{"id":"30997530-a196-49a3-b0b5-499d97324106","arxiv_id":"2608.10628","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"InSight-doc trains an 8B vision-language model to zoom into document sub-regions on demand, improving long-document VQA accuracy by up to 16.4 points while cutting latency by 41-68%.","lead":"InSight-doc is a system that lets a vision-language model read long documents by starting with low-resolution overview pages and selectively zooming into specific regions when it needs more detail. The trained 8B model improves document VQA accuracy while sharply reducing latency and hallucination compared with the base model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The arXiv training source may overlap MMLongBench-Doc evaluation documents, and no deduplication is reported, so the headline long-document gains could be inflated by layout or evidence memorization.","rationale":"The central claim is primarily an empirical accuracy-efficiency gain over Qwen3-VL-8B. The controlled comparisons in Table 2 and Appendix F.2, the released code, model, and data, and the honest discussion of page-cap and cross-paper limitations together provide genuine support for the framework. The reader's teacher-bias concern is legitimate but not the most load-bearing: the RL stage rewards final-answer correctness rather than imitation of the teacher's bounding boxes, so a biased teacher can in principle be corrected by the sparse-accuracy reward. A more direct threat is train/eval contamination. Section 4.1 builds a substantial part of the training corpus from a broad arXiv snapshot selected for long papers with dense figures and tables, which is exactly the population that MMLongBench-Doc samples. The paper never reports a deduplication step between its training documents and any evaluation benchmark. If the same arXiv PDFs appear in both, the model could recognize pages and visual elements from training, inflating the MMLongBench-Doc gain of 17.1 points and the associated hallucination and latency improvements relative to a baseline that has no such exposure. This is a concrete, checkable threat, and the absence of an overlap audit is a real gap in an otherwise carefully controlled evaluation. I therefore recommend keeping the reader's CONDITIONAL verdict, but for a different reason: acceptance should require the overlap check described above, while the teacher-audit concern can remain a secondary item for revision.","tokens_in":35402,"tokens_out":7713,"duration_ms":78986,"concrete_test":"Compute the overlap between the arXiv documents used for training (by arXiv ID, title, or PDF SHA-256) and the official MMLongBench-Doc evaluation documents, and similarly for LongDocURL if its source documents are public. If the overlap is non-zero, rerun InSight-doc-8B (SFT+RL) on the non-overlapping subset at r=0.25, 0.35, and 0.5, and compare accuracy, not-answerable F1, sequence length, and latency against the corresponding rows of Tables 2 and 16. If removing overlapping documents shifts the MMLongBench-Doc gains by more than a few points, the headline long-document improvement must be recomputed and reported separately for non-contaminated documents.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper constructs part of its training corpus from a broad arXiv snapshot, prioritizing longer papers with dense figures and tables (Section 4.1, Appendix C.1). MMLongBench-Doc, one of the four benchmarks behind the headline claim, is built from scientific papers with visualizations and contains documents up to 468 pages (Appendix E.1). No deduplication against MMLongBench-Doc or LongDocURL is reported anywhere in the data-construction, filtering, or evaluation sections. If the same PDFs appear in both the training corpus and the evaluation benchmark, InSight-doc has seen the document pages and their enriched visual descriptions during training, while the Qwen3-VL-8B baseline has not. That would inflate the +17.1 point gain on MMLongBench-Doc in Table 2, as well as the benchmark-specific hallucination and latency numbers in Appendix F, because the model could memorize layout and evidence locations rather than learn a generalizable zoom policy. The paper is careful to flag other threats to validity, such as the 40-page cap (Appendix F.2) and non-controlled cross-paper comparisons (Appendix F.3), but it never reports an overlap audit. This gap is the most load-bearing because MMLongBench-Doc anchors the 'long-document' part of the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InSight-doc, an agentic visual perception framework for long-document VQA in which the model starts from a low-resolution rendering of all pages and iteratively issues zoom_in tool calls (image index, region label, bounding box) to append high-resolution crops to its interleaved reasoning chain. The training corpus is built from six document sources, with an arXiv-derived branch that uses MinerU and Gemini to generate enriched visual descriptions and QA pairs, and an InSight-o3 two-agent teacher that produces zoom-in trajectories; this yields 17,913 SFT trajectories and 19,236 RL prompts. Starting from Qwen3-VL-8B-Instruct, the authors train with SFT then GRPO. On DUDE, MP-DocVQA, MMLongBench-Doc, and LongDocURL, the final model improves the baseline by 4.3-16.4 accuracy points, reduces hallucination on unanswerable questions by more than 40%, and reduces latency by 41%-68%. Additional experiments cover general high-resolution VQA (MME-RealWorld-Lite, O3-Bench), cross-paper comparisons, trajectory-quality metrics, an uncapped long-document setting, and an inference-cost analysis in Propositions 1 and 2.","tokens_in":35559,"tokens_out":14709,"duration_ms":134993,"significance":"If the results hold, this is a valuable empirical contribution: it demonstrates that a retriever-free, region-level zoom agent trained with SFT+RL can push the accuracy-efficiency Pareto frontier for long-document VQA. The controlled comparisons against Qwen3-VL-8B with and without the same zoom tool, the disclosure of the 40-page cap with uncapped results in Appendix F.2, the explicit caveats on cross-paper comparisons in Appendix F.3, the 150-example judge calibration, and the release of code, data, and model are all strengths. The main reservation is that the arXiv-derived training corpus may overlap with MMLongBench-Doc, and no deduplication is reported; since MMLongBench-Doc anchors the long-document part of the central claim, this issue must be resolved. A secondary concern is that the teacher trajectories are audited only on training data, leaving the localization-bias premise unverified on evaluation documents.","major_comments":[{"comment":"The training corpus uses an arXiv Kaggle snapshot that prioritizes longer papers with dense figures and tables (Section 4.1, Appendix C.1). MMLongBench-Doc is built from scientific papers with visualizations and contains documents up to 468 pages (Appendix E.1). The paper does not report any overlap or deduplication analysis between the arXiv training snapshot and MMLongBench-Doc or LongDocURL. Because Table 2's headline long-document gain is +17.1 points on MMLongBench-Doc at r=0.25, and the uncapped setting in Table 16 still shows a +17.0-point gap, partial overlap would inflate the central long-document claim: InSight-doc would have seen these document pages and their enriched descriptions during training, while the Qwen3-VL-8B baseline has not. Please report the overlap count (e.g., arXiv IDs or PDF hashes) against both long-document benchmarks and re-run the headline metrics on the non-overlapping subset (or exclude overlapping documents). This is necessary to support the long-document conclusions.","section":"Section 4.1 / Appendix C.1 / Appendix E.1 / Table 2"},{"comment":"The SFT trajectories are distilled from InSight-o3, in which a fine-tuned Qwen3-VL-8B vSearcher supplies the bounding box for every zoom-in call (Section 4.2, Appendix C.4). The trajectory-quality audit in Appendix D.2 is performed on training rows only, not on evaluation documents, so the transfer of the learned zoom policy depends on an unverified assumption that the teacher's localization is accurate and unbiased outside the training distribution. Table 6 reports evidence-box coverage for the final InSight-doc model, not for the teacher. Please add a small human audit of teacher bounding boxes on held-out evaluation-style documents, or report teacher-level coverage on a sample of LongDocURL evidence boxes, to show that the student is not inheriting systematic localization bias from the teacher.","section":"Section 4.2 / Appendix C.4 / Appendix D.2"}],"minor_comments":[{"comment":"The main text describes a universal three-stage cascade, but Appendix C.4 explains that arXiv-derived answerable rows bypass the Stage-1 prior-only filtering. This exception should be stated in Section 4.2.","section":"Section 4.2 vs. Appendix C.4"},{"comment":"The InSight-doc scores quoted in Table 5 for MMLongBench-Doc and LongDocURL (57.8 and 65.6 at r=0.5) correspond to the uncapped setting reported in Table 16, but Table 5's caption and the main-text discussion do not say so; add a footnote to prevent readers from comparing them with the capped numbers in Table 2.","section":"Table 5 / Section 5.2 / Appendix F.3"},{"comment":"The phrase 'consistent with a theoretical prediction of about 48%–81% (Proposition 2)' overstates the status of the cost model: Propositions 1 and 2 are upper bounds under hand-chosen parameters (δ, λ, γ, κ), not predictions fitted to or tested against the latency measurements. Suggest rewording to 'consistent with the upper-bound analysis under the stated parameter regime'.","section":"Section 3.2 / Section 5.3"},{"comment":"The Limitations section discusses only the choice of base model and RL method; it does not mention the two main threats to the long-document results, namely the potential training/evaluation overlap and the teacher-dependence of the SFT trajectories. A sentence or two acknowledging these would be appropriate.","section":"Limitations"},{"comment":"The first sentence of the Conclusion contains a typo: 'reducing, hallucination' should be 'reducing hallucination'.","section":"Section 6 / Conclusion"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the training/evaluation overlap with MMLongBench-Doc is the key uncertainty in this paper. The revision should prioritize the overlap audit; if the authors cannot exclude contamination, the headline long-document gains should be re-reported on a non-overlapping subset. The teacher-dependence of the SFT data (own prior work InSight-o3) is a second point to watch, but it is less severe because the paper audits training trajectories and reports final model coverage. I do not see any basis for rejection if the overlap can be ruled out."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a solid system paper. It trains an 8B MLLM to selectively zoom into sub-page regions of long documents without an external retriever, and the controlled comparisons against the same base model with and without the zoom tool are consistent and fairly presented. The gains are real on benchmarks that don't depend on the arXiv overlap question. But the paper never reports a dedup check between its arXiv training snapshot and MMLongBench-Doc or LongDocURL, and both are built from scientific papers. That's the thing I'd chase first.\n\nWhat's actually new: region-level adaptive zooming inside one MLLM, trained end-to-end with SFT then GRPO, no external retriever. Prior coarse-to-fine work either works at page level or needs a retriever. The paper also constructs a large training corpus and releases code, data, and model. Credit where due: the ablation of SFT vs RL, the trajectory quality statistics, the uncapped results in Appendix F.2, and the explicit caveats about cross-paper numbers are all honest. The latency analysis is presented as an upper bound, not a fit, and the measured numbers are reported separately; that's the right call.\n\nSoft spots: first, the overlap. Section 4.1 and Appendix C.1 say they sampled arXiv papers, prioritizing long ones with dense figures. MMLongBench-Doc is exactly that population. No dedup is reported. If the same PDFs appear in training, the +17.1 point gain on MMLongBench-Doc could be inflated, and the baseline hasn't seen them. That doesn't kill the paper -- the gains on DUDE and MP-DocVQA, which are not arXiv, are also large -- but it does mean the headline long-document numbers need an audit before I'd trust them. Second, the SFT teacher is InSight-o3, a two-agent system with a GPT-5-mini planner and a Qwen3-VL-8B searcher. The quality audit covers training data only, not evaluation benchmarks, so systematic teacher bias could transfer. The RL stage helps, but it remains a dependency. Third, the LLM judge is calibrated on 150 examples; that's thin, though they do report the false-positive/false-negative tradeoff and choose a conservative option.\n\nNet: this paper is for anyone working on document VQA or agentic visual perception. It deserves a serious referee. I'd send it out with a request for the dedup audit and a stronger teacher-bias check.","headline":"Real gains on document VQA from region-level adaptive zooming, but the missing dedup audit between the arXiv training snapshot and MMLongBench-Doc is the thing to chase first.","tokens_in":36253,"tokens_out":3329,"would_cite":true,"duration_ms":30042,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A model that reads documents by zooming in beats fixed high-resolution reading.","keywords":["agentic visual perception","long-document understanding","document visual question answering","adaptive resolution","zoom-in policy","reinforcement learning","hallucination reduction","multimodal large language models"],"falsifier":"Retrain InSight-doc on the same SFT and RL data but with every teacher bounding box randomly shifted by roughly ten percent of page width, then evaluate on LongDocURL, which has box-level evidence annotations. If accuracy and evidence-box coverage stay close to the reported numbers, the model is not relying on precise teacher localization; if they collapse toward the baseline, the adaptive-perception claim depends on memorized teacher boxes.","tokens_in":35089,"feed_emoji":"🔍","tokens_out":6034,"duration_ms":51263,"temperature":0.7,"pith_summary":"InSight-doc is an agentic visual perception method that lets a multimodal language model read long documents at low resolution first, then zoom into specific page regions when it needs finer evidence. The paper claims this turns visual resolution into an adaptive reasoning-time resource, allowing the model to match or beat fixed high-resolution processing while using many fewer tokens and less time. On document VQA benchmarks, the trained model improves accuracy by 4.3 to 16.4 points over its base model, cuts hallucination on unanswerable questions by more than 40 percent, and reduces inference latency by 41 to 68 percent. The core bet is that a model can learn where and when to zoom well enough to replace both external retrievers and full-page high-resolution encoding.","feed_headline":"Zoom agent reads long documents faster and more accurately","feed_subtitle":"A retriever-free model learns when to zoom in, beating fixed-resolution reading with fewer tokens and less hallucination.","key_machinery":"The load-bearing mechanism is the zoom-in tool call, formalized as zoom_in(k, d, b | context), where k indexes the page image, d is a natural-language description of the region of interest, and b is a bounding box; the crop is taken from the high-resolution source and appended to the visual context. The learned policy decides when to call it, which region to request, and how to integrate the returned crop. Training uses supervised fine-tuning on 17.9K trajectories generated by a two-agent teacher, followed by GRPO reinforcement learning on 19.2K hard examples with binary accuracy reward. A latency bound shows that aggressive downsampling plus a few zoom calls keeps total inference below the no-resize baseline.","core_discovery":"The central claim is that one end-to-end multimodal model can learn to acquire the right visual evidence at the right resolution through multi-round region cropping. Starting from a low-resolution overview of the whole document, the model emits a thought and a zoom-in tool call, receives a higher-resolution crop of a predicted bounding box, and repeats until it answers or abstains. The authors show this policy beats both the same backbone equipped with a zoom tool but no specialized training and the no-tool high-resolution baseline, while shortening sequence length and latency. It also abstains more accurately on unanswerable questions, which they attribute to the model actively checking regions, finding no evidence, and refusing to fabricate an answer.","pith_inferences":["The same zoom-as-resource idea could be applied to long videos, replacing spatial region cropping with temporal segment selection; the paper draws the analogy but does not test it.","If the teacher's localization quality is the bottleneck, then improving or scaling the teacher should directly improve the student, pointing to a distillation-style research program.","The token and latency savings could make long-document understanding feasible on smaller or edge-deployed models that cannot afford full-page high-resolution encoding.","A fully controlled head-to-head against retrieval-assisted page-level methods would settle whether region-level zooming is inherently more token-efficient; the paper's proxy comparison suggests so but is not conclusive."],"forward_implications":["Documents can be processed at very low initial resolution without sacrificing accuracy, because the model recovers detail on demand.","Unanswerable questions become safer: the model more often states that evidence is insufficient instead of hallucinating an answer.","The learned zoom policy transfers from documents to general high-resolution image VQA, indicating a general visual-search ability.","Reinforcement learning sharpens the policy beyond supervised fine-tuning, raising evidence-box coverage and nearly eliminating stuck trajectories.","Because the method needs no external retriever, it avoids retrieval indexing overhead and retriever errors."],"supporting_citations":[{"why":"Supplies the two-agent teacher that generates the zoom-in trajectories used as SFT targets.","marker":"Li et al., 2026"},{"why":"The base model and the teacher's searcher; all accuracy and efficiency gains are measured against it.","marker":"Bai et al., 2025a"},{"why":"Provides DUDE, a key document VQA benchmark with unanswerable questions used in hallucination evaluation.","marker":"Van Landeghem et al., 2023"},{"why":"Provides MMLongBench-Doc, the long-document benchmark for accuracy, hallucination, and latency comparisons.","marker":"Ma et al., 2024b"},{"why":"Provides LongDocURL, whose box-level evidence annotations are used to measure trajectory quality.","marker":"Deng et al., 2025"},{"why":"Provides the GRPO reinforcement-learning algorithm used to train the final model.","marker":"Shao et al., 2024"},{"why":"Closest coarse-to-fine method; used as the comparison point for retriever-free region-level evidence acquisition.","marker":"Zheng et al., 2026"}],"fun_headline_variants":["Adaptive zoom policy reads long docs with fewer tokens","InSight-doc: AI agent that zooms to reduce context rot","Zoom-in on evidence: better doc understanding agent","Retriever-free agent learns when to zoom for answers","Active perception: model zooms to cut hallucination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The teacher that produced the zoom-in training data must localize the right regions correctly enough; if its bounding boxes are systematically wrong or biased, the student inherits that bias and the claimed gains may not generalize beyond the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive zoom policy reads long docs with fewer tokens","InSight-doc: AI agent that zooms to reduce context rot","Zoom-in on evidence: better doc understanding agent","Retriever-free agent learns when to zoom for answers","Active perception: model zooms to cut hallucination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001084,"raw_usage":{"total_tokens":4493,"prompt_tokens":864,"completion_tokens":3629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":3548}},"tokens_in":480,"tokens_out":3629,"duration_ms":25681,"temperature":1.0,"reasoning_tokens":3548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:37:29.589293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain InSight-doc on the same SFT and RL data but with every teacher bounding box randomly shifted by roughly ten percent of page width, then evaluate on LongDocURL, which has box-level evidence annotations. If accuracy and evidence-box coverage stay close to the reported numbers, the model is not relying on precise teacher localization; if they collapse toward the baseline, the adaptive-perception claim depends on memorized teacher boxes.","supporting_citations":[],"review_version":1}