SEE: Unlocking the Hidden Perception of Frozen VLMs
Organizations: Amazon
Abstract
Vision-language models (VLMs) grasp the overall content of an image but often miss the fine details a question hinges on, such as a small object in a large photo or a number in a dense infographic. Yet even when their answer is wrong, the region they point to often contains the evidence. We introduce Self-Elicited Evidence (SEE), a training-free method that lets a frozen VLM look again where it points. When the initial answer has low confidence, SEE prompts the model to propose regions where the details may appear, answers the question on these enlarged crops, and aggregates the answers with the original prediction through confidence-weighted voting. The model also answers with its own transcript, which helps on text-rich images. SEE needs no detector or OCR engine. Across five backbones, SEE improves every small-object benchmark by 2.3-13.1 points and outperforms five training-free baselines on small objects. With SEE, Qwen3.5-4B surpasses the direct answer of Qwen3.5-9B and matches the 6 times larger Qwen3.5-27B.
Figures & tables
| Method | Needs | New views | Reads | Decides | Time |
|---|---|---|---|---|---|
| Self-cons. [ 22 ] | probs. | chains | – | conf. vote | |
| Zoom-Refine [ 25 ] | – | one box | – | 2nd pass | |
| ZoomEye [ 16 ] | probs. | tree search | – | chosen patches | |
| HiDe [ 10 ] | attention | packed regions | – | one pass | |
| DyFo [ 6 ] | detector | crops | – | reward vote | |
| SEE | probs. | own boxes, text | ✓ | conf. vote |
| Method | Q3.5-0.8B [ 15 ] | Q3.5-2B [ 15 ] | Q3.5-4B [ 15 ] | Q3.5-9B [ 15 ] | GLM-4.1V [ 2 ] | Overall ∘ | Small ∘ |
|---|---|---|---|---|---|---|---|
| Direct | 62.8 | 72.8 | 77.9 | 80.5 | 73.2 | 73.4 | 72.4 |
| Self-consistency [ 22 ] | 68.7 | 75.8 | 83.5 | 84.3 | 75.3 | 77.5 | 75.9 |
| Zoom-Refine [ 25 ] | 64.7 | 66.1 | 80.4 | 82.9 | 76.6 | 74.1 | 74.7 |
| ZoomEye [ 16 ] | 60.8 | 70.3 | 78.4 | 84.2 | 72.5 | 73.2 | 72.8 |
| HiDe [ 10 ] | 66.7 | 73.1 | 76.5 | 88.2 | 77.9 | 76.5 | 76.4 |
| DyFo [ 6 ] | 62.6 | 70.3 | 70.9 | 79.0 | 75.8 | 71.7 | 71.4 |
| Method (recovery / damage) | 0.8B | 2B | 4B | 9B | GLM-4.1V |
|---|---|---|---|---|---|
| Self-consistency | |||||
| Zoom-Refine | |||||
| HiDe | |||||
| DyFo | |||||
| SEE |
| Method | Text | V* | HR-4K | HR-8K | InfoVQA | Mean |
|---|---|---|---|---|---|---|
| vanilla | 78.5 | 77.6 | 73.4 | 82.0 | 77.9 | |
| self-consistency [ 22 ] | 82.2 | 84.9 | 80.0 | 87.0 | 83.5 | |
| HiDe [ 10 ] | 73.3 | 76.0 | 73.1 | 83.7 | 76.5 | |
| SEE | ✓ | 90.1 | 82.6 | 81.4 | 84.2 | 84.6 |
| controls: SEE with other zoomed views | ||||||
| random crops | ✓ | 73.8 | 74.6 | 70.0 | 81.7 | 75.0 |
| Candidates / rule | V* | HR-4K | HR-8K | InfoVQA | Mean | Time |
|---|---|---|---|---|---|---|
| direct answer only | 78.5 | 77.6 | 73.4 | 82.0 | 77.9 | 7.4 |
| + Zoom | 90.1 | 80.4 | 80.8 | 79.8 | 82.8 | 55.2 |
| + Read | 80.6 | 78.5 | 73.0 | 84.2 | 79.1 | 34.6 |
| + both, vote (Eq. 5 ) | 89.0 | 81.4 | 80.9 | 85.3 | 84.1 | 82.4 |
| + gate ( SEE ) | 89.5 | 82.3 | 81.0 | 84.7 | 84.4 | 43.5 |
| Answering mode | Q3.5-0.8B | 2B | 4B | 9B | GLM-4.1V |
|---|---|---|---|---|---|
| Small objects: mean accuracy over V*Bench, HR-Bench 4K and 8K | |||||
| plain | 64.5 | 72.7 | 76.5 | 78.6 | 69.7 |
| plain + SEE | 72.5 | 77.1 | 84.7 | 83.3 | 73.6 |
| thinking | 71.1 | 77.9 | 84.9 | 86.4 | 72.2 |
| thinking + self-consistency | 73.0 | 79.4 | 86.0 | 87.0 | 72.2 |
| thinking + SEE | 73.3 | 80.5 | 87.3 | 88.4 | 72.1 |
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
| All items | Without tuning items | ||||||
|---|---|---|---|---|---|---|---|
| Size | Benchmark | 95% CI | |||||
| 0.8B | V*Bench | 131 | |||||
| 0.8B | HR-Bench 4K | 740 | |||||
| 0.8B | HR-Bench 8K | 740 | |||||
| 0.8B | InfographicVQA | 440 | |||||
| 2B | V*Bench | 131 | |||||
| Run | V*Bench | HR-4K | HR-8K | InfoVQA | Mean |
|---|---|---|---|---|---|
| analysis run | 89.5 ( ) | 82.3 ( ) | 81.0 ( ) | 84.7 ( ) | 84.4 |
| comparison run | 90.1 ( ) | 82.6 ( ) | 81.4 ( ) | 84.2 ( ) | 84.6 |
| seed 2 | 90.6 ( ) | 83.4 ( ) | 81.9 ( ) | 85.6 ( ) | 85.4 |
| seed 3 | 90.6 ( ) | 83.5 ( ) | 81.0 ( ) | 85.4 ( ) | 85.1 |
| mean SD of the gain |
| Baseline | Cells | 95% CI | ||
|---|---|---|---|---|
| Self-consistency | Small ∘ (15 cells) | |||
| InfoVQA (5 cells) | ||||
| Overall ∘ (20 cells) | ||||
| Q3.5-0.8B | ||||
| Q3.5-2B | ||||
| Q3.5-4B |
| Analysis | Result |
|---|---|
| Elicitation gap (V*Bench; medoid of 5 box requests, hit: IoU ) | hits on wrong ( ), right ( ) |
| IoU of that box predicts that a zoom fixes the error | AUROC ( ) |
| Self-check “does this crop show what the question asks?” (pilots) | AUROC – ; answers changed |
| Most confident view instead of the vote (every size) | small objects to ; InfoVQA to |
| The zooms share one vote (weight ) instead of one each | mean vs. ( ) |
| Calibration error of zoomed views vs. direct answer (small objects) | ECE – vs. – |
| Method | Q3.5-0.8B | Q3.5-2B | Q3.5-4B | Q3.5-9B | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| V* | HR-4K | HR-8K | InfoVQA | V* | HR-4K | HR-8K | InfoVQA | V* | HR-4K | HR-8K | InfoVQA | V* | HR-4K | HR-8K | InfoVQA | |
| Direct | 66.0 | 66.1 | 61.3 | 57.6 | 75.9 | 72.5 | 69.8 | 72.8 | 78.5 | 77.6 | 73.4 | 82.0 | 85.9 | 76.7 | 73.3 | 85.9 |
| Self-consistency | 74.3 | 68.6 | 64.5 | 67.2 | 74.9 | 74.9 | 71.9 | 81.3 | 82.2 | 84.9 | 80.0 | 87.0 | 88.5 | 81.1 | 77.8 | 89.7 |
| Zoom-Refine | 74.3 | 69.2 | 65.5 | 49.8 | 71.2 | 68.1 | 65.1 | 60.1 | 81.7 | 80.5 | 78.1 | 81.1 | 83.2 | 81.1 | 80.0 | 87.2 |
| ZoomEye | 66.0 | 69.0 | 59.0 | 49.0 | 72.8 | 67.0 | 72.0 | 69.5 | 80.6 | 80.0 | 70.0 | 82.8 | 93.7 | 77.0 | 79.0 | 86.9 |
| HiDe | 74.3 | 70.8 | 66.1 | 55.6 | 73.8 | 73.2 | 71.2 | 74.2 | 73.3 | 76.0 | 73.1 | 83.7 | 92.7 | 87.4 | 85.8 | 86.9 |
| Benchmark | 0.8B | 2B | 4B | 9B | GLM-9B |
|---|---|---|---|---|---|
| V*Bench | |||||
| HR-Bench 4K | |||||
| HR-Bench 8K | |||||
| InfographicVQA |
| Method | 0.8B | 2B | 4B | 9B | GLM |
|---|---|---|---|---|---|
| Zoom-Refine | |||||
| ZoomEye | |||||
| DyFo | |||||
| HiDe | |||||
| SEE (zoom + read) | |||||
| Self-consistency |
| Backbone | Mode | V* | HR-4K | HR-8K | InfoVQA |
|---|---|---|---|---|---|
| Q3.5-0.8B | thinking | 72.8 | 71.6 | 68.9 | 71.3 |
| + self-consistency | 75.4 | 73.6 | 70.0 | 74.2 | |
| + SEE | 75.4 | 73.6 | 70.8 | 71.9 | |
| Q3.5-2B | thinking | 83.2 | 76.5 | 73.9 | 82.5 |
| + self-consistency | 83.2 | 78.8 | 76.1 | 85.1 | |
| + SEE | 86.4 | 77.9 | 77.1 | 82.5 |
| Benchmark | Direct | SEE | Gain | Gate stops (%) | |
| V*Bench | 89.0 | 92.1 | 0.18 | 46.1 | |
| HR-Bench 4K | 82.0 | 83.8 | 0.09 | 54.1 | |
| HR-Bench 8K | 78.4 | 81.6 | 0.002 | 48.5 | |
| InfographicVQA | 89.6 | 90.4 | 0.33 | 65.6 | |
| Small-object mean | 83.1 | 85.8 | |||
| Overall mean | 84.8 | 87.0 |
| Model | Bench. | Van. | HiDe | SEE | +HiDe | Either | Time |
|---|---|---|---|---|---|---|---|
| Q3.5-0.8B | V* | 66.0 | 74.3 | 79.1 | 78.5 † | 88.5 | 12.7 |
| Q3.5-0.8B | HR-4K | 66.1 | 70.8 | 70.5 | 72.1 | 81.6 | 27.4 |
| Q3.5-0.8B | HR-8K | 61.2 | 66.1 | 67.9 | 69.1 † | 78.2 | 30.2 |
| Q3.5-0.8B | InfoVQA | 57.6 | 55.6 | 58.6 | 61.5 | 70.0 | 11.7 |
| Q3.5-2B | V* | 75.9 | 73.8 | 81.7 | 83.8 † | 91.6 | 10.5 |
| Q3.5-2B | HR-4K | 72.5 | 73.2 | 75.9 | 76.1 † | 86.2 | 33.4 |
VLM Inference
Vision-language model inference studies how multimodal models turn images and video into answers efficiently, accurately, and reliably.
See all 86 papers