{"id":"8d3d3100-e8b0-4e68-92cf-357df3e5d185","arxiv_id":"2607.24554","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Decoupling semantic anchoring from high-res cropping fixes VLM visual attention sinks in document Graph RAG, gaining up to 12.5 SPR points and 40.8% fewer prompt tokens.","lead":"DeCoRAG rebuilds multimodal Graph RAG so VLMs first write a global semantic summary, then crop only dense regions before extracting knowledge graphs. That split cuts graph-construction tokens by about 41% and raises answer pass rates on sparse charts, slides, and tables.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline +12.5-point SPR gain rests on a custom judge metric that was never calibrated on PaperTab — the very benchmark where the 12.5 figure comes from — and whose rubric may structurally favor DeCoRAG's denser outputs.","rationale":"I read the paper as a mechanism-motivated systems paper whose empirical core is: (i) a boundary-driven attention-sink diagnosis supported by genuinely decent controlled interventions (Table 1's paired design with sign test is above the norm for this genre), and (ii) end-to-end SPR gains plus a 40.8% token reduction. The token claim is internally well-documented (Table 8: request counts flat, output tokens flat, so the saving is real per-request visual pruning; the SPR parity 90.3 vs 90.6 is noise-level but at least directionally fine). The mechanistic claim is correlational at its core, but the authors do the right paired interventions and I would not make mechanism-attribution the load-bearing attack — the systems claims stand even if the sink story is only partially causal. The anchor-cascade concern the reader picked is legitimate and paper-acknowledged (§6), but Table 7's branch ablation shows the system degrades gracefully (semantic-only branch loses coverage, not accuracy: 84% vs 86%), so anchor failure costs recall that Stream A partially backstops. The least-secured condition for the strongest claim is therefore the measurement instrument itself: the single biggest number in the paper is produced by an uncalibrated-on-that-benchmark, rubric-weighted, single-judge, no-variance metric. This is a correctness-risk concern, not a circularity or novelty one. The reader flagged \"custom SPR judge\" in passing but ranked anchor fragility first — hence partial agreement. Verdict stays CONDITIONAL (UNCHANGED): the paper is honest about limitations and the interventions are real, but acceptance should be conditioned on independent-judge/human re-scoring of the PaperTab comparison and variance reporting.","tokens_in":17908,"tokens_out":2087,"duration_ms":79740,"concrete_test":"Re-score a stratified sample of ≥200 PaperTab and SlideVQA answers (balanced across DeCoRAG and ColQwen2/MMGraphRAG, method labels blinded) with (a) human annotators and (b) an independent judge from a different model family than the in-house 32B judge. Recompute the SPR gap under each scorer and report per-run variance over ≥3 generation seeds. If the DeCoRAG–best-baseline gap on PaperTab shrinks below ~5 points or the method ordering changes under the independent judge, the \"+12.5 points\" headline does not survive; if the gap is stable across scorers, the central systems claim is substantially de-risked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim's headline number (+12.5 SPR over the strongest baseline) comes from PaperTab: 79.3% vs ColQwen2's 66.8% (Table 4). SPR is a bespoke metric: a \"32B LLM-as-a-judge\" (§5.1) scoring Fact Consistency 40%, Information Completeness 30%, Logical Structure 20%, Expression Quality 10%, with a pass threshold of ≥0.5. Three weaknesses compound here. (1) Calibration (Table 3) covers only SPIQA (95%) and SlideVQA (86%) — 100 samples each — and never PaperTab, so the benchmark carrying the headline gain has no human-agreement evidence at all; 86% agreement on SlideVQA also means ~1 in 7 judge verdicts diverges from humans, large relative to some reported gaps. (2) The rubric hands 40% of weight to Completeness + Expression, dimensions on which DeCoRAG's pipeline (dense anchor, richer graphs — Table 2 shows more L2/L3 triples) is structurally advantaged in producing verbose, well-organized answers; a judge with known verbosity/completeness biases could inflate DeCoRAG's pass rate relative to terse retriever pipelines independently of factual correctness. (3) No variance, seeds, or judge-identity disclosure, and no code — so neither run-to-run noise nor judge-model sensitivity can be assessed. Note the reader's chosen concern (Phase-I anchor cascade) is real but partially self-mitigated inside the paper: Table 7 shows the visual-density branch alone holds 86.0% QA Pass with full graph coverage, so Stream A provides a recall floor when the anchor errs. The judge-metric concern has no such internal mitigation for the PaperTab headline.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper argues that monolithic multimodal Graph RAG construction fails on visually dense, structurally sparse documents because coupling localization, semantic interpretation, and relation extraction in one VLM pass triggers a \"Visual Attention Sink\" — attention mass escaping to boundary tokens (top-left), producing shallow, hallucinated graphs. Evidence: attention visualizations on 500 SPIQA images, a sink-severity vs. extraction-yield scatter, and a 100-page paired intervention study (Table 1) showing the sink is boundary-driven rather than content-driven and that a generated global description (\"Semantic Anchor\") reduces COCR by 77% and TL mass by 69%. The proposed system, DeCoRAG, operationalizes \"Cognitive Decoupling\": Phase I generates a Semantic Anchor from a downsampled view; Phase II (RAP-Crop) fuses morphological density boxes with OCR-grounded semantic boxes to crop Clusters-of-Interest; Phase III performs anchor-guided high-resolution extraction. Reported results: Semantic Pass Rate gains over text-only RAG, ColPali/ColQwen2, and MMGraphRAG on SPIQA/SlideVQA/PaperTab (up to +12.5 points on PaperTab), DocVQA generalization, and a 40.8% reduction in offline graph-construction prompt tokens with stable end-to-end SPR.","tokens_in":18352,"tokens_out":3729,"duration_ms":128111,"significance":"If the central claims hold, the paper contributes both a reusable diagnosis (the Visual Attention Sink as a boundary-driven failure of coupled multimodal graph construction) and a practical system (DeCoRAG) with a credible efficiency story: a 40.8% offline prompt-token reduction on FetaTab at essentially unchanged request counts and stable SPR (Table 8). Particular strengths worth naming: (1) the paired controlled interventions with a sign test (p≈1.9e-15) rather than anecdotal attention maps; (2) the honest reporting that masking controls can amplify rather than mitigate the sink; (3) cross-architecture ablation of Phase I over six VLMs (Table 6), showing the anchor is not a Qwen-specific prompt trick; (4) a branch-level RAP-Crop ablation (Table 7) that transparently shows the visual-density stream alone matches dual-stream QA Pass, giving a recall floor when the anchor errs — which partially mitigates the cascade risk flagged in §6. Weaknesses in the evaluation protocol (uncalibrated judge on the headline benchmark, rubric weights favoring verbose systems, no released code or prompts) currently prevent full confidence in the magnitude of the reported gains, though the direction (","major_comments":[{"comment":"The headline +12.5-point claim comes from PaperTab (79.3% vs ColQwen2's 66.8%), but the evaluator calibration in Table 3 covers only SPIQA (95% agreement) and SlideVQA (86%) — PaperTab has no human-agreement evidence at all, so the benchmark carrying the largest gain is exactly the one where the judge is unvalidated. This is fixable: extend the 100-sample human calibration to PaperTab (and ideally FetaTab/DocVQA). Relatedly, the 86% SlideVQA agreement implies roughly 1-in-7 judge verdicts diverge from humans; several reported gaps (e.g., SPIQA 89.2% vs 86.8%, a 2.4-point margin) are within plausible judge noise. Please report binomial confidence intervals or per-run variance for all SPR numbers in Table 4, and disclose the judge model's identity — §5.1 says only '32B LLM-as-a-judge'; if it is the same Qwen3-VL-32B family used as the system backbone, that self-evaluation risk must be addr","section":"§5.1–5.3, Tables 3–4"},{"comment":"The SPR rubric weights Information Completeness (30%) and Expression Quality (10%) at 40% combined. DeCoRAG is structurally advantaged on these dimensions: it produces denser graphs (Table 2: RDI 70.5% vs 57.3%; Table 7: 6.2K nodes/11.2K edges dual-stream vs 2.2K/3.3K semantic-only) and richer intermediate descriptions, and LLM judges have documented verbosity/completeness biases. The SPR gain could therefore partly reflect answer style rather than factual correctness, independently of the authors' intent. A load-bearing check: report the dimension-wise judge scores per system (Fact Consistency separately from Completeness/Expression), or a fact-only pass-rate variant, to show the Table 4 margins survive under a correctness-dominated weighting.","section":"§5.1, Evaluation Protocol"},{"comment":"The abstract and §5.4 claim generalization to conventional document VQA, but Table 5 reports DeCoRAG's DocVQA numbers (0.968 ANLS, 97.0% QA Pass) with no baseline whatsoever. Without ColQwen2, MMGraphRAG, or at minimum the raw backbone VLM on the same 100-sample split, the result cannot be attributed to cognitive decoupling rather than to the underlying 32B model, which is strong on DocVQA-style inputs out of the box. n=100 also warrants a confidence interval. Please add baseline rows to Table 5.","section":"§5.4, Table 5"},{"comment":"The boundary-driven (vs content-specific) causal diagnosis rests on this table, and two entries are underexplained. The control mask (masking a non-TL corner) does not merely fail to mitigate the sink — it doubles COCR from 3.49 to 6.62, and TL shift raises it to 12.57. The text glosses this as 'generic masking does not mitigate the sink,' but a twofold amplification under an off-target intervention is itself a mechanistic fact (e.g., attention renormalization over reduced visual area?) that bears on the interpretation of the anchor condition's COCR reduction (3.49→0.80). Please provide an explanation or additional controls (e.g., mask area matched to the anchor's effective information reduction) so the semantic-anchor effect can be separated from a generic 'reduce visual entropy' effect.","section":"§3.5, Table 1"}],"minor_comments":[{"comment":"Fig. 2c claims a 'negative trend' between sink density and extraction yield but reports no correlation coefficient or regression fit; please quantify (Spearman/Pearson with p-value) or soften the language.","section":"§3.4, Fig. 2c"},{"comment":"The ρ sweep reports foreground/OCR/reference recalls 'at 1.00' for all three threshold values — perfect recall across the board is surprising and suggests the recall metric may be too coarse to detect degradation; please clarify the metric definition and report at least one precision-side statistic.","section":"§5.6, Threshold Robustness"},{"comment":"'Structured-output validity 75%→85%' is cited as evidence of improved graph quality but the validity criterion is never defined (schema conformance? parseability?). Please define it.","section":"§3.5, Table 1"},{"comment":"DSSR (Dense Semantic Subsumption Rate) is defined as core L3 fact retention, but Table 2 reports it only for the anchored condition ('-' for baseline), making it uninterpretable as a comparison; either give both values or explain why only one side is meaningful.","section":"Table 2"},{"comment":"Algorithm 1, lines 9–16: after popping the largest box b and merging a single nearby b′, the merged box is appended to B_CoI without re-checking it against remaining candidates, so the result depends on pop order and does not implement full transitive merging. Either clarify that single-pass merging is intended or fix the pseudocode.","section":"Algorithm 1"},{"comment":"Presentation: 'intoGraph RAG' (§1) missing space; 'DeC-oRAG' line-break hyphenation throughout reads as a different system name; Fig. 4 contains garbled OCR-like placeholder text ('ounodone cndgh bondbo...') that should be cleaned before publication; the footnote about zeroing the bottommost attention row (§3.4) deserves to be in the main text since it affects interpretation of all attention visualizations.","section":"General"},{"comment":"No code, prompts, or judge rubric files are released. Given that the central metric is author-defined and the pipeline has several free parameters (ρ, padding α, SPR threshold), an artifact release would substantially strengthen reproducibility; at minimum please release the judge prompt and the SPR rubric.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The headline figure (+12.5 SPR) rests entirely on an author-defined judge metric whose calibration does not cover the benchmark where that figure is measured, and no code or prompts are released, so independent verification is currently impossible. The mechanistic analysis (§3, Table 1) is the stronger part of the paper and would stand even if the SPR numbers were discounted; I would encourage the editor to require (a) PaperTab calibration, (b) dimension-wise judge decomposition, and (c) artifact release as conditions of acceptance rather than treating any single one as disqualifying. The systems contribution is solid enough that these are fixable within one revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is not another RAG brand name. It is the diagnosis that open-schema multimodal graph construction over sparse high-res pages drives a boundary-sensitive visual attention sink, plus paired interventions (mask/blank/shift vs semantic anchor) that separate boundary effects from top-left content artifacts. That section is better than most systems papers bother to do.\n\nWhat they ship on top is coherent: Phase-I global anchor, dual-stream RAP-Crop, local extraction, then ordinary hybrid retrieval. Cross-backbone Phase-I gains, the branch ablation (visual-only still holds QA pass while semantic-only thins the graph), and the FetaTab token audit (≈41% prompt cut, stable end-to-end SPR, request count unchanged) are the parts I trust. DocVQA generalization is a sensible check that they did not overfit charts.\n\nSoft spots, in proportion. The stress note is right on the metric: SPR is a weighted 32B judge; human agreement is only reported for SPIQA/SlideVQA, not PaperTab—the table that carries the +12.5 vs ColQwen2. Completeness/expression weight can favor denser graph-backed answers. No variance, seeds, or code. Phase I can cascade; they admit it, and Stream A is a partial floor, not a fix. Crop ρ/α are hand-set; the small sweep is reassuring, not definitive.\n\nNone of that kills the central argument. The sink evidence and the efficiency story stand without the headline delta. This is for people building document multimodal Graph RAG who care about construction cost and grounding failure modes—not for theory people.\n\nI would send it to peer review. Engage the mechanism and RAP-Crop sections; treat SPR rankings as directional until someone re-scores with string metrics or a calibrated judge on PaperTab.","headline":"Real mechanism work plus a usable describe→crop→extract pipeline; the +12.5 SPR headline is softer than the abstract implies because PaperTab was never human-calibrated.","tokens_in":19289,"tokens_out":479,"would_cite":true,"duration_ms":18489,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Forcing a vision-language model to localize, read, and extract relations from a dense page in one step triggers a boundary-driven visual attention sink; decoupling those jobs with a global semantic anchor and region crops restores graph qua","keywords":["multimodal RAG","Graph RAG","visual attention sink","cognitive decoupling","semantic anchor","region-aware cropping","complex document understanding","vision-language models"],"falsifier":"Run the same controlled top-left mask, blank-margin, and layout-shift tests on a held-out dense-document set and check whether semantic anchoring still cuts corner over-concentration and raises valid triple yield while RAP-Crop keeps end-to-end semantic pass rate flat as prompt tokens fall ~40 percent; if anchoring fails to move attention or pass rate drops with the crops, the central claim fails.","tokens_in":19007,"feed_emoji":"📄","tokens_out":964,"duration_ms":27694,"temperature":0.7,"pith_summary":"Multimodal Graph RAG systems often ask a vision-language model to find sparse evidence, interpret a page, and emit knowledge triples in a single pass. On charts, tables, and slides this coupled stage fails: attention collapses onto image boundaries (especially the top-left), relations become shallow or hallucinated, and full-page high-resolution encoding wastes tokens on empty background. The paper shows the failure is driven by layout boundaries and structural sparsity, not by any special content in the corner, and that a brief global description can redirect attention. DeCoRAG therefore rebuilds graph construction as cognitive decoupling: first write a macroscopic semantic anchor from a cheap downsampled view, then use that anchor plus simple visual density cues to crop only clusters of interest (RAP-Crop), then extract triples from the purified high-resolution patches. The resulting graphs improve end-to-end semantic pass rate by up to 12.5 points over strong baselines and cut offline prompt tokens by about 41 percent without hurting answer quality.","feed_headline":"Decoupling stops VLMs from staring at page corners","feed_subtitle":"A global semantic anchor plus targeted crops lifts Graph RAG pass rates and cuts construction tokens 41%.","key_machinery":"Cognitive Decoupling via a Semantic Anchor plus RAP-Crop: a downsampled global description redirects attention and supplies keywords; dual-stream (morphology + OCR) proposals are greedily merged into padded clusters-of-interest that discard background before local triple extraction.","core_discovery":"When a vision-language model must simultaneously localize sparse visual evidence, interpret page semantics, and induce open-schema relations, its attention mass sinks to boundary tokens rather than information-bearing regions. That sink is boundary-driven, systematically correlates with collapsed extraction yield, and is mitigated by an explicit semantic prior. Reorganizing construction into global anchoring, dual-stream region pruning, and local high-resolution extraction therefore recovers richer graphs and cheaper encoding.","pith_inferences":["If anchor errors cascade, the natural next control loop is a cheap verify-and-rewrite step on the global description before cropping, rather than larger end-to-end VLMs.","Boundary-driven sinks may appear in other high-resolution sparse canvases (schematics, UI screenshots, maps), so the same decoupling pattern is a candidate outside financial and scientific documents.","Token savings concentrated in offline construction suggest indexing cost, not online latency, is the first place this design pays off at industrial scale."],"forward_implications":["Multimodal Graph RAG builders should stop treating full-page coupled extraction as the default graph-construction stage.","A short global description can serve as a reusable control signal that redirects VLM attention away from sink regions across model families.","Dual-stream crop-and-prune can cut offline graph-construction prompt tokens by roughly two-fifths while preserving answer quality on high-resolution pages.","The same describe-then-crop pattern generalizes from sparse charts and tables to broader document VQA layouts.","Hybrid retrieval over the recovered graphs yields higher semantic pass rates than dense visual retrievers or prior multimodal graph pipelines on the reported benchmarks."],"fun_headline_variants":["Semantic anchors stop VLM attention from sinking to page edges","DeCoRAG decouples vision and semantics to fix Graph RAG losses","Region-aware crops cut construction tokens 41% without accuracy drop","Global anchor plus RAP-Crop lifts document Graph RAG pass rates","Boundary-driven attention sinks fixed by cognitive decoupling"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The cheap global description written in the first phase is reliable enough to steer attention and cropping; if that anchor is badly wrong, later stages inherit the error.","fun_headline_variants_meta":{"raw":{"variants":["Semantic anchors stop VLM attention from sinking to page edges","DeCoRAG decouples vision and semantics to fix Graph RAG losses","Region-aware crops cut construction tokens 41% without accuracy drop","Global anchor plus RAP-Crop lifts document Graph RAG pass rates","Boundary-driven attention sinks fixed by cognitive decoupling"]},"model":"grok-4.5","effort":"low","cost_usd":0.003983,"raw_usage":{"total_tokens":1315,"prompt_tokens":868,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":39828000,"prompt_tokens_details":{"text_tokens":868,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":380,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":868,"tokens_out":67,"duration_ms":6536,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T11:45:39.618808+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same controlled top-left mask, blank-margin, and layout-shift tests on a held-out dense-document set and check whether semantic anchoring still cuts corner over-concentration and raises valid triple yield while RAP-Crop keeps end-to-end semantic pass rate flat as prompt tokens fall ~40 percent; if anchoring fails to move attention or pass rate drops with the crops, the central claim fails.","supporting_citations":[],"review_version":1}