InSight-doc: Agentic Visual Perception for Long-Document Understanding
Kaican Li, Weiyan Xie, Lewei Yao, Jiannan Wu, Lanqing Hong, Yongxiang Huang, Nevin L. Zhang
cs.CV, cs.CL, cs.LG
2026-08-11
InSight-doc trains Qwen3-VL-8B to start low-res and zoom in on demand; it beats GPT-5.4-mini on long-document VQA while cutting latency 41–68% and hallucination over 40%.
Long-document visual question answering hits two recurring snags. Feeding every page at high resolution makes the token count blow up roughly with the square of the input, so inference is slow and expensive. And as the prompt grows, models suffer "context rot," degrading even though more information is in scope. The usual fixes are either end-to-end high-resolution input (costly) or bolting on an external retriever to pre-filter pages (which needs indexing and can miss the page that matters).
The authors want to mimic how a person reads a long document: skim the whole thing first, zoom in only where detail is needed. The hard part is teaching the model when to zoom and where, with no external retriever.
InSight-doc starts from Qwen3-VL-8B-Instruct and trains it into an agent that calls a zoomin tool. At inference it first gets every page at low resolution (50–100 DPI), then can repeatedly request a zoom: it specifies which image, a natural-language description of the region of interest, and a bounding box, and the cropped high-resolution patch is spliced back into the reasoning chain as evidence. Zoomed regions can be zoomed again, with the scale factor accumulating as r ← c·r.
This behavior is trained, not innate, in two stages:
The biggest gains come when the model starts at low resolution (r=0.25, 50 DPI):
| Benchmark | Qwen3-VL-8B | InSight-doc-8B | Gain |
| DUDE | 52.9 | 70.1 | +17.2 |
| MP-DocVQA | 65.1 | 83.4 | +18.3 |
| MMLongBench-Doc | 33.7 | 50.8 | +17.1 |
| LongDocURL | 50.5 | 63.3 | +12.8 |
| Average | 50.5 | 66.9 | +16.4 |
At medium resolution (r=0.5) the gain narrows to +4.3 on average, because high resolution is already near the ceiling. At r=0.25 this 8B model averages 66.9%, beating GPT-5.4-nano (51.2%), GPT-5.4-mini (64.3%), and GPT-5-mini (65.9%); at r=0.5 it roughly ties them.
The efficiency story is the bigger payoff. At 50 DPI it reaches 66.9% accuracy using less than half the tokens of the 100-DPI baseline's 68.3% (a 58% cut). On the longest-document subset, 70 DPI reaches 56.2% versus the baseline's 53.2%, with tokens down from 136.8k to 42.4k, a 69% cut. Latency follows: 41–68% lower on the long-document benchmarks, and on the longest subset per-example time drops from 39.3s to 11.2s (71% faster) at higher accuracy.
On unanswerable questions, measured by F1 (whether the model will admit it cannot answer), DUDE rises from 44.5 to 69.1 and MMLongBench-Doc from 48.5 to 74.4, a hallucination reduction of more than 40%. Against dedicated methods, InSight-doc is the only one that is simultaneously retriever-free, coarse-to-fine, iterative, and region-level. It scores 57.8 on MMLongBench-Doc and 65.6 on LongDocURL, ahead of the previous best Doc-V by 15.7 and 9.3 points respectively.
Long-document multimodal understanding is a real engineering need (papers, financial reports, contracts), and feeding all pages at high resolution is the common, expensive default. InSight-doc offers a clean alternative: an open 8B agent that, without losing to and sometimes beating closed models, drives inference cost down to a fraction. Code, data, and model are all released. The RL stage cuts "stuck" trajectories from 9.7% to 0.1% and lifts evidence-box coverage from 27.5% to 82.3%, which suggests the model genuinely learned where to zoom rather than clicking at random.
The authors acknowledge that only Qwen3-VL-8B-Instruct was validated as the backbone, and they did not try more advanced RL algorithms or reward designs, so how the recipe transfers to other models and larger scales is unmeasured. The related-work comparison table is explicitly not controlled: backbones, training data, input resolution, page budgets, and evaluation protocols differ across methods, so those cross-paper numbers are qualitative only. The main evaluation uses a "capped" setting that affects only MMLongBench-Doc and LongDocURL, but capping does change absolute scores.