ResVision adds two constraints to visual-reasoning agents: set sampling scale from image spectrum, keep the full working view beside every zoom
Synopsis
ResVision is a training-free framework that fills two gaps when a vision-language model zooms in to inspect a region: it picks a global sampling scale from the image's own fine structure, and it returns every enlarged crop together with the complete working view. Across four frontier vision-language models and 29 benchmarks, the complete configuration (Agentic Base+SMR+CPI) is the best main setting for every model, with a four-model mean Macro-29 of 75.93, 3.88 points above Direct Base and 2.28 points above Agentic Base.
Interpretation
When a visual-reasoning agent repeatedly enlarges local regions, no new scene content is added, yet answers change; ResVision splits that change into two separately addressable steps, the scale at which the whole image is sampled into visual tokens and whether an enlarged crop is still presented alongside the full scene. Earlier approaches mostly set visual-token allocation from image geometry, a fixed resolution policy, or importance estimated after tokenization, while a smooth photograph, a dense diagram and a fine texture of similar dimensions need different sampling densities. On four vision-language models and 29 benchmarks, complete ResVision beats both Direct and Agentic baselines for every model: 3.36–4.94 points over Direct Base and 1.87–2.94 points over Agentic Base, raising the four-model mean Macro-29 from 72.05 and 73.65 to 75.93.
The two components act at different points along the evidence path: one fixes the working scale of the whole image before tokenization, the other places local detail and the full reference in the same input on each tool turn. Spectrum-Matched Resizing (SMR) first estimates an image's high-frequency demand, then selects, within a visual-token budget, the minimum global scale that brings representative spatial frequencies into the encoder's usable band. Context-Preserving Inspection (CPI) returns each enlarged crop with the same complete working view as separate, labeled image inputs; on 50 BabyVision images, joint presentation holds 72% grounding accuracy across all tested scales, whereas isolated crops reach 40–58% at the corresponding scales.
Each component helps on its own and the combination is stronger, indicating that adaptive sampling and context preservation address distinct constraints. Within agentic inference, CPI alone raises Agentic Base by 1.10–1.97 points without SMR; SMR-only and CPI-only variants average 75.10 and 75.07. Complete ResVision reaches 75.93 and exceeds the stronger single-component variant by 0.40–0.94 points on every model.
Perspective
The conclusions rest on evaluation across four frontier vision-language models (Kimi-K3, Qwen3.8-Flash-Next, GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017) and 29 benchmarks, with agentic configurations sharing a four-call tool ceiling and three independent runs per setting. Gains are uneven: relative to Agentic Base, Math & Logic rises 4.59–8.34 points and Hallucination & Count 2.08–6.19 points, while General Visual is comparatively flat.
The fixed-token control shows accuracy is non-monotonic in budget; the strongest fixed target is 2,048 visual tokens for both tested models, and adaptive SMR improves on it by 3.31 points on Kimi-K3 and 3.36 on Qwen3.8. Reasoning effort is the largest factor in the configuration sweep, raising accuracy by 22.68 points on Kimi-K3 and 8.68 on Qwen3.8 from Low to Max; temperature changes produce a spread of at most 3.01 points, and raising the generation ceiling beyond 32K does not consistently improve accuracy.
