible but removes the row and column headers. The defining question is which entity or location does this evidence belong to?
F2—candidate-contrast omission is a comparison error: the model accepts one plausible candidate without checking alternatives that could distinguish it. In the horse reconstruction case (Appendix § 16.§ 4), the extra yellow bar in A becomes decisive only when the candidate inventories are compared. The defining question is what distinguishes this candidate from the others?
F3—spatial/transform inconsistency is a relational error: the model recognizes relevant local parts but combines them under an incorrect direction, rotation, topology, ordering, or cross-view correspondence. In the rotation case (Appendix § 16.§ 2), wedge colors can be read correctly while the requested transformation and answer option become mismatched. The defining question is how must these parts correspond after the stated relation or transformation?
F4—local-detail overinterpretation is an interpretation error: the model identifies the correct entity and region but gives a local attribute the wrong meaning because the crop omits scale, pose, or object-level context. In the volleyball case (Appendix § 16.§ 5), the arm-and-ball crop supports a plausible “set” interpretation, while the full pose supports “bump.” The defining question is what does this cue mean in the complete object or scene?
F5—coverage/count aggregation collects failures to exhaustively enumerate repeated instances or reconcile partial counts. Its defining question is has the reasoning covered and reconciled every relevant instance?
F6—other errors contains residual failures outside the preceding five classes.
Together, F1–F6 separate identity, comparison, relation, interpretation, aggregation, and other residual causes. C1P supplies grounding for the first four, while the complete interaction can also support systematic coverage and reconciliation.

Figure 14. Error incidence over all predictions. Each bar uses the same fixed total of 20,148 predictions. Colored segments show the six failure classes and the pale remainder is correct. ResVision reduces the overall error mass from 27.6% to 19.8%.
Table 25: Absolute incidence of system-level errors. Percentages use the fixed 20,148 predictions per method. Negative favors ResVision.
| Error category | VISTA (%) | ResVision (%) | Rel. reduction | |
| F1 Referent binding | 597 (3.0) | 419 (2.1) | −178 | 29.8% |
| F2 Candidate contrast | 569 (2.8) | 406 (2.0) | −163 | 28.6% |
| F3 Spatial consistency | 701 (3.5) | 514 (2.6) | −187 | 26.7% |
| F4 Detail overinterpretation | 834 (4.1) | 730 (3.6) | −104 | 12.5% |
| F5 Coverage/count aggregation | 1,867 (9.3) | 1,201 (6.0) | −666 | 35.7% |
| F6 Other | 1,002 (5.0) | 718 (3.6) | −284 | 28.3% |
| All errors | 5,570 (27.6) | 3,988 (19.8) | −1,582 | 28.4% |
Table 25.
Table 26: Question-matched error transitions from VISTA to ResVision. Fixed denotes VISTA-wrong/ResVision-correct predictions; regressed is the reverse. The category is assigned on the failing side.
| Error category | VISTA errors | Fixed | Regressed | Diff | Fix rate | Exact |
| F1 Referent binding | 597 | 301 | 99 | +202 | 50.4% | |
| F2 Candidate contrast | 569 | 313 | 108 | +205 | 55.0% | |
| F3 Spatial consistency | 701 | 335 | 156 | +179 | 47.8% | |
| F4 Detail overinterpretation | 834 | 325 | 188 | +137 | 39.0% | |
| F5 Coverage/count aggregation | 1,867 | 855 | 223 | +632 | 45.8% | |
| F6 Other | 1,002 | 389 | 162 | +227 | 38.8% | |
| All errors | 5,570 | 2,518 | 936 | +1,582 | 45.2% |
Table 26.
ResVision removes 1,582 errors overall, with declines in all six classes: 29.8% for referent binding, 28.6% for candidate contrast, 26.7% for spatial consistency, 12.5% for local-detail overinterpretation, 35.7% for coverage/count aggregation, and 28.3% for other errors.
In the matched analysis, ResVision corrects 38.8–55.0% of VISTA errors within each class and introduces fewer reverse transitions in all six. Across the taxonomy, 2,518 predictions are fixed and 936 regress, a net gain of 1,582 question-matched predictions (, exact two-sided McNemar test). Table 27 reports the per-benchmark distribution.
Table 27: Complete error taxonomy by benchmark. Entries are count (percentage of all predictions for that benchmark).
| Benchmark | Method | F1 Bind | F2 Contrast | F3 Spatial | F4 Detail | F5 Coverage | F6 Residual | All errors |
| BabyVision | VISTA | 380 (8.2) | 342 (7.3) | 407 (8.7) | 96 (2.1) | 564 (12.1) | 420 (9.0) | 2,209 (47.4) |
| BabyVision | ResVision | 219 (4.7) | 199 (4.3) | 291 (6.2) | 78 (1.7) | 277 (5.9) | 262 (5.6) | 1,326 (28.5) |
| VLMsAreBiased | VISTA | 26 (0.5) | 56 (1.2) | 62 (1.3) | 215 (4.5) | 1,229 (25.6) | 330 (6.9) | 1,918 (40.0) |
| VLMsAreBiased | ResVision | 6 (0.1) | 49 (1.0) | 42 (0.9) | 216 (4.5) | 863 (18.0) | 272 (5.7) | 1,448 (30.2) |
| BLINK | VISTA | 163 (3.4) | 166 (3.5) | 126 (2.6) | 124 (2.6) | 16 (0.3) | 214 (4.5) | 809 (16.9) |
| BLINK | ResVision | 158 (3.3) | 152 (3.2) | 102 (2.1) | 108 (2.2) | 9 (0.2) | 152 (3.2) | 681 (14.2) |
| VStarBench | VISTA | 26 (1.1) | 0 (0.0) | 56 (2.4) | 83 (3.6) | 5 (0.2) | 8 (0.3) | 178 (7.8) |
| VStarBench | ResVision | 33 (1.4) | 0 (0.0) | 38 (1.7) | 58 (2.5) | 8 (0.3) | 7 (0.3) | 144 (6.3) |
| MMVP | VISTA | 2 (0.1) | 5 (0.1) | 50 (1.4) | 316 (8.8) | 53 (1.5) | 30 (0.8) | 456 (12.7) |
| MMVP | ResVision | 3 (0.1) | 6 (0.2) | 41 (1.1) | 270 (7.5) | 44 (1.2) | 25 (0.7) | 389 (10.8) |
Table 27.
Representative Failure Cases
Five cases illustrate the dominant context-sensitive failure classes in Sec. G. They are ordered by the consequence of losing global context: referent, spatial relation, coverage, candidate contrast, and local interpretation. The cases were selected after the quantitative taxonomy was completed and do not determine its categories or statistics.
| Case | Category | Benchmark / ID | Global constraint | Direct | VISTA | ResVision |
| A | F1 Referent binding | VLMsAreBiased / 62280 | cell address | 1/3 | 1/3 | 3/3 |
| B | F3 Spatial consistency | BabyVision / 510 | row–transform mapping | 0/3 | 1/3 | 3/3 |
| C | F5 Coverage aggregation | BabyVision / 4552 | complete lattice | 0/3 | 0/3 | 3/3 |
| D | F2 Candidate contrast | BabyVision / 2353 | shared inventory | 1/3 | 0/3 | 3/3 |
| E | F4 Detail interpretation | MMVP / 90100 | whole-body pose | 3/3 | 0/3 | 3/3 |
Table 28. Severity-ordered qualitative cases. Entries are exact-match successes out of three runs.
The common requirement. Local evidence must retain a role in the complete task: an address, a transform, a position in a coverage ledger, a competing hypothesis, or an object-level context. ResVision preserves that role while inspecting the detail.
Case A (F1): Bind the Detail to the Right Referent
The question asks how many vertical marks appear in cell D6. Enlargement makes the marks easier to distinguish, while the full table supplies their row and column address. A correct local count answers the question only if it is bound to D6.
Direct answers 4, 3, 4. It is correct once, but the count changes across runs because the small marks are difficult to read reliably in the full table.
VISTA answers 2, 4, 3. Its crop makes the marks larger, but removes the row and column headers. The model can therefore read a nearby cell clearly while losing confidence about whether that cell is D6.
ResVision answers 3, 3, 3. It first locates column D and row 6 in the full table, then enlarges only their intersection and counts the three marks there.
Why this matters. A clear local observation is useful only when it remains attached to the correct global referent.

Figure 15. F1—referent binding. The crop resolves the marks; the full table proves that they belong to D6. The figure includes the labels “Original question,” “Q: How many vertical lines are in cell D6?”, “A: 3,” “VISTA,” “1. Inspect locally,” “Enlarge a cell-sized region,” “2. Read the marks,” “Count visible strokes without reliable row and column anchors,” “3. Commit,” “Return a locally plausible count,” “Outcome: wrong cell/count in 2 of 3 runs,” “ReVision,” “Global context,” “1. Anchor globally,” “Locate column D and row 6 in the complete table,” “2. Inspect locally,” “Magnify their intersection only,” “3. Re-ground,” “Bind the three visible marks back to D6,” “Local detail,” and “Outcome: 3, correct in all 3 runs.”
Case B (F3): Preserve the Spatial Contract
The question asks for three different rotations and expects an ordered triple. Each local match must therefore stay linked to its source row, requested angle, and final option number.
Direct answers (6, 4, 5), (6, 4, 5), and (5, 1, 3). It can recognize some rotated patterns, but does not keep all three row-specific correspondences correct at once.
VISTA answers (4, 6, 2), (6, 2, 3), and (6, 4, 2). Enlarged wedges clarify the local color order, yet the row-to-angle-to-option bookkeeping is lost in two runs.
ResVision answers (6, 4, 2) in all three runs. It uses local views to resolve wedge order while the full layout keeps row 1, row 2, and row 3 tied to the right transformations and options.
Why this matters. A correct local match is not enough; it must be returned to the right place in the ordered answer.

Figure 16. F3—spatial consistency. Each local match remains attached to its source row, requested transform, and option index.
Case C (F5): Make Coverage Auditable
The question asks for the total number of dark cells in a grid. Individual cells are easy to identify, but the final count is correct only if every position is visited once, with no omissions or duplicates.
Direct answers 37, 39, 46. All three totals are too small, which is consistent with losing cells while reading a dense grid at its original scale.
VISTA answers 59, 60, 52. Zooming makes a local group of cells easier to read, but the totals move above and below the truth because there is no explicit record of which positions have already been counted.
ResVision answers 54, 54, 54. It treats the full lattice as a ledger, records the row totals 2, 3, 5, 7, 3, 7, 10, 8, 4, 5, 0, and checks that every cell contributes exactly once before summing them.
Why this matters. Precise local readings still need a global coverage plan to become a reliable aggregate.

Figure 17. F5—coverage aggregation. Global indexing turns local readings into a complete, nonoverlapping count.
Case D (F2): Compare the Competing Hypotheses
The question asks which option contains exactly the same pieces as the target horse. Options A and C look similar, so the decision depends on comparing their complete piece inventories rather than describing the target alone.
Direct answers A, A, and C. It notices the overall resemblance, but only one run distinguishes the candidates correctly.
VISTA answers A in all three runs. Its close view gives a detailed description of the target horse, yet does not place A and C in the same comparison; the extra yellow bar in A is therefore overlooked.
ResVision answers C in all three runs. It compares the target, A, and C together, checks the piece inventory item by item, and rejects A after finding its extra yellow bar.
Why this matters. Inspection should test competing hypotheses against one shared reference and compare every plausible option.

Figure 18. F2—candidate contrast. The decisive evidence appears only when target, A, and C share one comparison.
Case E (F4): Interpret the Detail in the Whole Pose
The question asks whether the player is more likely to bump or set the volleyball. A close view of the arms and ball can suggest either action; the whole-body pose is needed to decide which interpretation is physically consistent.
Direct answers A (bump) in all three runs. In this example the full image already preserves the decisive cue: the player is leaning forward with joined forearms held low.
VISTA answers B (set) in all three runs. The crop emphasizes a bent arm and a high ball, which supports a plausible local story about setting, but hides the rest of the posture that contradicts it.
ResVision answers A (bump) in all three runs. It inspects the arm-and-ball region while retaining the full pose, then checks that the local cue agrees with the joined forearm platform and forward body geometry.
Why this matters. This example separates detail accessibility from interpretation: Direct already reads the complete pose correctly, but the local inspection supports a misleading action hypothesis. Keeping the whole beside the detail provides the context for checking that hypothesis.

Figure 19. F4—detail interpretation. The crop is sharp, but the whole-body geometry determines the action. Graphic text: Original question. Q: What is the woman more likely to do: (A) bump or (B) set? A: A (bump). VISTA. 1. Inspect contact Zoom around the player, arms, and ball. 2. Read locally Treat the bent arm and high ball as an overhead setting cue. 3. Commit Choose set without checking the complete body geometry. Outcome: B (set), incorrect in all 3 runs. ResVision. Global context. 1. Keep the whole pose Retain the player's body axis and both arms. 2. Inspect the platform See joined forearms extended forward and low. 3. Reconcile Interpret the local cue under the complete posture. Local detail. Outcome: A (bump), correct in all 3 runs.
References
- [1]Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022.arxiv.org/abs/2204.14198
- [2]Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023.arxiv.org/abs/2308.12966
- [3]Ruzena Bajcsy. Active perception. Proceedings of the IEEE, 76(8):996–1005, 1988.DOI
- [4]Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. FlexiViT: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14496–14506, 2023.arxiv.org/abs/2212.08013
- [5]Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In International Conference on Learning Representations, 2023.arxiv.org/abs/2210.09461
- [6]Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P. Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. ViP-LLaVA: Making large multimodal models understand arbitrary visual prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12914–12923, 2024.arxiv.org/abs/2312.00784
- [7]Liang Chen, Weichu Xie, Yiyan Liang, Hongfeng He, Hans Zhao, Zhibo Yang, Zhiqi Huang, Haoning Wu, Haoyu Lu, et al. BabyVision: Visual reasoning beyond language. arXiv preprint arXiv:2601.06521, 2026.arxiv.org/abs/2601.06521
- [8]Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. PaLI: A jointly-scaled multilingual language-image model. In International Conference on Learning Representations, 2023.arxiv.org/abs/2209.06794
- [9]Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Leiwei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024.
- [10]Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023.
- [11]DeepSeek-AI. DeepSeek-V4.1-Flash: Pushing the limits of KV cache compression. arXiv preprint arXiv:2609.19969, 2026.
- [12]Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey Gritsenko, Mario Lucic, and Neil Houlsby. Patch n’ pack: NaViT, a vision transformer for any aspect ratio and resolution. In Advances in Neural Information Processing Systems, 2023.arxiv.org/abs/2307.06304
- [13]Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.
- [14]Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024.arxiv.org/abs/2404.12390
- [15]Jun Gao, Yongqi Li, Ziqiang Cao, and Wenjie Li. Interleaved-modal chain-of-thought. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19520–19529, 2025.arxiv.org/abs/2411.19488
- [16]Xingang Guo, Utkarsh Tyagi, Advait Gosai, Paula Vergara, Jayeon Park, Ernesto Gabriel Hernández Montoya, Chen Bo Calvin Zhang, Bin Hu, Yunzhong He, Bing Liu, and Rakshith Sharma Srinivas. Beyond seeing: Evaluating multimodal LLMs on tool-enabled image perception, transformation, and reasoning. arXiv preprint arXiv:2501.12712, 2025.
- [17]Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953–14962, 2023.arxiv.org/abs/2211.11559
- [18]Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu, and Kaiming He. VISTA: A visual harness for reasoning in an interactive world. arXiv preprint arXiv:2610.02200, 2026.arxiv.org/abs/2610.02200
- [19]Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazhen Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. CogAgent: A visual language model for GUI agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281–14290, 2024.arxiv.org/abs/2312.08914
- [20]Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. In Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2406.09403
- [21]Woojun Jung, Jaehoon Go, Mingyu Jeon, Sunjae Yoon, and Junyeong Kim. Visual funnel: Resolving contextual blindness in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Findings, pages 8962–8971, 2026.arxiv.org/abs/2512.10362
- [22]Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. VisualWebArena: Evaluating multimodal agents on realistic visual web tasks. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 881–905, 2024.
- [23]Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understanding. In Proceedings of the 40th International Conference on Machine Learning, pages 18893–18912, 2023.arxiv.org/abs/2210.03347
- [24]Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, 2023.arxiv.org/abs/2301.12597
- [25]Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763–26773, 2024.arxiv.org/abs/2311.06607
- [26]Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023.
- [27]Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, and ChunYuan Li. LLaVA-Plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, 2024.arxiv.org/abs/2311.05437
- [28]Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021.arxiv.org/abs/2103.14030
- [29]Volodymyr Mnih, Nicolas Heess, Alex Graves, and Koray Kavukcuoglu. Recurrent models of visual attention. In Advances in Neural Information Processing Systems, pages 2204–2212, 2014.arxiv.org/abs/1406.6247
- [30]Moonshot AI. Kimi K3: Open frontier intelligence. Official technical blog, https://www.kimi.com/en/blog/kimi-k3, 2026.arxiv.org/abs/2607.24653
- [31]Namuk Park and Songkuk Kim. How do vision transformers work? In International Conference on Learning Representations, 2022.arxiv.org/abs/2202.06709
- [32]Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In International Conference on Learning Representations, 2024.
- [33]Zihan Qiu, Zekun Wang, Xiao Li, Yanpeng Li, Yang Xu, Yixuan Wang, Huaqing Zhang, Rui Men, Boqiao Mao, Chengruodong Zhang, et al. On the design of Qwen3.8-Next architecture: Evaluation, efficiency, and training stability, 2026.arxiv.org/abs/2608.30320
- [34]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748–8763, 2021.arxiv.org/abs/2103.00020
- [35]Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. DynamicViT: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems, 2021.arxiv.org/abs/2106.02034
- [36]Michael S. Ryoo, A. J. Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. TokenLearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems, 2021.
- [37]Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual CoT: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. arXiv preprint arXiv:2403.16999, 2024.arxiv.org/abs/2403.16999
- [38]Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual inference via Python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888–11898, 2023.arxiv.org/abs/2303.08128
- [39]Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal LLMs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568–9578, 2024.arxiv.org/abs/2401.06209
- [40]An Vo, Khai-Nguyen Nguyen, Mohammad Reza Taesiri, Vy Tuong Dang, Anh Totti Nguyen, and Daeyoung Kim. Language models are biased. arXiv preprint arXiv:2505.23941, 2025.
- [41]Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual ChatGPT: Talking, drawing and editing with visual foundation models, 2023.arxiv.org/abs/2303.04671
- [42]Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal LLMs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084–13094, 2024.arxiv.org/abs/2312.14135
- [43]Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments. In Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2404.07972
- [44]Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. LLaVA-UHD: an LMM perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024.arxiv.org/abs/2403.11703
- [45]Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Jiujuan Wang. MM-REACT: Prompting ChatGPT for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023.arxiv.org/abs/2303.11381
- [46]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations, 2023.arxiv.org/abs/2210.03629
- [47]Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In International Conference on Learning Representations, 2024.arxiv.org/abs/2310.07704
- [48]Z.ai. GLM-5.3-Flash: Frontier intelligence, flash cost. Official technical blog, https://z.ai/blog/glm-5.3-flash, 2026.
- [49]Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. GPT-4V(ision) is a generalist web agent, if grounded. In Proceedings of the 41st International Conference on Machine Learning, pages 61349–61385, 2024.arxiv.org/abs/2401.01614






