acceptodds
Under review as a conference paper at ICLR 2027

The Cost of a Near Miss: Compute-Matched Evaluation of Training-Free Visual Cropping

Abstract

Training-free visual cropping aims to improve fine-grained visual question answering, but an appended crop, and the passes that localize it, also add compute. We count visual tokens per question and compare cropping methods to no-crop baselines given the same tokens. ViCrop's reported gain reproduces, and HiDe's keeps its sign at a smaller size. On LLaVA-1.5, which reads images at a fixed resolution, ViCrop's crop beats centered, random or repeated second views of identical token count by 18.9 to 21.5 percentage points on V\*Bench, and two higher-resolution tiles by 13.1. At HiDe's own budget of 16,384 tokens, a baseline allowed all of HiDe's tokens still trails it by 3.4 and 6.5 points on HR-Bench-4K and half of HR-Bench-8K. At low budgets on a backbone that can spend the same tokens on resolution, neither gain survives: ViCrop's once the answering call is matched, HiDe's once its localization is charged. Pooled over 30 Qwen2.5-VL configurations, matching the answering call lowers ViCrop's gain of 1.81 points by 2.65, leaving no gain distinguishable from zero on natural scenes and a loss of 2.03 on documents. Whether cropping pays depends on whether its box finds the target often enough to beat what the baseline gains by spending the same compute on resolution. Annotated-box crops on V\*Bench still gain 19.9 to 30.4 points when matched, but at ViCrop's 256-token setting attention boxes hit the target on only 26.7% to 39.3% of questions. In-place magnification adds no tokens, but pasting an enlarged region over the image can overwrite the target when the box narrowly misses it. Displacing annotated boxes on V\*Bench reveals a near-miss trough: nearby misses score 7.8 to 10.8 points below distant ones at comparable hit rates. Real attention boxes pay this price too. On the 44% of their misses whose pasted patch still covers the target, pasting costs 12.5 points more than on the rest, a gap that resampling the same boxes does not show. Resampling, which compresses the surroundings instead, shows no trough on any of the three models, yet with detector boxes it is significantly worse than pasting on five of six document configurations. Cropping should therefore be compared at matched compute, and in-place methods should report how their primitive responds to localization errors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
vision-language models, visual cropping, compute-matched evaluation, localization, high-resolution visual question answering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZGo9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper evaluates the benefits of two existing “thinking with images” methods, ViCrop and HiDe, from the perspective of token costs. It reports that the gains claimed by both methods diminish when comparisons are made under matched token budgets. The paper also examines how crop accuracy and the placement of cropped regions affect performance.

Critical strengths
  1. The paper identifies unfair comparisons between existing “thinking with images” methods and their baselines in terms of token costs, and develops a comprehensive evaluation and associated metrics under constrained token budgets.
  2. The paper explains why accurate bounding boxes are important for improving visual understanding.
Critical weaknesses
  1. The writing is difficult to follow. Many sentences are hard to understand. For example: “At low budgets on a backbone that can spend the same tokens on resolution, neither gain survives: ViCrop’s once the answering call is matched, HiDe’s once its localization is charged.”
  2. Reallocating a multi-round token budget to a single baseline call does not constitute a fair comparison. In my view, pooling the input tokens used across ViCrop’s multiple rounds and allocating that budget to a single baseline call raises the baseline’s visual processing capacity, but does not establish that ViCrop itself is less effective. Under this setup, the two methods receive images at different resolutions, so differences in performance may arise from image resolution rather than the methods themselves. The baseline’s performance is thus tied to the artificially imposed resolution limit: relaxing that limit can improve its results without demonstrating an inherent advantage over ViCrop.
Questions for authors
  1. How could the effects of different token budgets be evaluated while ensuring a fair comparison in terms of input image resolution?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ZGo9

Thank you for the precise objection. Your strengths state the starting point: published comparisons give the method more tokens than its baseline. Your concern calls for two comparisons, placement at fixed resolution and budget allocation; the submission ran the first on LLaVA-1.5, and the revision adds it on the three Qwen models. On two points we respectfully differ.

1. The main comparison does not pool ViCrop's rounds. The baseline receives only the tokens of ViCrop's answering call, the main image plus the crop, bracketed from below and above (ratios 0.92–0.99 and 1.03–1.11). ViCrop's two localization passes are not charged, which favours ViCrop (Section 3.2, as in the submission), and Appendix A.4 checks that the brackets also enclose the answering call's FLOPs. Pooling across calls appears only in the total-compute comparisons (Section 5.3; Appendix B.3 for HiDe).

2. At fixed resolution, the crop's placement wins. We do not claim that ViCrop is ineffective in itself. On LLaVA-1.5, whose resolution is fixed, ViCrop's crop beats centered, random and repeated second views by 18.9 to 21.5 points at identical tokens, and two higher-resolution tiles by 13.1; the submission concluded that "its published gain stands" (now Section 4.1, Table 8). On Qwen (Table 10) the main image is pixel-identical and only the second view changes. R costs exactly the crop's tokens; D and T are within 2% of them on Qwen2.5-VL and 1–13% above them on Qwen3-VL.

ModelSetting− repeat (D)− two tiles (T)− random box (R)− higher-res baseline (G)
Qwen2.5-VL-3BHR-4K b512+3.5±3.3*+3.9±3.3*+3.4±2.9*+2.9±4.2
Qwen2.5-VL-7BV* b256+6.8±6.8+7.9±6.8*+7.3±6.5*+4.7±6.8
Qwen3-VL-8BV* b256+13.6±6.5*+16.8±6.8*+14.7±6.8*-1.0±7.3
Qwen3-VL-8BHR-4K b256+4.4±3.6*+7.2±3.8*+6.9±3.4*-4.4±4.5

Selected rows; crop minus condition, points ± rounded 95% paired-bootstrap interval half-width (HR-Bench resampled by question), * p < 0.05 uncorrected. Over all 18 cells the crop is significantly better than D in 4, T in 9 and R in 9, with one borderline loss against D on Qwen3-VL-8B, HR-8K b256 (−2.88, p ≈ 0.05). On Qwen3-VL-8B this series uses localization layer 26, re-selected by hit rate, which favours the crop (Appendix B.9).

3. What the resolution comparison establishes. We agree that raising the baseline's resolution changes the information it receives. G tests how to allocate answering tokens; it does not isolate localization at fixed resolution. R does: the main image is pixel-identical, and the random box has ViCrop's dimensions and exactly its tokens. On Qwen3-VL-8B, V*Bench b256, ViCrop gains +14.7±6.8 over R but −1.0±7.3 against G (Table 13: 61.3% at 346 tokens against 62.3% at 328). Placement helps without the crop beating resolution, and the interval against G does not establish equivalence. The asymmetry lies in the published comparison, which gives the method a second image while holding the baseline at the original cap.

Our conclusions concern the tested budgets. ViCrop's published LLaVA-1.5 gain stands, HiDe keeps a gain at its published 16,384-token setting (+4.00 [0.50, 7.62] on HR-Bench-4K, Appendix B.11), and saturation alone does not make a crop pay (Appendix I.2).

4. Your question: how to vary the budget while keeping resolution fair. Three designs:

  • (a) Placement: fix the main image, vary only the second view (Figure 3b, Table 10).
  • (b) Allocation: sweep the baseline over eleven budgets and place the method on its accuracy–token curve (Figure 3a for V*Bench, Figure 8 for all five benchmarks); ViCrop lies below it in 23 of 30 cells.
  • (c) Box quality: fix the token count exactly (522 per question) and vary only the crop's content (Section 5.1). The gain turns positive between coverage 0.17 and 0.25 on Qwen2.5-VL and 0.36 and 0.49 on Qwen3-VL-8B. Since the box is scaled about the target, this is a sensitivity analysis, not a validated threshold for real attention boxes.

5. Beyond "accurate boxes are important". Two results go further. Under pasting, accuracy is non-monotone in box position: a box beside the target scores 7.8 to 10.8 points below a distant one (Section 6.1, Figure 11). And Table 3 states in which regimes a crop is worth its tokens at the tested budgets.

Writing. The sentence you quoted is gone. Its two claims now stand separately, each with its condition: Section 4.2 opens with "Given a choice, the whole image at higher resolution usually wins", and the HiDe result is the third row of Table 3.

Reviewer_ZGo9OpenReview ↗

Response to authors

Thank you for your response and the additional experiments, which have partially addressed my concerns. However, as noted in Weakness 1, I still find much of the manuscript difficult to read and understand. Substantial restructuring and revision are needed before it meets the standard for acceptance.

AuthorsOpenReview ↗

Revised PDF uploaded

Thank you for reading the revision and for the clear verdict. We took "substantial restructuring" literally: the revised PDF just uploaded rewrites the main text for readability, with no change to results or claims (summary in our comment to all reviewers). In brief, the prose is about 30% shorter, with confidence intervals moved into the tables. The findings are organized as three regimes stated in the abstract and introduction. Sections 4–6 have one subsection per finding, each titled with its conclusion. If specific passages are still hard to follow, we would be grateful to know which ones.

Reviewer_ZGo9OpenReview ↗

Response to authors

Thank you for the authors' efforts. I have read the revised paper, the long sentences have been improved. However, I believe that beyond the revisions to wording and length, the overall structure of the paper also needs to be revised. I think this requires time and effort to polish, rather than making some textual edits within just a few days. I maintain my score.

Reviewer_rYJnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper identifies an intriguing phenomenon termed the "near miss." For certain training-free visual cropping methods, a slight deviation in the crop’s accuracy can actually be worse than a larger deviation. This occurs because a "near miss" crop—when enlarged—ends up encroaching upon the space that should have been occupied by the correct region, thereby degrading model performance. Furthermore, the authors observe that while many methods claim performance gains, these are frequently achieved at the cost of higher computational requirements. In the context of newer VLMs—such as Qwen2.5-VL—this strategy of trading compute for performance is actually less effective than simply increasing the resolution of the original input image. Although the paper does not introduce a groundbreaking new method, it offers a fresh perspective and a new line of reasoning for this category of visual cropping techniques.

Critical strengths

This paper features a sound experimental setup and an extensive range of experiments, revealing many interesting phenomena. It allows readers to re-examine previous approaches in visual cropping area and evaluate their net contributions. Furthermore, the paper establishes more rigorous evaluation criteria, which will help guide future research.

Critical weaknesses

My primary concern regarding this paper stems from its definition of a "near miss." The paper identifies a "near miss" scenario occurring when an image is cropped and then pasted back into its original position; however, in practice, such an approach is rarely used—most methods instead provide a separate image as an additional input. Furthermore, the paper notes that only 7.5%–10% of "missed" crops fall into this "near miss" category, which makes the phenomenon seem rather trivial and of little practical significance.

Questions for authors
  1. Has the issue of the relatively low replication results for ZoomEye and Hide been resloved so far? What is the reason for this?
  2. Lines 296–298 state that a gain must be greater than 8.98 to be considered significant on Vbench; however, in Table 3 on page 16, why is a value of +7.85 also marked as significant?
1 comment
AuthorsOpenReview ↗

Response to Reviewer rYJn

Thank you for the careful reading. Your summary states the paper's two findings as we intended them. We take the scope question first, since it bears on the title.

W1 — the trough needs pasting, which crop-and-append methods do not use. We agree on the scope of the trough, and Section 6 and Figure 11 show this explicitly: it belongs to the pasting primitive, and resampling removes it. Two points place it in context.

  • For separate-image methods, the analogue of a near miss is a box that covers only part of the target. Section 5.1 and Figure 4 measure its cost by scaling the annotated box about its centre at a fixed 522 tokens: the gain turns positive between coverage 0.17 and 0.25 on Qwen2.5-VL and between 0.36 and 0.49 on Qwen3-VL-8B. Attention boxes cover 0.181 to 0.327 of the target, near that crossing, and on Qwen3-VL-8B a crop covering about 4% of the target scores 28.3 points below the same tokens spent on resolution.
  • In-place magnification is a published design, attractive because it adds no tokens. AttWarp's released warp, given a box displaced by three diagonals, falls 12.6, 11.0 and 24.1 points below no crop on the three models (Section 6.2). It has no trough, but the same dependence on the box.

W2 — only 7.5–10% of misses fall in the near-miss band. That figure is correct, and it bounds how often the displacement-defined trough occurs. A separate analysis (Appendix D.12, also in the submitted version) asks a different question: under the tested k = 3 pasting, a real box's patch is three times its size, so it can cover the target even when the box lies outside the band. It does on 42.9% to 45.1% of attention misses, and those misses lose 12.46 [3.12, 21.89] points more under pasting, while resampling shows no such gap, +2.16 [−7.76, +12.01]. The two groups are observational, and this result concerns pasting. For crop-and-append methods, the fixed-token coverage sweep of Section 5.1 tests the cost of incomplete coverage directly.

Q1 — replication.

HiDe: reproduced in its pinned environment. Our submitted run executed HiDe's code inside our harness (transformers 5.15.1, eager attention) and reached 84.82 on V*Bench. In the environment its code pins (transformers 4.57.1, FlashAttention-2), the same code reaches 90.58 against the published 92.1, with the baseline at 79.06 against 79.1. Both evaluation scripts score the outputs identically, so scoring is not the cause; the library version and the attention kernel changed together, and we do not attribute the gap to either alone (Appendix B.11). At this setting HiDe's gain survives total-compute matching on HR-Bench-4K, +4.00 [0.50, 7.62] (p = 0.024); on HR-Bench-8K it is +4.50 [−0.25, 9.25] (p = 0.064).

ZoomEye: not resolved. The library versions its code pins, its FlashAttention-2 kernel and the option order each leave the gap in place: 45.03 and 51.83 without and with search on Qwen2.5-VL-7B, and 42.93 and 67.02 on the 3B model, against the published 76.96 and 89.01 for the 3B model (Appendix B.2). We therefore report ZoomEye's compute as it ran and do not interpret its accuracy.

Q2 — 8.98 versus +7.85*. 8.98 points is the minimum detectable effect at 80% power for n = 191, not a significance threshold. An effect below it can still reach p < 0.05, as +7.85 does. Section 3.3 now says so.

Reviewer_xkYXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a protocol that focuses on compute budget and the accuracy gains, with either where the token budget is distributed in the question-answering or localization. Studies two methods, ViCrop and HiDe based on this protocol. It also experiments on the cropped boxes impact by displacing them. Overall, the paper tried to obtain token matched visual cropping protocol where deciding which part of image (localization) and using this crop for question answering use the same token budget.

Critical strengths
  1. Studies an interesting aspect of efficient visual question answering and high resolution VQA by providing a joint protocol that keeps cropping budget (finding/localizing) what to crop and further using this to perform VQA under the same token budget.
  2. Provides Visual tokens processed per question for Qwen2.5-VL-7B baseline with different settings like: no crop, ViCrop (with/without localization), HiDe, ZoomEye, and show that even though these methods use crops, they use more forward passes than the baseline.
Critical weaknesses
  1. The paper is very poorly written and is very hard to follow, even after multiple careful readings.
  2. The paper's protocol only works with older models like LLaVA-1.5 where AnyRes was not introduced. It doesn’t provide any solution to the models currently used like LLaVA-OneVision, Intern3.5VL, Qwen3VL, etc.
  3. Section 4, is very dense, hard to understand and contains only readouts of the performances, but doesn’t provide any additional information. It is hard to understand if by matching the compute-budget are the cropping methods better than baseline in which configurations.
  4. Overall, the paper needs a lot of restructuring for presentation of results, protocol definition and usage of the protocol. Without these keys components, and readability, it is hard to evaluate how good the proposed protocol works and can be scaled to the new models.
Questions for authors
  1. Please specify which three models Line 302 is referring to. Also, what is the oracle crop? Is it something that is present in the datasets annotations? Please clarify.
  2. Can the authors comment on attention sink’s role with reference to section 5.1 about why most attention-derived crops miss the target.
  3. Section 5.2 needs restructuring. Can the authors first provide a reference table containing different models being tested, their original setup, cropped setup, matched-cropped setup, oracle setup (assuming annotations of the bounding box are known). The setup needs to contain what is the token budget, and for each token budget what is the performance of the model. By doing so, we can see a clear trend of where to use the budget to obtain better performance.
  4. Can the authors comment on reusing caching for localization? How will this change the protocol?
  5. Since the protocol is based on inference time token-budget evaluation, will the similar observations hold true for bigger models?
  6. It would be better if the protocol is presented as a part of the evaluation pipeline in existing evaluation frameworks like VLMEvalKit or lmms_eval.
  7. Why are all the important tables like Table 7 only referred to in the main text and moved to supplementary.
3 comments
AuthorsOpenReview ↗

Response to Reviewer xkYX

Thank you for the detailed questions. Your summary of strengths is the paper's aim: one protocol that charges localization and answering against the same token budget, and the observation that these methods use more forward passes than their baselines.

W2 — applicability beyond LLaVA-1.5. The protocol is built for dynamic-resolution backbones: for each question it reads the crop condition's realized visual tokens and runs the full-image baseline at the resolutions just below and above that count (Section 3.1). The submitted main grid (Section 3.3; Table 3 of the submission) is Qwen2.5-VL-3B, Qwen2.5-VL-7B and Qwen3-VL-8B, and LLaVA-1.5 is the fixed-resolution control. The revision adds InternVL3.5-8B and LLaVA-OneVision-7B (AnyRes) at the attainable tile settings around the crop's count (Table 14; in one InternVL3.5 cell the next setting has fewer tokens than the crop), and Qwen2.5-VL-32B, Qwen3-VL-32B and Qwen2.5-VL-72B (Table 12). LLaVA-OneVision's crop beats a full-image repeat by +7.9 on both benchmarks and trails the next tile budget by 7.9 and 4.0 points; across both tile backbones the next tile budget is significantly ahead in 4 of 7 cells.

W3 — in which configurations does cropping win? Table 3, new in the body:

  • LLaVA-1.5 (fixed resolution), V*Bench, equal tokens: the crop leads by 13.1 to 21.5;
  • HiDe at 16,384 tokens (resolution exhausted), HR-Bench-4K, total compute matched: +4.00;
  • annotated box on V*Bench, answering call matched: +19.9 to +30.4 on the main-grid models;
  • attention boxes on the Qwen models: ViCrop lies below the baseline's accuracy–token curve in 23 of 30 cells.

Section 4 asks the two underlying questions in two subsections: 4.1, placement at fixed resolution, and 4.2, the same tokens spent on resolution.

W1, W4, Q7 — writing and tables. The abstract runs problem, gap, result, implication, and each paragraph of Section 4 opens with its conclusion. The 45-configuration table (Table 3 of the submission, in its appendix) is now Table 2 in the body; the regime table (Table 3), the accuracy–token and control figure (Figure 3), and the coverage figure (Figure 4) are now in the body. Table 7 of the submission (LLaVA-1.5 same-token controls) is summarized in Section 4.1 and Table 3, in full in Appendix B.7.

Q1. Line 302 refers to Qwen2.5-VL-3B, Qwen2.5-VL-7B and Qwen3-VL-8B. The oracle crop is V*Bench's annotated box (part of the dataset), padded by 25% on each side and appended as the second image; it is compared with a baseline at its own token count. Section 4.1 now defines it at first use.

Q2 — attention sinks. Masking visual tokens whose generic-prompt attention exceeds five times the mean does not reliably improve localization (Appendix C.5). Counting a hit when the box contains the target's centre, Qwen2.5-VL-7B's V*Bench hit rate goes from 24.1% to 23.0%, and Qwen3-VL-8B's falls from 48.7% to 31.9%, where the redrawn crops score 8.9 to 14.0 points below their token-matched baselines. What limits the boxes is coverage: they cover 0.181 to 0.327 of the target, near where a crop starts to pay (Section 5.2).

Q3 — reference table. Table 13 (Appendix B.13) gives accuracy and realized visual tokens for no crop, ViCrop and the matched baseline, for the three main-grid models on V*Bench and HR-Bench-4K at budgets 256 and 512, plus the oracle on V*Bench at 256. For Qwen2.5-VL-7B on V*Bench at 256: no crop 51.8 at 245 tokens, ViCrop 58.6 at 389, matched baseline 53.9 at 379, oracle 80.6 at 294 against 48.7 for its own baseline at 280. Table 2 gives all 45 gains; Figure 8 sweeps the baseline over eleven budgets.

Q4 — caching. Caching the image prefix can make localization cheaper, but not the answering call (Appendix A.4). The main comparison (Table 2) matches the answering call alone, so its verdicts are unchanged; caching would reduce the localization cost charged in the total-compute comparisons (Section 5.3, Appendix B.3). We count tokens and did not time cached execution.

Q5 — larger models. Yes, in the direction the paper predicts: the gain follows the box (Table 12). On V*Bench, counting the target's centre inside the box, Qwen2.5-VL-72B hits 49.2% and Qwen3-VL-32B 56.5%, against 30.4% for Qwen2.5-VL-32B. The crop is ahead of the lower-bracket matched baseline in all six V*Bench cells, and on the 72B model one comparison at each bracket is significant. On HR-Bench-4K the boxes hit 6.6% to 11.5% and no matched gain is significant. Two qualifications (Appendix B.12): these layers were chosen by hit rate, which favours the crop, and with the released layer 22 the 72B crop is −2.09±6.82 against a higher-resolution baseline, not significant.

Q6 — lmms-eval. The protocol is an lmms-eval task for V*Bench (Appendix B.15, code in the artifact). On Qwen2.5-VL-7B it returns 48.2 for no crop, 57.1 for ViCrop, and 51.8 and 50.3 for the lower and upper brackets.

I would like to thank the authors for the explanation and answers. However, the main concern of the paper is that it is very hard to read and understand, and in its current state, it is not ready for acceptance, I would like to keep the my original score.

AuthorsOpenReview ↗

Revised PDF uploaded

Thank you for confirming that the AnyRes concern is resolved. On readability, the revised PDF just uploaded rewrites the main text without changing results or claims (summary in our comment to all reviewers). For your specific point that Section 4 "contains only readouts", Section 4 now has two subsections, each titled with its conclusion. Each paragraph states the finding before the number that supports it. The configuration-level readouts are in Table 2 and Figure 3. We would be grateful if you could take another look.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers. All three agree on the problem the paper addresses: published cropping gains come with more compute than their baselines receive. xkYX notes that these methods "use more forward passes than the baseline", ZGo9 that the paper "identifies unfair comparisons ... in terms of token costs" and "explains why accurate bounding boxes are important", and rYJn that it "establishes more rigorous evaluation criteria, which will help guide future research". The revision keeps that contribution and makes its answer easier to find.

What is unchanged. The protocol (Section 3), the three dynamic-resolution models of the main grid (Qwen2.5-VL-3B, Qwen2.5-VL-7B, Qwen3-VL-8B), the benchmarks and the main measurements are those of the submission. LLaVA-1.5 remains the control for a backbone that cannot buy resolution.

The answer, in one table (Table 3, new in the body). At the tested budgets a crop leads in three regimes:

  • LLaVA-1.5, fixed resolution, V*Bench: +13.1 to +21.5 over other second views at equal tokens;
  • HiDe at 16,384 tokens, resolution exhausted, HR-Bench-4K: +4.00 [0.50, 7.62] over a baseline given all of HiDe's tokens;
  • an annotated box on V*Bench, answering call matched: +19.9 to +30.4 on the main-grid models.

Otherwise the whole image at higher resolution leads: with attention boxes, ViCrop lies below the baseline's accuracy–token curve in 23 of 30 cells (Figure 8). The protocol can credit cropping, but saturation alone does not make a crop pay: all six ViCrop cells above saturation are negative (Appendix I.2).

New experiments, all in the revised PDF.

  1. Resolution-fixed controls on Qwen (Section 4.1, Table 10), for ZGo9: with the main image pixel-identical, ViCrop's crop is significantly better than a full-image repeat in 4 of 18 cells, two tiles in 9 and a same-size random box in 9, with one borderline loss against the repeat on Qwen3-VL-8B, HR-Bench-8K b256 (−2.88, p ≈ 0.05).
  2. Coverage at a fixed token count (Section 5.1, Figure 4), for rYJn: the gain turns positive between coverage 0.17 and 0.25 on Qwen2.5-VL and between 0.36 and 0.49 on Qwen3-VL-8B; attention boxes cover 0.181 to 0.327 of the target.
  3. More backbones (Tables 12 and 14), for xkYX: Qwen2.5-VL-32B, Qwen3-VL-32B, Qwen2.5-VL-72B, InternVL3.5-8B and LLaVA-OneVision-7B, plus TextVQA and an lmms-eval task (Appendix B.15).
  4. Replication (Appendix B.11), for rYJn: HiDe's released code, run in the environment it pins, reaches 90.58 on V*Bench against the published 92.1.
  5. A published in-place warp (AttWarp, Section 6.2) under displaced boxes, and attention-sink masking (Appendix C.5).

Presentation. The abstract runs problem, gap, result, implication. Section 4 answers its two questions in two subsections, and each paragraph opens with its conclusion. The 45-configuration table moved from the appendix into the body (Table 2), grouped by budget to retain every cell compactly. Figure 3 now shows accuracy–token curves and the fixed-resolution controls together; Figure 4 plots the gain against measured target coverage. The sentence ZGo9 quoted is replaced by two plain claims.

One updated number. With HiDe's code run in its own environment, its HR-Bench-8K lead over a baseline given all of its tokens is +4.50 [−0.25, 9.25] (p = 0.064), down from +6.50 in the submission and no longer significant; on HR-Bench-4K it is +4.00 (p = 0.024).

Open. Our ZoomEye run stays below its published accuracy after checking library versions, attention kernel and option order (Appendix B.2), so we report its compute and do not interpret its accuracy.

AuthorsOpenReview ↗

Revised PDF focused on readability

We thank the reviewers for the follow-up. Reviewers ZGo9 and xkYX both name readability as the remaining obstacle, so we used the last revision to rewrite the main text for that alone. No result, number or claim changed; the changes are in how the paper is organized and written.

  • Shorter, with fewer numbers in the prose. The main-text prose is about 30% shorter and carries about 40% fewer decimal values. Confidence intervals and secondary readouts now sit in the tables, figures and appendix, and each paragraph states its finding first, then gives one representative number. Sentences over 35 words dropped from 31 to 13.
  • One story stated up front. The abstract and introduction now present the results as three regimes: (1) cropping pays when whole-image resolution cannot be raised; (2) on dynamic-resolution models at low budgets, spending the same tokens on resolution usually beats attention crops, because attention boxes cover too little of the target; (3) when magnifying in place without extra tokens, the primitive decides what a near miss costs.
  • Simpler structure. Sections 3–7 keep their numbers and topics, and their subsections are merged from 13 to 8. Each subsection title in Sections 4–6 now states its finding (for example, "On dynamic-resolution backbones, resolution usually beats attention crops").
  • Clearer setup and transitions. Section 3.2 now lists the benchmarks by type and states which model each method runs on. The introduction states why the paper turns to in-place magnification. The Discussion now gives separate takeaways for crop-and-append and for in-place methods.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.