RSieve: Suppressing Unconditional Information from Attention to Answers for High-Resolution VQA
Organizations: School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai 201210, China · Jiangnan University, Wuxi, China
Abstract
High-resolution visual question answering (VQA) requires models to identify question-relevant small objects in high-resolution images and select the correct answer based on what they recognize, demanding fine-grained perception. Previous methods focus on obtaining better visual evidence (image crops), but obtaining it alone is insufficient, using it effectively is equally crucial. Visual evidence should therefore be utilized as effectively as possible throughout inference. We argue that small-target evidence is obscured by unconditional information that remains observable in subsequent processing even after the condition on which it should depend is removed. Such information hinders the effective use of small-target evidence. We propose RSieve (Residual-Region-Response Sieve), a training-free framework. We divide inference into three stages---localization, cropping, and answering---and introduce distinct counterfactual references to suppress interference at each stage. During localization, the Shared-Subspace Canceller (SSC) uses a generic image description to construct a counterfactual reference and obtain question-relevant attention maps. During cropping, the Unique-Contribution Spotter (UCS) constructs counterfactual references from visual and non-visual features to identify distinctive, complementary crop regions. During answering, the Content-Aligned Voter (CAV) constructs an image-free branch to calibrate answer responses. Experiments on three high-resolution VQA benchmarks show that RSieve achieves the highest overall accuracy among compared methods across diverse settings, improving base models by up to 21.5 percentage points.
Figures & tables
| Backbone | Method | TF | V ∗ Bench | HRBench-4K | HRBench-8K | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Attr. | Spa. | All | FSP | FCP | All | FSP | FCP | All | |||
| Qwen3-VL-4B | w/o reasoning | 85.2 | 76.3 | 81.7 | 87.3 | 64.3 | 75.8 | 86.3 | 62.0 | 74.1 | |
| BVS [ 11 ] (ICML’26) | 97.4 | 82.9 | 91.6 | 92.0 | 71.5 | 81.8 | 93.3 | 63.5 | 78.4 | ||
| SPARC † [ 2 ] (ICML’26) | – | – | 87.0 | – | – | 80.4 | – | – | 78.6 | ||
| ZwZ [ 31 ] (ICML’26) | 94.8 | 89.5 | 92.7 | 89.0 | 71.8 | 80.4 | 86.8 | 70.8 | 78.8 | ||
| R 3 Sieve (ours) | 98.3 | 90.8 | 95.3 | 94.0 | 74.0 | 84.0 | 94.0 | 72.0 | 83.0 | ||
| Components | Attr. | Spa. | All |
|---|---|---|---|
| R 3 Sieve | 98.26 | 90.79 | 95.29 |
| w/o SSC | 97.39 | 88.16 | 93.72 |
| w/o UCS | 98.26 | 86.84 | 93.72 |
| w/o CAV | 98.26 | 85.53 | 93.19 |
| w/o (SSC + UCS) | 97.39 | 85.53 | 92.67 |
| w/o (SSC + CAV) | 97.39 | 85.53 | 92.67 |
| Method | V ∗ Bench | HRBench-4K | HRBench-8K |
|---|---|---|---|
| HiDe [ 17 ] | 91.62 | 79.38 | 78.38 |
| HiDe + CAV | 92.67 | 82.50 | 82.00 |
| BVS [ 11 ] | 91.62 | 81.75 | 78.38 |
| BVS + CAV | 92.15 | 81.50 | 82.00 |
| R 3 Sieve | 95.29 | 84.00 | 83.00 |
| Method | V ∗ Bench | HRBench-4K | HRBench-8K |
|---|---|---|---|
| Base | 0.67 (0.25) | 3.88 (1.21) | 5.41 (0.22) |
| BVS [ 11 ] | 18.74 (5.47) | 114.11 (28.89) | 94.83 (29.98) |
| HiDe [ 17 ] | 3.74 (0.81) | 12.77 (2.21) | 16.82 (3.28) |
| R 3 Sieve | 8.20 (1.11) | 10.33 (0.94) | 12.07 (0.42) |
| MMVP | MMStar | MME-RW-Lite | ||||||
|---|---|---|---|---|---|---|---|---|
| Method | Single | Pair | Coarse | Fine | All | Per. | Rea. | All |
| Base | 79.00 | 59.33 | 72.00 | 58.40 | 61.60 | 52.52 | 41.73 | 48.31 |
| BVS [ 11 ] | 77.67 | 57.33 | 74.00 | 55.60 | 62.53 | 57.66 | 44.53 | 52.53 |
| HiDe [ 17 ] | 79.33 | 60.00 | 70.00 | 56.40 | 60.27 | 55.60 | 40.80 | 49.82 |
| R 3 Sieve | 79.33 | 60.67 | 70.00 | 66.00 | 63.40 | 60.14 | 45.07 | 54.25 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Errors | ||
|---|---|---|---|
| All 191 questions | |||
| Qwen3-VL-4B | 35 | 28/35 (80.0%) | 27/35 (77.1%) |
| Qwen3-VL-8B | 33 | 31/33 (93.9%) | 30/33 (90.9%) |
| InternVL3.5-4B | 52 | 44/52 (84.6%) | 43/52 (82.7%) |
| InternVL3.5-8B | 51 | 41/51 (80.4%) | 40/51 (78.4%) |
| 54 spatial questions with both queried targets annotated | |||
| Crop input | Original order | Pooled |
|---|---|---|
| Automatic | 84.82 | 86.91 |
| Annotated target | 88.48 | 91.62 |
| Target detail views | Original | Blurred |
|---|---|---|
| Absent | 86.96 | 93.91 |
| Fixed | 97.39 | 98.26 |
| Full-image caption |
| Full source image. |
| Crop caption (repeated for each crop) |
| A detail view of the same image, source rectangle [l, t, r, b] in [left, top, right, bottom] coordinates on a 0-1000 grid. Use the full image to determine global spatial relationships. |
| Shared answering instruction |
| Answer with only the option letter. |
| Branch | Images | Captions | Question + options |
|---|---|---|---|
| Visual | Yes | Yes | Yes |
| Metadata-only | No | Yes | Yes |
| Question-only | No | No | Yes |
| Module | Hyperparameter | Qwen3-VL-4B | Qwen3-VL-8B | InternVL3.5-4B | InternVL3.5-8B |
|---|---|---|---|---|---|
| SSC | Response layer | 18 | 20 | 18 | 18 |
| Preview tokens | 4096 | 8192 | 4096 | 8192 | |
| UCS | Nonvisual weight | 1 | 2 | 1.5 | 3 |
| Detail-gain exponent | 1 | 1 | 0.375 | 0.325 | |
| CAV | Visual pooling exponent | 0.375 | 3 | 24 | 8 |
| Text pooling exponent | 0.3 | 3 | 0.3 |
| Location score | Independent ranking | GCS |
|---|---|---|
| Residual energy | 92.67 | 93.72 |
| Leverage score | 90.05 | 95.29 |
| ID | Variant | Attr. | Spa. | All | ||||
|---|---|---|---|---|---|---|---|---|
| D0 | Full CAV | 98.26 | 90.79 | 95.29 | ||||
| D1 | Without metadata correction | 98.26 | 86.84 | 93.72 | ||||
| D2 | Probability mean with both references | 98.26 | 88.16 | 94.24 | ||||
| D3 | Probability mean with question reference | 99.13 | 85.53 | 93.72 | ||||
| D4 | Unit branch temperatures | 98.26 | 86.84 | 93.72 | ||||
| D5 | Pooling without calibration | 99.13 | 85.53 | 93.72 |
| Case | Visual margin | Calib. advantage | |
|---|---|---|---|
| 4K | 8K | ||
| Camera color | 0.800 | 10.195 | 0.974 |
| Zaragoza sign direction | 0.198 | 2.490 | 1.019 |
| Circled label meaning | 2.073 | 4.976 | 2.816 |
Visual Question Answering
Visual Question Answering (VQA) develops systems that answer natural-language questions about images, videos, and visual documents.
See all 275 papers