- 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
-
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.
-
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
-
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.
-
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.
-
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.
-
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
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.
| Configuration | REL-SFT | GRPO | Inference tools | LongBench | MRCR (2 needles) | MRCR (4 needles) | MRCR (8 needles) |
|---|---|---|---|---|---|---|---|
| Full FocusVTC | ✅ | ✅ | ✅ | 56.40 | 60.76 | 45.21 | 30.71 |
| w/o tools | ✅ | ✅ | ❌ | 36.82 | 52.00 | 37.65 | 23.50 |
| w/o SFT | ❌ | ✅ | ✅ | 49.30 | 57.05 | 39.36 | 26.66 |
| w/o GRPO, without tools | ✅ | ❌ | ❌ | 37.86 | 44.94 | 28.83 | 21.30 |
| w/o GRPO, with tools | ✅ | ❌ | ✅ | 25.41 | 22.25 | 15.09 | 8.96 |
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.
| Configuration | Inference tools | RULER v1 | RULER v2 | LongBench |
|---|---|---|---|---|
| Full REL-SFT | ❌ | 32.28 | 43.58 | 37.86 |
| T1: SFT without bbox | ❌ | 31.65 | 41.01 | 33.40 |
| Full REL-SFT + GRPO | ✅ | 87.38 | 75.94 | 56.40 |
| T1: SFT without bbox + GRPO | ✅ | 78.58 | 67.08 | 47.86 |
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.
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 (Table 9). At 144 DPI, 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.
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.
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.
Compression is the arithmetic mean of Text/(P+O) across the three benchmarks, using the text-token references in Table 10. Here, P denotes initial prompt tokens and O 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.