acceptodds
Under review as a conference paper at ICLR 2027

FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution

Abstract

Long-context reasoning in large language models incurs substantial computation and memory costs. Visual text compression (VTC) reduces input length by rendering text as images, but fixed-resolution rendering creates a compression–performance trade-off: low DPI saves tokens at the expense of legibility, whereas high DPI spends tokens on irrelevant content. We introduce \model, which breaks this trade-off through adaptive resolution while preserving general multimodal capabilities. It combines compressed low-DPI global views with selective region enhancement, integrating enhanced views into ongoing reasoning. We construct 29.4K high-quality Reasoning–Evidence Localization (REL) chain-of-thought examples (REL-CoT) that link reasoning traces to page indices and bounding boxes. Multi-resolution REL supervised fine-tuning (REL-SFT) teaches the model to localize relevant regions, and Group Relative Policy Optimization learns when to enhance resolution and how to use the resulting observations, without a separate continual-pretraining stage. At 72 DPI on RULER v1, \model scores 87.4 at input compression, including tool observations, versus 57.5 for Glyph at input compression. It surpasses its text-input backbone on LongBench (56.40 versus 55.86), improves the MRCR macro-average by 13.91 points, and achieves a 51.19 macro-average on VTCBench. The MRCR latency evaluation also shows a online end-to-end speedup over Text. General multimodal capabilities are preserved, with MMMU increasing from 65.12 to 66.73 and MME from 2424.02 to 2457.62.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Visual Text Compression, Long-Context Reasoning, Adaptive Resolution, Evidence Localization, Vision-Language Models, Reinforcement Learning.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_n7uVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FocusVTC, an adaptive-resolution framework for visual text compression. It combines compressed low-DPI global views with selective region enhancement, integrating enhanced views into ongoing reasoning.

Critical strengths

The paper addresses the compression–performance trade-off in visual text compression, which is an important and practical problem.

Critical weaknesses
  1. The technical contributions are largely incremental. The core idea of adaptive visual text reading has already been explored in AGAR and SEER, while DeepEyes has demonstrated RL-based active visual inspection. FocusVTC mainly integrates SFT and GRPO with region-level high-resolution enhancement.

  2. The reward design does not explicitly optimize the actual benefit of resolution enhancement, potentially encouraging unnecessary tool calls. Moreover, the call-efficiency reward measures average localization quality without explicitly promoting complete evidence coverage. The GRPO data filtering further excludes samples where no trajectory invokes tools, potentially biasing the policy toward tool usage. These design choices raise concerns about whether the model genuinely learns when enhancement is necessary.

Questions for authors
  1. The contribution of the constructed training data is insufficiently isolated. The method relies on 29.4K externally generated and verified reasoning–evidence localization examples, expanded across multiple resolutions. However, the paper does not adequately disentangle the benefits of additional training data, evidence localization annotations, and the proposed adaptive strategy.

  2. The reinforcement learning design lacks sufficient ablation. The reward combines answer correctness, localization IoU, resolution-dependent weighting, and call efficiency. However, the individual contributions of these components are not systematically evaluated.

  3. The reported 2.79× end-to-end speedup excludes offline document rendering. Additionally, the text baseline generates substantially more output tokens than FocusVTC, making it difficult to separate input-compression benefits from differences in generation behavior. A more comprehensive cost analysis covering rendering, tool interactions, and matched inference conditions would strengthen the efficiency claims.

  4. The experiments primarily rely on a single 9B backbone and rendered text documents. The effectiveness across different architectures, languages, and realistic document layouts remains unclear. Moreover, FocusVTC underperforms the original Qwen3.5-9B visual baseline on VTCBench Reasoning (36.25 vs. 41.57), while its improvement over the text-input backbone on LongBench is relatively marginal (56.40 vs. 55.86). These results suggest that the benefits are task-dependent rather than uniformly generalizable.

8 comments
AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Question 1 (Part 1/2)

Thank you for examining our supervision, reward design, and efficiency analysis in detail. In response, we provide additional ablations and timing experiments, and clarify the scope and limitations of the evaluation. We address each concern below.

W1 & Q1. The contribution appears incremental relative to AGAR, SEER, and DeepEyes. The gains from the 29.4K examples expanded across resolutions, evidence annotations, and the adaptive strategy are insufficiently separated.

A: We agree that prior work has explored adaptive visual reading and RL-based active inspection. In multi-page, long-context VTC, the trade-off between text legibility and input cost requires identifying relevant pages and regions before acquiring clearer observations. Our contribution is coupling cross-resolution localization alignment (REL-SFT) with cost-aware policy learning (GRPO) to enable selective reading of low-resolution, blurred multi-page text. We distinguish this contribution through method comparisons, internal controls, and a bbox-supervision ablation.

(1) Methods comparisons. The cited methods use distinct mechanisms. AGAR [1] uses training-free attention-guided re-rendering for VTC, enlarging selected text spans. SEER [2] learns page-level selection followed by retrieval of the selected pages' original text. DeepEyes [3] learns active crop/zoom behavior through RL without cold-start SFT.

For this rebuttal, we add new evaluations of AGAR and DeepEyes. AGAR is applied to the original Qwen3.5-9B. Our multi-page adaptation of DeepEyes v1 accepts complete page lists and routes regional operations to the appropriate page, retaining its released weights and core mechanism without additional training. These direct evaluations use 72-DPI document inputs and the same questions and scoring protocols as the paper, covering 10 RULER v1 tasks, 12 RULER v2 task–difficulty combinations, and LongBench tasks.

Tab. R12. Adaptive-reading comparisons: scores and average compression.

MethodCompressionRULER v1RULER v2LongBench
Qwen3.5-9B (vision input, no tools)3.83×37.4044.6435.86
AGAR [1] (training-free transfer)1.86×51.6351.5837.06
DeepEyes v1 [3] (multi-page adaptation)1.10×82.1553.8048.20
SEER [2] (published scores, recomputed)———55.11
FocusVTC (Ours)2.93×87.3875.9456.40

Compression is the arithmetic mean of Text/(P+O)\mathrm{Text}/(P+O) across the three benchmarks, using the text-token references in Table 10. Here, PP denotes initial prompt tokens and OO denotes tokens from all additional observations. Qwen3.5-9B and FocusVTC scores were already reported in Table 1 and Appendix Tables 15–16; FocusVTC token costs are from Appendix Table 10. SEER scores are recomputed from published results; its average compression under this 72-DPI protocol is unavailable (—). We discuss its published cost separately because it uses a different evaluation setting.

FocusVTC outperforms the evaluated AGAR and DeepEyes variants while using fewer input tokens on all three benchmarks, including a 60.2–66.2% reduction relative to DeepEyes v1.

Under the same LongBench task configuration, FocusVTC scores 56.40 versus SEER's 55.11 unweighted average (SEER, Tables 1 and 10 [2]). We use SEER's published results because its checkpoint release is pending. Reported input costs are 3,737 tokens (3.0× compression) for FocusVTC (Table 10) and 7,623.8 tokens (1.28×) for SEER (Table 2 [2]), though the cost evaluations use different backbones, task sets, and aggregation rules.

AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Question 1 (Part 2/2)

(2) Coupled training and adaptive inference. REL-SFT establishes cross-resolution localization alignment through reasoning-linked page indices and evidence boxes. GRPO learns how to acquire and use high-resolution evidence from this aligned initialization while accounting for input cost (Secs. 3.2–3.4).

The internal ablations already reported in Table 3, reproduced in Tab. R13, explain why REL-SFT and GRPO are coupled and both are necessary in our framework. All configurations use the same 9B model and 72-DPI inputs.

Tab. R13. Training-stage and inference-tool ablations on LongBench and MRCR.

ConfigurationREL-SFTGRPOInference toolsLongBenchMRCR (2 needles)MRCR (4 needles)MRCR (8 needles)
Full FocusVTC✅✅✅56.4060.7645.2130.71
w/o tools✅✅❌36.8252.0037.6523.50
w/o SFT❌✅✅49.3057.0539.3626.66
w/o GRPO, without tools✅❌❌37.8644.9428.8321.30
w/o GRPO, with tools✅❌✅25.4122.2515.098.96

Here, “w/o GRPO, with tools” corresponds to the row labeled “w/o GRPO+tools” in Table 3: the REL-SFT checkpoint with tools enabled, without GRPO.

  • REL-SFT initialization: direct GRPO without cold-start REL-SFT reaches 49.30 on LongBench, versus 56.40 for the full model. This control removes the entire SFT stage and measures the combined contribution of its data and supervision.
  • GRPO policy learning: enabling tools immediately after REL-SFT lowers LongBench performance from 37.86 to 25.41. Coupling REL-SFT with GRPO raises it to 56.40, showing why localization alignment must be followed by learning how to acquire and use observations in our framework.
  • Inference-time enhancement: disabling tools on the fully trained checkpoint lowers LongBench from 56.40 to 36.82 and reduces all three MRCR scores, isolating the contribution of selective observations at inference.

(3) Evidence-localization supervision. We add a new bbox-supervision ablation (T1) using the same SFT examples, multi-resolution inputs, and training settings, removing only bounding boxes from the SFT targets while retaining reasoning, page indices, and answers. Both the full REL and T1 branches then undergo 150 GRPO updates, with unchanged GRPO data, rewards (including localization rewards), and tool interfaces. Tab. R14 compares the new T1 results with the full-REL controls from Table 3 at 72 DPI.

Tab. R14. Effect of SFT bounding-box supervision before and after GRPO.

ConfigurationInference toolsRULER v1RULER v2LongBench
Full REL-SFT❌32.2843.5837.86
T1: SFT without bbox❌31.6541.0133.40
Full REL-SFT + GRPO✅87.3875.9456.40
T1: SFT without bbox + GRPO✅78.5867.0847.86

With the training examples and resolution exposure held fixed, the full model's advantage across all three benchmarks supports the contribution of explicit region supervision during SFT to subsequent tool learning. The difference therefore cannot be attributed to the full model receiving more training examples.

T1 specifically isolates SFT bbox supervision; it does not separately quantify the effects of reasoning/page annotations or multi-resolution expansion.

AuthorsOpenReview ↗

Rebuttal of the Weakness 2

W2. The reward does not explicitly optimize enhancement benefit and may encourage unnecessary calls. Call efficiency averages localization quality without explicitly promoting complete evidence coverage. Excluding all-no-tool trajectory groups may also bias the policy toward tool use, raising doubts about whether it learns when enhancement is necessary.

A: We address tool utility, evidence coverage, and filtering separately.

(1) Tool utility and cost. The reward balances accuracy and input cost while allowing verification; it does not explicitly measure each call's marginal benefit. The tool bonus requires a correct answer and combines localization, call efficiency, and resolution weighting. In our new ablations (Tabs. R15 and R16), removing resolution weighting or call efficiency increases RULER v1 observation tokens by 27.4% and 128.4% at 72 DPI, with similar scores; removing localization shaping lowers scores and raises observation cost. Our new comparison using the same checkpoint also shows 86.45 with 3,427 total input tokens at 84 DPI with tools, versus 61.20 with 3,544 tokens at 96 DPI without tools on RULER v1.

(2) Evidence coverage. The IoU reward encourages coverage of annotated regions, while one-to-one matching prevents repeated credit for the same evidence. Replacing redundant crops with well-localized crops of uncovered evidence can improve reward without adding calls. Together with answer-correctness gating, this encourages useful, non-redundant evidence acquisition, although it does not explicitly maximize recall over all annotated regions.

(3) Filtering and tool-use bias. The no-tool filter selects prompts with observed opportunities to learn tool use: one tool-using trajectory among eight is enough to pass this rule. Online rollouts are sampled afresh, and a correct, well-formatted answer without tools still receives R=1R=1 (Table 9). At 144 DPI, g(144)=0g(144)=0 removes the additional tool bonus. Thus, direct answering remains allowed, while unnecessary tool use receives less incentive at high resolution. The filtering step itself has not been separately ablated.

Reviewer_szj7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets the compression–performance trade-off in visual text compression for long-context reasoning, where lower-resolution rendering reduces token cost but may lose locally important textual details. The proposed FocusVTC adopts an adaptive-resolution strategy, where the model reasons over low-DPI rendered pages and invokes a region-enhancement tool to inspect selected areas at higher resolution when needed. The authors construct 29.4K Reasoning–Evidence Localization multi-resolution examples and train the model with SFT followed by GRPO to learn when and where to request higher-resolution evidence. Experimental results demonstrate substantial gains over fixed-resolution baselines across multiple benchmarks.

Critical strengths
  1. The paper is clearly written, with the data construction, method, and experimental results presented in a straightforward manner. The figures are well-illustrated and helpful.
  2. The experimental design is straightforward and easy to follow.
  3. The training process and ablation analyses are sufficiently detailed, providing a clear picture of how SFT, GRPO, and tool access contribute to the final performance.
Critical weaknesses
  1. My main concern is the limited novelty of the overall approach. Feeding low-resolution inputs to the model and allowing it to selectively retrieve higher-resolution evidence appears to be a common strategy in recent work [1–3], while this paper mainly extends this similar idea to text-intensive multimodal reasoning. Although the authors construct 29.4K multi-resolution training examples and apply SFT followed by GRPO, I am not fully convinced that the paper offers a sufficiently novel methodological component or deeper analysis under this specific task setting.
  2. The empirical comparison is somewhat incomplete. Since FocusVTC is closely related to prior work, these recent methods [1-3] should also be included as baselines for comparison. Specifically, a direct comparison with [2] is helpful, which appears to be the most closely related method in terms of coarse-to-fine visual acquisition.
  3. The current ablation does not fully isolate the benefit of adaptive region enhancement from the effect of input resolution itself. A more controlled comparison would be helpful. Specifically, comparing (i) low-resolution page inputs with selective region enhancement and (ii) uniformly higher-resolution page inputs without enhancement under matched token or computational budgets (using the same multi-DPI post-trained checkpoint) would help isolate the value of adaptive enhancement from that of higher resolution alone.

[1] VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning. NIPS 2025.
[2] AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition. CVPR 2026.
[3] Q-Zoom: Query-Aware Adaptive Perception for Efficient Multimodal Large Language Models. Arxiv: 2604.06912.

Questions for authors

See weakness

5 comments
AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Weakness 2 (Part 1/2)

Thank you for your thoughtful comments on the relation to prior methods and the evaluation of adaptive enhancement. We appreciate your suggestions for direct baselines and a comparison at similar input-token costs. Below, we address the methodological contribution, the requested baselines, and the resolution control in turn.

W1 & W2. Selective high-resolution retrieval is already common in prior work. The 29.4K multi-resolution examples and SFT+GRPO pipeline do not clearly establish a novel component or deeper task-specific analysis. Comparisons with the cited methods [1–3] are missing, especially the closely related AdaptVision [2].

A: We agree that using low-resolution inputs with selective high-resolution acquisition is an established strategy. In multi-page, long-context VTC, the trade-off between text legibility and input cost requires selecting evidence from low-resolution views with blurred text. Our contribution is coupling cross-resolution localization alignment (REL-SFT) with cost-aware policy learning (GRPO). As the internal ablations in Tab. R8 show, neither direct RL without cold-start REL-SFT nor SFT followed by immediate tool access achieves the full selective-reading performance. REL-SFT establishes localization alignment across resolutions, and GRPO learns how to acquire and use high-resolution evidence from that initialization while accounting for input cost. We explain this coupling through method comparisons, internal controls, the bbox-supervision ablation, and reward-component ablations.

(1) Methods comparisons.

VisionThink [1] requests higher-resolution full-image views, AdaptVision [2] learns crop acquisition through DTPO, and Q-Zoom [3] uses learned gating and self-distilled region proposals. Our REL-SFT links reasoning steps to page indices and normalized evidence boxes, with shared targets across seven rendering resolutions (Secs. 3.2–3.3). The resulting release contains approximately 29.4K verified REL-CoT examples. We use task-specific GRPO to train the model to acquire and use high-resolution observations (Sec. 3.4). For this rebuttal, we add evaluations of all three methods, using VisionThink-Efficient for VisionThink. For AdaptVision and Q-Zoom, we extended their released single-image region-processing interfaces to multi-image documents, retaining their weights and core mechanisms without additional training. Each method retains its own inference budget. Evaluation uses 72-DPI document inputs and the same questions and scoring protocols as the paper, covering 10 RULER v1 tasks, 12 RULER v2 task–difficulty combinations, and LongBench tasks.

Tab. R7. VisionThink-Efficient, AdaptVision, and Q-Zoom at 72 DPI: scores and average compression.

MethodCompressionRULER v1RULER v2LongBench
Qwen3.5-9B (vision input, no tools)3.83×37.4044.6435.86
VisionThink-Efficient [1]0.98×79.0550.9346.61
AdaptVision [2] (multi-page adaptation)2.26×14.6828.0938.41
Q-Zoom [3] (multi-page adaptation)2.09×53.3539.4742.17
FocusVTC (Ours)2.93×87.3875.9456.40

Compression is the arithmetic mean of Text/(P+O)\mathrm{Text}/(P+O) across the three benchmarks, using the text-token references in Table 10. Here, PP denotes initial prompt tokens and OO denotes tokens from all additional observations. Qwen3.5-9B and FocusVTC scores were already reported in Table 1 and Appendix Tables 15–16; FocusVTC token costs are from Appendix Table 10.

FocusVTC scores higher and uses fewer input tokens than all three cited methods. Compared with VisionThink-Efficient, which requests higher-resolution full-image views, it uses 62.7–69.7% fewer input tokens across the three benchmarks, including all additional observations. AdaptVision's low RULER scores may reflect limited transfer of its single-image crop policy to evidence selection across many compressed pages.

AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Weakness 2 (Part 2/2)

(2) Methodological components and deeper analysis under this specific task setting.

(i) Internal ablations in Table 3. Table 3 explains why REL-SFT and GRPO are coupled and both are necessary in our framework. Its same-backbone controls also isolate the contribution of inference-time tool access. Tab. R8 reproduces the RULER v1/v2 and LongBench results already reported in Table 3; all configurations use Qwen3.5-9B and 72-DPI document inputs.

Tab. R8. Effects of REL-SFT, GRPO, and inference-time tools across three benchmarks.

ConfigurationREL-SFTGRPOInference toolsRULER v1RULER v2LongBench
Full FocusVTC✅✅✅87.3875.9456.40
w/o SFT❌✅✅73.2162.7749.30
w/o GRPO, without tools✅❌❌32.2843.5837.86
w/o GRPO, with tools✅❌✅22.9839.4125.41
w/o tools✅✅❌36.6044.8236.82

Here, “w/o GRPO, with tools” is labeled “w/o GRPO+tools” in Table 3. On low-resolution, blurred multi-page text, direct RL without cold-start REL-SFT lowers LongBench performance from 56.40 to 49.30, while SFT followed by immediate tool access without GRPO falls to 25.41, below 37.86 without tools. Neither approach alone achieves the full result. Coupling cross-resolution localization alignment with cost-aware policy learning reaches 56.40: REL-SFT grounds evidence localization across resolutions, and GRPO turns this alignment into effective acquisition and use of high-resolution observations. These ablations show that the full selective-reading performance depends on coupling localization alignment with cost-aware policy learning.

(ii) Bbox-supervision ablation (T1). To examine the contribution of explicit evidence-region supervision, we additionally conduct a new bbox-supervision ablation (T1) using the same SFT data and settings, removing only bounding boxes from the SFT targets while retaining reasoning, page indices, and answers. Both branches then undergo the same GRPO training with unchanged data, rewards, and tool interfaces. This preserves the selective-reading strategy while testing the contribution of explicit region supervision. The SFT-only variants in Tab. R9 are evaluated without tools, while the post-GRPO variants use tools. The full-REL controls are from Table 3 of the paper.

Tab. R9. Effect of SFT bounding-box supervision before and after GRPO.

ConfigurationInference toolsRULER v1RULER v2LongBench
Full REL-SFT❌32.2843.5837.86
T1: SFT without bbox❌31.6541.0133.40
Full REL-SFT + GRPO✅87.3875.9456.40
T1: SFT without bbox + GRPO✅78.5867.0847.86

The consistent post-GRPO advantage across all three benchmarks supports the contribution of explicit evidence-region supervision during SFT to subsequent tool learning.

(iii) GRPO reward-component ablations (T2–T4). We additionally ablate the task-specific tool-reward components: T2 removes DPI-dependent weighting by setting g(d)=1g(d)=1; T3 removes the call-efficiency factor EcallE_{\mathrm{call}}; T4 removes the combined IoU- and area-based localization shaping by setting Qloc=1Q_{\mathrm{loc}}=1. All variants start from the same REL-SFT checkpoint and undergo 150 GRPO updates, with answer-correctness and format rewards unchanged.

Tab. R10 compares new T2–T4 experiments with the paper-reported full-reward baseline (Table 3; Appendix Table 10). All configurations use 72-DPI inputs and tools enabled. Here, OO denotes additional tool-observation tokens on RULER v1; the percentage is the increase over the full-reward baseline. Initial prompt tokens are identical across these configurations.

Tab. R10. GRPO reward ablations at 72 DPI: benchmark scores and RULER v1 observation-token costs.

ConfigurationRULER v1RULER v2LongBenchRULER v1 observation tokens OO
Full reward (FocusVTC)87.3875.9456.40723
T2: w/o DPI weighting86.8374.2256.27921 (+27.4%)
T3: w/o call-efficiency factor87.1675.4155.971,651 (+128.4%)
T4: w/o localization shaping (IoU + area)80.8768.2450.761,450 (+100.6%)

T2 and T3 retain relatively similar benchmark averages but increase observation cost by 27.4% and 128.4%, respectively. T4 produces the largest score drop on all three benchmarks while increasing observation cost by 100.6%. These results clarify the components' roles in the accuracy–input-cost trade-off: DPI weighting and call efficiency primarily control observation expenditure in this setting, while localization shaping supports both task performance and input efficiency.

AuthorsOpenReview ↗

Rebuttal of the Weakness 3


W3. The current ablation does not fully separate adaptive enhancement from higher input resolution. A controlled comparison should use the same multi-DPI-trained checkpoint and matched token or computational budgets for low-resolution inputs with enhancement versus uniformly higher-resolution inputs without enhancement.

A: Your proposed control helps distinguish the benefit of selective enhancement from simply increasing input resolution.

We have added the requested comparison using the same multi-DPI-trained FocusVTC checkpoint, pairing lower-resolution inputs with selective enhancement against uniformly higher-resolution inputs with tools disabled. We compare 72 DPI with tools vs. 84 DPI without tools, and 84 DPI with tools vs. 96 DPI without tools, which have approximately matched input-token costs.

Following Table 10, input cost is P+OP+O for adaptive enhancement, including all additional tool observations, and PP for the tool-disabled setting. The adaptive results were already reported in the paper's resolution-sweep experiments (plotted in Fig. 6); the 84/96-DPI tool-disabled scores are new evaluations for this rebuttal. Token counts are from Appendix Table 10. RULER v1 uses the same 10-task average as the paper.

Tab. R11. Selective enhancement versus uniformly higher-resolution inputs at similar input-token costs.

ConfigurationInitial DPIRULER v1 scoreRULER v1 input tokensRULER v2 scoreRULER v2 input tokens
Adaptive enhancement7287.382,87775.943,251
Uniform higher resolution, no tools8451.232,72959.833,050
Adaptive enhancement8486.453,42777.233,884
Uniform higher resolution, no tools9661.203,54468.153,731

The input-token differences are +5.4% / +6.6% for the first pair and −3.3% / +4.1% for the second pair on RULER v1/v2, respectively, relative to the tool-disabled setting. Adaptive enhancement achieves higher scores on both benchmarks at both budget levels. In particular, on RULER v1, 84-DPI adaptive enhancement outperforms 96-DPI uniform input while using fewer input tokens. These results support the value of selectively acquiring high-resolution evidence beyond uniformly increasing page resolution at similar input-token costs.


Thank you again for helping us clarify the relation to prior methods and evaluate selective enhancement at comparable input cost. We hope these comparisons and ablations address your concerns.


[1] VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning. NeurIPS 2025.

[2] AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition. CVPR 2026.

[3] Q-Zoom: Query-Aware Adaptive Perception for Efficient Multimodal Large Language Models. Arxiv 2026.

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

This paper proposes an adaptive visual text compression method, FocusVTC. The idea of FocusVTC is: retain a low-DPI global page view as the compressed context, while during inference letting the model selectively read relevant regions from aligned high-DPI pages through tool calls, thereby combining the benefits of both global coverage and local refinement. The authors construct REL-CoT data with reasoning and evidence localization annotations, and combine REL-SFT and GRPO to train the model to learn evidence localization, region enhancement, and tool-call policies. It performs well on multiple benchmarks.

Critical strengths
  1. Accurate problem positioning, simple and self-consistent method design. The core idea is very simple: decouple the resolution required for "global coverage" and "local fine reading," and implement it with high-DPI crops using aligned coordinates. The training part is also very clearly designed, and it explains the necessity of the two training stages. Due to the clarity of the design and writing, I was able to make judgments more accurately.

  2. Relatively comprehensive experimental coverage, balancing long-context and general capabilities. This paper does not sacrifice general performance specifically for improving performance on its own task, and honestly reports issues such as font-selection sensitivity. This improved my overall impression of the paper.

Critical weaknesses
  1. Insufficient methodological novelty. The two-stage "global first, then local" approach is not uncommon in visual reasoning. The current experiments mainly demonstrate the effectiveness of this mechanism in VTC and long-context tasks, but lack direct comparison with the closest adaptive visual reading/region-selection methods. This directly affects my evaluation rating of the paper.

  2. Lack of key controls, unclear attribution of gains. All baselines (Glyph, GLM-4.1V) did not undergo equivalent 29.4K data SFT + GRPO training, and the "w/o tools" and "w/o GRPO+tools" in Table 3 also respectively lack GRPO or the tool interface, making it impossible to distinguish whether the gains come from adaptive resolution, the GRPO tool policy, or REL-SFT supervision. This makes it difficult to judge the true reason for the method's improvement, affecting my judgment of the method's true performance.

Questions for authors
  1. This paper mainly compares with fixed-resolution VTC methods and internal ablation experiments. Then how do the other methods that also use a global + local two-step approach, especially training-free methods, compare in performance?

  2. Could multiple-run results or significance tests be added for some experiments, especially for those experiments with small improvements?

  3. Could experiments be conducted on models of more scales to prove that this method is not only effective for Qwen3.5-9B?

  4. I think the current training cost is overall relatively high (a lot of LLM annotation + large data volume). Is there a scheme that can reduce the cost, for example, can the training data volume be reduced by methods such as "selecting excellent samples"?

6 comments
AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Question 1 (Part 1/2)

Thank you for your careful review and constructive suggestions. We have added comparisons with adaptive reading methods, repeated LongBench evaluations, and experiments on a 2B model. We also clarify the baseline training and ablation settings below to address your concerns about the sources of the gains.

W1 & Q1. The global-first/local-second approach has limited novelty, and comparisons with the closest adaptive reading methods are missing. How does FocusVTC compare with other global-local methods, particularly training-free approaches?

A: We agree that global-to-local reading with selective high-resolution acquisition is an established strategy. In multi-page, long-context VTC, the trade-off between text legibility and input cost requires the model to select relevant pages and regions from low-resolution views with blurred text. Our contribution is coupling cross-resolution localization alignment (REL-SFT) with cost-aware policy learning (GRPO) to make selective reading effective under this trade-off (Secs. 3.2–3.4).

Table 3 explains why these two stages are coupled and both are necessary in our framework. Direct RL without cold-start REL-SFT (w/o SFT) lowers the LongBench score from 56.40 to 49.30. SFT followed by immediate tool access, without GRPO (w/o GRPO+tools) falls to 25.41, even below 37.86 with tools disabled. Thus, neither direct RL nor SFT with tool access alone is sufficient to achieve the full selective-reading performance on blurred multi-page text. REL-SFT establishes localization alignment across resolutions; GRPO learns how to acquire and use high-resolution evidence from this aligned initialization while accounting for input cost. Coupling them raises performance to 56.40. The comparisons and bbox-supervision ablation in Tabs. R1 and R2 further examine this mechanism.

(1) Methods comparisons. For this rebuttal, we add evaluations of VisionThink-Efficient [1], which adaptively requests higher-resolution full images; AGAR [2], a training-free attention-guided re-rendering method for VTC, applied to the original Qwen3.5-9B; and AdaptVision [3], Q-Zoom [4], and DeepEyes v1 [5], which selectively process regional details. For the latter three, we extended the single-image region-processing interfaces in their released codebases to support multi-image document inputs, retaining their released weights and core mechanisms without additional training.

Evaluation uses 72-DPI document inputs and the same questions and scoring protocols as the paper, covering 10 RULER v1 tasks, 12 RULER v2 task–difficulty combinations, and LongBench tasks. Tab. R1 also includes the paper-reported Qwen3.5-9B visual baseline without tools.

Tab. R1. Adaptive visual reading methods at 72 DPI: scores and average compression.

MethodCompressionRULER v1RULER v2LongBench
Qwen3.5-9B (vision input, no tools)3.83×37.4044.6435.86
AGAR [2] (training-free transfer)1.86×51.6351.5837.06
VisionThink-Efficient [1]0.98×79.0550.9346.61
AdaptVision [3] (multi-page adaptation)2.26×14.6828.0938.41
Q-Zoom [4] (multi-page adaptation)2.09×53.3539.4742.17
DeepEyes v1 [5] (multi-page adaptation)1.10×82.1553.8048.20
FocusVTC (Ours)2.93×87.3875.9456.40

Compression is the arithmetic mean of Text/(P+O)\mathrm{Text}/(P+O) across the three benchmarks, using the text-token references in Table 10. Here, PP denotes initial prompt tokens and OO denotes tokens from all additional observations. Qwen3.5-9B and FocusVTC scores were already reported in Table 1 and Appendix Tables 15–16; FocusVTC token costs are from Appendix Table 10.

FocusVTC achieves the highest average scores across all three benchmarks while using fewer input tokens than all five adaptive-reading baselines. It uses 62.7–69.7% fewer input tokens than VisionThink-Efficient and 60.2–66.2% fewer than DeepEyes v1. AdaptVision's low RULER scores may reflect difficulties in transferring its single-image crop policy to multi-page documents, where evidence must be located across pages. The comparisons distinguish three cases:

(i) Training-free adaptation. AGAR improves over the Qwen3.5-9B visual baseline without additional training, but delivers smaller gains than FocusVTC across all three benchmarks.

(ii) Transfer of regional reading methods. AdaptVision, Q-Zoom, and DeepEyes v1 remain below FocusVTC, despite DeepEyes' competitive RULER v1 performance.

(iii) Multi-image adaptive reading. VisionThink-Efficient supports multiple input images and adaptive full-image resolution, but trails FocusVTC on all three benchmarks, most notably RULER v2.

AuthorsOpenReview ↗

Rebuttal of the Weakness 1 & Question 1 (Part 2/2)

(2) Bbox-supervision ablation (T1). To examine the contribution of region-level supervision, we additionally conduct a new bbox-supervision ablation (T1) using the same SFT data and settings, removing only bounding boxes from the SFT targets while retaining reasoning, page indices, and answers. Both the full REL and T1 branches are followed by GRPO, with unchanged GRPO data, rewards, and tool interfaces. The full-REL controls are from Table 3 of the paper.

Tab. R2. Effect of SFT bounding-box supervision before and after GRPO.

ConfigurationInference toolsRULER v1RULER v2LongBench
Full REL-SFT❌32.2843.5837.86
T1: SFT without bbox❌31.6541.0133.40
Full REL-SFT + GRPO✅87.3875.9456.40
T1: SFT without bbox + GRPO✅78.5867.0847.86

With subsequent GRPO settings unchanged, removing SFT bbox supervision lowers the final LongBench score from 56.40 to 47.86, with declines on both RULER benchmarks as well. This further supports the role of evidence-aligned initialization in subsequent acquisition learning.

AuthorsOpenReview ↗

Rebuttal of the Weakness 2

W2. Key controls are missing: Glyph and GLM-4.1V did not receive equivalent 29.4K-example SFT+GRPO training, and the Table 3 ablations leave the contributions of adaptive resolution, the GRPO tool policy, and REL-SFT unclear.

A: The sources of the gains need to be clearly separated. We clarify two aspects below: the training of the Glyph baseline and the configurations of our ablation studies.

(1) Training of the Glyph baseline. Glyph is built on GLM-4.1V-9B-Base and trained for approximately 4,000 CPT steps (batch size 170), 1,500 SFT steps (batch size 32), and 500 GRPO iterations (batch size 32, with 16 sampled responses per group), as reported in Glyph, Appendix B [6]. Thus, Glyph provides a comparison with another extensively trained VTC method, while GLM-4.1V serves as a backbone baseline. This is an appropriate method-level comparison between VTC-trained systems; controlled attribution of our gains comes from the same-backbone ablations in Tab. R4.

Tab. R3 compares Glyph and FocusVTC with their visual backbone baselines, using results already reported in the paper (Table 1; Appendix Tables 15–16).

Tab. R3. Glyph and FocusVTC with their visual backbone baselines at 72 DPI.

Benchmark / aggregateGLM-4.1V-9B (vision)GlyphQwen3.5-9B (vision)FocusVTC
LongBench Avg38.4552.3435.8656.40
RULER v1 Avg28.6857.5337.4087.38
RULER v2 Avg28.9447.9144.6475.94

FocusVTC outperforms Glyph on all three benchmarks.

(2) Configurations of the Table 3 ablations. Table 3 separates REL-SFT, GRPO, and inference-time tool access. In particular, “w/o tools” retains both REL-SFT and GRPO, disabling tools only at inference; “w/o GRPO+tools” denotes the REL-SFT checkpoint with tools enabled, without GRPO. Tab. R4 clarifies the configurations and reproduces the results already reported in Table 3.

Tab. R4. Training-stage and inference-tool ablations on RULER v1/v2 and LongBench.

ConfigurationREL-SFTGRPOInference toolsRULER v1RULER v2LongBench
Full FocusVTC✅✅✅87.3875.9456.40
w/o tools✅✅❌36.6044.8236.82
w/o SFT❌✅✅73.2162.7749.30
w/o GRPO, with tools✅❌✅22.9839.4125.41
w/o GRPO, without tools✅❌❌32.2843.5837.86

The results show why localization alignment and policy learning need to work together on low-resolution, blurred multi-page text:

  • Direct RL without cold-start REL-SFT: our w/o SFT variant follows a direct-RL strategy similar to DeepEyes [5], but removing localization-aligned initialization lowers LongBench performance from 56.40 to 49.30.
  • SFT followed by immediate tool access: without GRPO, enabling tools lowers the REL-SFT result from 37.86 to 25.41. Localization alignment alone does not teach effective tool use.
  • Coupled localization alignment and policy learning: REL-SFT followed by cost-aware GRPO reaches 56.40, exceeding both single-stage variants. Disabling tools on the fully trained checkpoint lowers the score to 36.82, confirming the contribution of selective observations at inference.

These controls show that neither stage alone delivers the full benefit in this setting. FocusVTC depends on coupling cross-resolution localization alignment with cost-aware policy learning: REL-SFT provides the grounded initialization, and GRPO turns that alignment into effective decisions about acquiring and using high-resolution evidence.

Comments on the paper

1 comment

We appreciate the reviewers’ feedback. During the rebuttal, we provided additional experiments and clarifications addressing the following concerns.

1. Methodological contribution and attribution. We clarified that our core contribution is an adaptive-resolution strategy for multi-page, long-context VTC. REL-SFT develops reasoning and localization across resolutions, enabling GRPO to learn when and where to acquire higher-resolution evidence. Internal controls, bbox-supervision ablations, and reward-component ablations support the coupling between localization alignment and cost-aware policy learning.

2. Comparisons and efficiency. We expanded comparisons with adaptive visual-reading methods and added same-checkpoint resolution controls at comparable input costs. These results support the performance–compression benefits of selective enhancement. A fixed-generation experiment retains a 1.47× end-to-end speedup, including initial rendering. We also clarified baseline training, annotation and data-selection procedures, reward design, and inference settings.

3. Robustness, generalization, and limitations. We added ten repeated LongBench evaluations, experiments on Qwen3.5-2B, and tool-enabled DocVQA/InfoVQA evaluations, alongside analyses of existing English/Chinese and context-length results. We acknowledged the modest LongBench improvement and weaker short-context reasoning, while emphasizing approximately 3× LongBench input compression and the advantages in long-context retrieval and memory.

Manuscript revisions. The revised paper incorporates the external method comparisons (Section 4.3, Table 3), combined localization-supervision and reward ablations (Section 4.4, Table 5), and fixed-generation latency experiment (Appendix C.2, Table 14). Detailed input costs and resolution-dependent reward results are provided in the appendix. All additions and revisions are marked in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.