- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Fine-grained VQA on high-resolution scenes involves a trade-off. Global resizing loses fine details. Uniformly raising the resolution of the whole image wastes computation on irrelevant background. Active-cropping methods (V*, Mini-o3) need multiple LLM forward passes, which makes inference expensive and ties evidence acquisition to the answering policy.
The paper proposes EviViT, a lightweight learnable module inserted into the intermediate layers of a frozen ViT. A question-conditioned "evidence density" is distilled from human visual-search trajectories and supervises the Prompt-conditioned Token–Evidence Aligner (PTEA), which predicts where high-resolution detail should be acquired. A planner then selects at most two decisive regions plus one complementary contextual region and re-encodes them from the original pixels. A coordinate-aware Sparse Bridge fuses the resulting features with the global tokens. The original ViT, merger, and LLM stay frozen throughout. The attachment can be reused across compatible post-trained models (e.g., from Qwen3-VL to its Qwen3.5/Vision-OPD descendants).
Experiments cover nine hosts (4B–9B). EviViT improves the seven-benchmark average by 1.98–7.40 points. Under matched token budgets from 1K to 20K, it outperforms global-only processing across the board while using only 90–97% of the tokens. General capabilities (MMBench/MMStar/Visual-CoT) are largely preserved, and the method is compatible with language-side LoRA SFT.
Critical strengths
Novel perspective. Unlike the iterative cropping of V*/Mini-o3 and the distillation approaches of ZwZ/Vision-OPD, this work treats "evidence acquisition" as a separate problem and implements it as a standalone attachment on a frozen ViT.
Human-trajectory supervision. Supervision comes from complete, real human search trajectories. The full-trace vs. final-box comparison in Table 5 (VP-Hard +3.77) directly supports this core design choice.
Rigorous controls and thorough ablations.The location controls in Table 8 (random translation, cross-image shuffling) show that the gains come from "choosing the right locations." Table 14 shows that the deployed crop regions cover only 12.39% of the image area yet achieve an average coverage of 73.69% of the official targets. Table 9 uses per-question rescue/loss statistics and McNemar tests to rule out "random answer flipping" as an explanation.
Good reproducibility and practicality. Training time is short, the host is fully frozen, and the attachment transfers directly to post-trained derivatives that share the same vision tower.
Critical weaknesses
The appendix experiments (Table 13, pooled) undermine the core claims. (a) The Sparse Bridge provides no gain over simply feeding the crops as multi-image input. The contribution that "coordinate-aware fusion helps understand details in scene context" is not supported by the data; its only remaining role is engineering convenience at the interface level. (b) Cropping from the already-resized global image performs exactly the same as "re-reading from the original pixels at native detail." This contradicts the central claim in the Introduction that "selecting encoded tokens cannot recover detail lost during initial resizing."
Missing key comparisons. Comparing only against global-only is insufficient. Under the same host and the same token budget, the method should be compared at least against training-free question-guided cropping (ViCrop, ZoomEye, DC², etc.) and region selection based on VLM internal attention or relevance (FOCUS, etc.).
In-distribution advantage and insufficient evidence of generalization. The PTEA trajectories and the Bridge's answer cross-entropy both come from the VisualProbe training set, and the largest gains appear on VisualProbe Medium/Hard (average +8.2 and +8.6). Gains on HR-Bench are much smaller. In Table 9, the McNemar p-values for the Easy and Hard tiers are not significant on several hosts.
Questions for authors
Regarding Table 13: cropping from the resized global image yields exactly the same result as re-reading from the original image (66.57). How do you reconcile this with the core claim of "recovering detail lost to resizing"? Please rerun this control in a setting where the global view is heavily downscaled (e.g., 1K/2K budgets).
Under the same host and the same token budget, how does EviViT compare with "Qwen3-VL's own grounding → crop → multi-image input" and with ViCrop/ZoomEye?
Do the VisualProbe training set, its test set, and V* overlap in image sources? What exactly does the "186-image overlap" in the appendix refer to?