DR Pixel: Page-Native Deep Research with Vision-Language Models
Organizations: NYU Shanghai · University of California, Berkeley · Rational Dynamics · Toyota Technological Institute at Chicago ·
Abstract
Complex documents such as scientific papers contain rich information in tables, figures, and equations; however, existing deep research (DR) agents typically reduce papers to OCR-extracted text. We instead ask whether vision-language models (VLMs) can enable page-native deep research: directly retrieving and reasoning over rendered document page images. We introduce DR Pixel, the first open-source end-to-end multimodal DR agent that uses rendered page images as its sole retrieval and reasoning unit, operating over a corpus of 1.06M pages. To evaluate this setting, we construct DRPixelBench, a benchmark on scientific papers requiring reasoning over text, figures, tables, and equations. Surprisingly, page-native DR already works well without training, outperforming OCR-based baselines across state-of-the-art models like GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, while requiring fewer tool calls. We next turn to DR training. A key challenge is that teacher trajectories for agentic supervised finetuning are noisy, mixing critical reasoning and action steps with redundant or failed attempts. To address this, we introduce counterfactual credit reweighting, a fine-grained credit assignment method that weights individual reasoning and action steps by their counterfactual contribution to the final answer. Our training improves retrieval recall by about 15 points and final accuracy by over 10 points on DRPixelBench, and likewise outperforms the untrained baseline on four existing benchmarks. On out-of-distribution ViDoRe-V3, DR Pixel comes within 1.8 points of an oracle that is given the gold pages directly. Our results suggest that the conventional OCR-then-retrieve pipeline is not only unnecessary but suboptimal: directly retrieving and reasoning over visual document pages makes complex DR both more effective and more efficient.
Figures & tables
| Benchmark | Agentic | MM | Non-templ. | Verif. |
|---|---|---|---|---|
| Deep research | ||||
| BrowseComp-Plus | ✓ | ✗ | ✗ | ✓ |
| DeepResearch Bench | ✓ | ✗ | ✓ | ✗ |
| ScholarQABench | ✓ | ✗ | ✓ | ✗ |
| Multimodal QA | ||||
| AirQA (comp.) | ✓ | ✓ | ✗ | ✓ |
| Retrieval Setup | GPT-5.4 | Gemini-3.5-Flash | Qwen3.6-27B | |||
|---|---|---|---|---|---|---|
| Acc ( ) | # Calls ( ) | Acc ( ) | # Calls ( ) | Acc ( ) | # Calls ( ) | |
| OCR-based pipelines | ||||||
| Text | 43.5 | 12.4 | 40.0 | 17.5 | 33.2 | 11.0 |
| + Figure ( txt + fig ) | 46.3 | 10.8 + 3.0 | 41.3 | 16.7 + 1.0 | 34.9 | 10.9 + 0.6 |
| DR Pixel | 51.9 | 9.2 | 46.4 | 13.7 | 40.1 | 8.8 |
| Model | Tool/q | PgHit | Acc Pg | Acc |
|---|---|---|---|---|
| Open-weight (Qwen3.6-27B) | ||||
| DR Pixel | 8.8 | 36.3 | 63.9 | 40.1 |
| + vanilla SFT | 14.0 | 53.6 | 67.2 | 49.0 |
| DR Pixel | 8.9 | 52.0 | 67.7 | 52.1 |
| Proprietary | ||||
| GPT-5.4 | 9.2 | 45.7 | 71.2 | 51.9 |
| IRPAPERS | M3SciQA | AirQA | ||||||||||
| Model | T/q | DocHit | Acc Doc | Acc | T/q | DocHit | Acc Doc | Acc | T/q | DocHit | Acc Doc | Acc |
| Open-weight (Qwen3.6-27B) | ||||||||||||
| DR Pixel | 5.6 | 89.4 | 68.8 | 65.4 | 7.8 | 75.3 | 63.4 | 49.8 | 7.9 | 84.3 | 46.0 | 42.7 |
| DR Pixel | 5.0 | 95.6 | 82.0 | 79.4 | 9.4 | 90.3 | 69.6 | 64.2 | 7.0 | 89.7 | 55.6 | 52.3 |
| Proprietary | ||||||||||||
| GPT-5.4 | 5.4 | 92.2 | 80.1 | 77.8 | 7.7 | 90.3 | 73.3 | 67.5 | 6.2 | 87.9 | 59.5 | 56.6 |
| English | French | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Retriever | Mod. | Generator | C.S. | Fin. | Phar. | H.R. | Ind. | Phys. | Ener. | Fin. | Global |
| Oracle ∗ | Text | Gemini 3 Pro | 80.9 | 70.2 | 71.4 | 72.3 | 66.4 | 71.2 | 69.2 | 62.8 | 70.6 |
| Page | 86.5 | 70.6 | 76.1 | 71.1 | 68.2 | 74.5 | 69.8 | 64.1 | 72.6 | ||
| Qwen3-Emb-8B | Text | DR Pixel | 86.5 | 64.7 | 60.7 | 63.8 | 53.7 | 77.4 | 66.6 | 51.9 | 65.7 |
| DR Pixel | 88.8 | 71.2 | 67.3 | 65.4 | 66.1 | 77.5 | 69.5 | 52.8 | 69.8 | ||
| Qwen3-VL-Emb-8B | Page | DR Pixel | 87.9 | 64.4 | 67.9 | 64.2 | 59.7 | 76.2 | 62.7 | 50.9 | 66.7 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Property | Text | Table | Figure | Equation | All |
|---|---|---|---|---|---|
| Questions | |||||
| Query Num | 216 | 135 | 102 | 47 | 500 |
| Query Len. | 35.0 | 34.4 | 32.7 | 37.3 | 34.6 |
| Answer Len. | 39.8 | 50.7 | 44.4 | 37.5 | 43.5 |
| Gold Page | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| Corpus | |||||
| Filter | Criterion | Remaining |
|---|---|---|
| Drafted questions | ||
| Coined-name removal | The paper’s own product or benchmark name is replaced by a generic class; questions where it cannot be removed are discarded. | |
| Self-contained reference | No dangling pointers such as “this method” or “the figure above”. | |
| No internal codes | No enumeration of identifiers meaningful only inside the paper. | |
| Single focus | Exactly one answer is required; choosing among listed options counts as one such as “How does it do on ImageNet, and on COCO?”. | |
| Answer shape | The gold matches its declared type: a comparison names a winner, a trend a direction, a regime a condition, and none is a disguised yes/no. | |
| Papers | Pages | |||
|---|---|---|---|---|
| Venue | Count | Share | Count | Share |
| NeurIPS | 21,448 | 21.7% | 232,136 | 21.9% |
| ACL | 12,870 | 13.0% | 114,720 | 10.8% |
| CVPR | 10,312 | 10.4% | 96,869 | 9.1% |
| ICML | 8,654 | 8.7% | 92,765 | 8.7% |
| ICLR | 8,561 | 8.6% | 86,688 | 8.2% |
| Outcome | Count | Share |
|---|---|---|
| Accepted as generated | 107 | 10.3% |
| Kept after editing | 393 | 37.9% |
| Entering DRPixelBench | 500 | 48.3% |
| Rejected | 536 | 51.7% |
| Reviewed | 1,036 | 100.0% |
| Target | R@1 | R@5 | R@10 | R@30 | R@100 | MRR |
|---|---|---|---|---|---|---|
| Page | 13.8 | 26.8 | 32.8 | 41.0 | 48.8 | 0.200 |
| Setting | Retrieved unit | Words per search | |
|---|---|---|---|
| Text | -word chunk | 5 | 1,257 |
| Page-native | page image | 3 | 1,359 |
| GPT-5.4 | DR Pixel | DR Pixel | |||
| Score | Score | Score | |||
| Retrieval | |||||
| Page Hit | 45.7 | 36.3 | 52.0 | ||
| Repeated Queries | 0.0 | 5.3 | 0.1 | ||
| Reading | |||||
| Acc Page Hit | 71.2 | 63.9 | 67.7 | ||
| Model | Tool/q | PgHit | Acc Pg | Acc |
|---|---|---|---|---|
| Open-weight (Qwen3.6-27B) | ||||
| DR Pixel | 8.8 | 36.3 | 63.9 | 40.1 |
| + vanilla SFT | 14.0 | 53.6 | 67.2 | 49.0 |
| DR Pixel | 8.9 | 52.0 | 67.7 | 52.1 |
| Ablation of the credit | ||||
| + answer weighting | 11.6 | 54.9 | 67.2 | 48.9 |
| Model | Tool/q | PgHit | Acc Pg | Acc |
|---|---|---|---|---|
| Reference | ||||
| Global rank | 11.0 | 51.9 | 68.7 | 50.9 |
| Weight magnitude | ||||
| Mean-weight matching | 10.1 | 50.4 | 63.1 | 46.8 |
| Rank support | ||||
| Within-trajectory rank | 10.0 | 50.3 | 70.9 | 50.1 |
Visual Document Retrieval
Visual document retrieval finds relevant pages or regions in image-based documents using visual and multimodal representations.
See all 69 papers