{"id":"51bdede7-3b00-4f14-b42d-591c3db66e57","arxiv_id":"2508.16974","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper sketches a coarse-to-fine LVLM with local detail enhancement and semantic consistency validation, and claims improved accuracy on GQA, A-OKVQA, and RefCOCO, though the experimental evidence is not presented.","lead":"This paper proposes a hierarchical two-stage vision-language model that first grasps the whole image and then zooms into specific regions to answer detailed questions. The authors claim it beats several popular models on visual question answering and referring expression tasks, but they provide no code, data, or full experimental details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on Table I baseline numbers that conflict with the cited papers; absent code or protocol, the comparison cannot be verified.","rationale":"The paper's sole quantitative evidence for its headline claim is Table I, and the comparator rows are not reproducible from the cited literature. The reader's weakest assumption identifies the same issue: unsupported baseline numbers. The claimed margins are smaller than the discrepancies between Table I and published results, so the central comparison cannot support the SOTA assertion. The absence of code, checkpoints, test-set details, and a formal hallucination-evaluation protocol reinforces the rejection, but the baseline inconsistency is the most load-bearing defect.","tokens_in":11947,"tokens_out":4956,"duration_ms":51041,"concrete_test":"Reconstruct Table I from the cited primary sources: extract BLIP-2's GQA and RefCOCO/A-OKVQA scores from [10], MiniGPT-4's from [11], and Flamingo's from [9], using the exact splits and metrics the paper claims. If any baseline differs from Table I by more than 1 point, the claimed improvement margins (0.9-3.2 points) collapse. As a secondary check, request the authors' evaluation script and checkpoints and rerun the three benchmarks; if neither baseline alignment nor an executable evaluation can be supplied, the central claim remains unverified.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that HCG-LVLM outperforms Flamingo, BLIP-2, and MiniGPT-4 on GQA, A-OKVQA, and RefCOCO by 0.9-3.2 points (Table I). The load-bearing condition is that the baseline rows are accurate. That condition is not met. The numbers are inconsistent with the cited sources: BLIP-2's published GQA accuracy is in the low 40s, not 59.5; MiniGPT-4's ICLR 2024 paper does not report 60.1 on GQA or 67.3 on RefCOCO; and Flamingo has no standard GQA/RefCOCO evaluation matching 58.2/65.1. Because the claimed margins are small, replacing the baseline rows with published values would erase or invert the reported gains. The evaluation section also omits the test split, evaluation protocol, code, and checkpoints, so HCG-LVLM's own numbers cannot be independently reproduced. The hallucination comparison (9.5% vs 18.2%) comes from 200 samples with no annotation protocol, item-level scores, or inter-annotator agreement, and the abstract promises RefCOCO/+/g results that never appear in Table I. Thus the only quantitative support for the central claim is a comparison that is internally and externally inconsistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HCG-LVLM, a hierarchical architecture for fine-grained visual question answering and referring expression comprehension, combining a global contextual perception layer with a fine-grained local grounding layer that includes a Local Detail Enhancement Module, a Semantic Consistency Validator, and an adaptive fusion mechanism. The authors report state-of-the-art results on GQA (61.3%), A-OKVQA (35.0%), and RefCOCO (68.2% IoU), outperforming Flamingo, BLIP-2, and MiniGPT-4, along with a lower hallucination rate in a human evaluation. The core claim is that the hierarchical design improves fine-grained understanding and grounding. However, the experimental evidence presented is insufficient to support this claim: the baseline numbers conflict with published results, the experimental protocol is incomplete, no code or data are provided, and the human evaluation lacks statistical grounding.","tokens_in":12263,"tokens_out":3811,"duration_ms":34150,"significance":"If the reported results were reliable, the hierarchical grounding architecture would be a useful and plausible contribution to the LVLM literature, as the idea of combining global context with local verification to reduce hallucination is conceptually attractive. The paper also includes a human evaluation and an error analysis, which are welcome in principle. However, the manuscript's central quantitative claim rests on baseline numbers that appear inconsistent with the cited sources, and the absence of a reproducible experimental protocol means the significance cannot be assessed. The paper is better characterized as a position/architecture proposal than as an empirical demonstration, and in its current form it does not establish the claimed state-of-the-art performance.","major_comments":[{"comment":"The baseline numbers for Flamingo, BLIP-2, and MiniGPT-4 on GQA, RefCOCO, and A-OKVQA are not consistent with the cited publications. For example, BLIP-2's GQA accuracy in its original paper is in the low 40s, not 59.5, and MiniGPT-4's ICLR 2024 paper does not report GQA 60.1 or RefCOCO 67.3. Since the claimed improvements over these baselines are only 0.9-3.2 points, replacing the baseline rows with published values would erase or invert the reported gains, making the central claim of consistent state-of-the-art performance unsupported.","section":"Table I"},{"comment":"The experimental setup omits essential protocol details: the specific test splits for GQA and RefCOCO, the evaluation metric definitions (e.g., exact match vs. open-ended accuracy for GQA), the number of runs, random seeds, and model checkpoint sources are not specified. No code or trained models are provided, so the reported HCG-LVLM numbers cannot be independently reproduced, and the claimed improvements are therefore not verifiable.","section":"Section IV-A"},{"comment":"The human evaluation is based on 200 samples with no description of the annotation instructions, item-level score distributions, inter-annotator agreement, or a precise definition of the hallucination rate. No error bars or statistical tests are reported, so the claimed hallucination reduction (9.5% vs. 18.2%) is not statistically grounded and may not be reliable.","section":"Section IV-D and Table III"},{"comment":"The abstract and conclusion claim results on RefCOCO, RefCOCO+, and RefCOCOg, but Table I reports only RefCOCO, and no results for RefCOCO+ or RefCOCOg appear anywhere in the manuscript. This mismatch means the stated scope of the empirical claim is not met.","section":"Abstract and Section V"},{"comment":"Equations (1)-(5) are high-level functional definitions rather than a concrete model specification. The paper does not describe how the region proposals A_G are generated, how the local encoder is parameterized, how the similarity function and fusion weights are computed, or how the consistency loss L_consistency is combined with task losses. The ablation study in Table II reports single numbers without variance or significance testing, so the individual contributions of LDE, SCV, and adaptive fusion are not established.","section":"Sections III-A to III-C and Table II"}],"minor_comments":[{"comment":"The paper repeatedly calls the baselines 'state-of-the-art' but does not compare against any of the more recent methods listed in the related work (e.g., InstructBLIP, HiVG, patch-aligned training), which limits the relevance of the comparison.","section":"Section I and IV"},{"comment":"The efficiency comparison does not specify the hardware, batch size, or model variant for the latency measurements, and the Flamingo parameter count (80B) is given without stating which Flamingo variant is used; these details are needed to interpret the comparison.","section":"Table VI"},{"comment":"Several references appear mismatched to their citing context, such as a schizophrenia bulletin article used to support hallucination in LVLMs and a CVPR 2025 reference for REC evaluation that is not clearly related to the claims made; the reference list should be checked for accuracy.","section":"References"},{"comment":"The human evaluation reports five annotators but does not disclose their qualifications, whether they were authors, or how disagreements were resolved; this information is necessary to assess the reliability of the scores.","section":"Section IV-D"},{"comment":"Table I would benefit from a caption note specifying the exact metric for each dataset (e.g., GQA test-dev accuracy, RefCOCO testA/testB IoU) and from error bars or confidence intervals, as the small margins between models are otherwise difficult to interpret.","section":"Table I"}],"recommendation":"reject","confidential_remarks":"The reported baseline numbers in Table I conflict with publicly available results from the cited papers, and this discrepancy is load-bearing for the paper's central claim. The authors should be asked to provide the exact sources of the baseline numbers and to release code and checkpoints before any resubmission is considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is not ready for circulation. The core idea — a coarse-to-fine LVLM with a local detail enhancement module and a semantic consistency validator, fused adaptively — is coherent and clearly explained. The authors also set up a reasonable ablation logic. But the central SOTA claim is unsupported: the numbers in Table I for Flamingo, BLIP-2, and MiniGPT-4 do not match the published results for those models on GQA, A-OKVQA, and RefCOCO. BLIP-2's own GQA accuracy is in the low 40s, not 59.5; MiniGPT-4's ICLR paper does not report 60.1 on GQA or 67.3 on RefCOCO; Flamingo has no standard GQA/RefCOCO evaluation that gives 58.2/65.1. Because the claimed margins are only 0.9–3.2 points, replacing the baseline rows with correct values would erase or invert the gains. That is a load-bearing flaw, not a minor one.\n\nThere is also no code, no checkpoints, no test split or metric definition, and no annotation protocol for the human study (200 samples, five annotators, no inter-annotator agreement). The ablation table has no error bars. The manuscript even contains a textual artifact ('referencing the provided summary's data'), and the abstract promises RefCOCO/+/g results that never appear in Table I. The equations (1)–(5) are generic formal definitions; they do not describe how the region proposals, local encoder, or consistency loss are actually instantiated. On novelty: crop-and-attend for local features and contrastive validation are established techniques, which the related work itself acknowledges [23–25]; the contribution is a new packaging, not a new mechanism.\n\nWhat the paper does well: the high-level design is sensible, the failure-mode analysis is honest, and the computational efficiency table, while also unreferenced, is at least plausible. The writing is structured and readable. This looks like a solid workshop-paper skeleton, not a finished claim.\n\nBottom line: desk reject. If the authors release code, correct the baseline table to numbers traceable to the cited papers, and provide a complete evaluation protocol, the work could be worth a second look. As submitted, there is no verifiable evidence for the main claim, and the comparison with baselines is internally inconsistent.","headline":"A sensible coarse-to-fine LVLM sketch whose SOTA claim collapses because the baseline numbers in Table I don't match the cited papers, and no code or protocol is provided.","tokens_in":12717,"tokens_out":3450,"would_cite":false,"duration_ms":32120,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two-pass coarse-to-fine architecture lifts fine-grained VQA and grounding while halving hallucination.","keywords":["HCG-LVLM","visual grounding","fine-grained visual question answering","referring expression comprehension","hallucination reduction","hierarchical vision-language model","local detail enhancement","semantic consistency validation"],"falsifier":"Run the exact GQA, A-OKVQA, and RefCOCO evaluations using the public checkpoints of Flamingo, BLIP-2, and MiniGPT-4 with the same prompts, decoding settings, and preprocessing, and recompute Table I; if any baseline meets or exceeds 61.3% on GQA, 68.2% IoU on RefCOCO, or 35.0% on A-OKVQA, the claimed state-of-the-art result is not supported.","tokens_in":11805,"feed_emoji":"🔍","tokens_out":7648,"duration_ms":71366,"temperature":0.7,"pith_summary":"The paper sets out to show that a large vision-language model becomes more accurate and less prone to hallucination on fine-grained tasks if it processes an image twice: first to grasp the whole scene, then to zoom into the regions that matter and verify the query's details against the actual pixels. The proposed HCG-LVLM couples a pre-trained global LVLM with a local detail enhancement module and a semantic consistency validator, fusing both passes adaptively. On GQA, A-OKVQA, and RefCOCO it reports gains over Flamingo, BLIP-2, and MiniGPT-4, plus a human-evaluated hallucination rate roughly half that of MiniGPT-4. The contribution, if it holds, is a general architectural pattern rather than a single new dataset or loss.","feed_headline":"Zoom-and-verify beats one-pass models on fine-grained tests","feed_subtitle":"Two-layer model checks image regions against the query: 61.3% on GQA, 9.5% hallucination.","key_machinery":"The load-bearing mechanism is a cascade of three parts: Global Contextual Perception, a pre-trained ViT-plus-LLM backbone that outputs coarse region proposals; Local Detail Enhancement, which crops each proposed region and re-encodes it at higher resolution; and a Semantic Consistency Validator, which computes a similarity or contrastive score between each local feature and the text query embedding. An adaptive fusion layer then combines global and local evidence. The design's purpose is to force fine-grained claims to be checked against image patches rather than generated from global statistics.","core_discovery":"The central discovery claimed is that coarse-to-fine hierarchical processing improves fine-grained visual-language understanding: a first pass over the whole image produces coarse region proposals, and a second pass crops those regions, extracts higher-resolution features, and scores them against the text query, so the final answer is grounded in verified local evidence. Concretely, the paper reports 61.3% accuracy on GQA, 35.0% on A-OKVQA, and 68.2% IoU on RefCOCO, exceeding the quoted scores for Flamingo, BLIP-2, and MiniGPT-4 on all three, and a hallucination rate of 9.5% versus MiniGPT-4's 18.2%. The ablations attribute the gain to the local detail enhancer and the semantic consistency validator together, not to the global backbone alone.","pith_inferences":["If the central claim holds, the same coarse-to-fine two-pass pattern could be added to other single-pass vision-language models as a modular post-processor: crop attention-highlighted regions, re-encode them, and gate the answer on local consistency, without retraining the whole backbone.","A testable extension the paper does not run is varying the global layer's proposal quality, for example comparing ground-truth boxes against noisy attention crops; the architecture predicts that better proposals amplify the final gains on RefCOCO more than on GQA.","The hallucination reduction is reported from one 200-sample human study; a natural follow-up is to decompose hallucinations into object-presence, attribute, and spatial-relation errors, since local verification should suppress the object-presence and attribute types more strongly."],"forward_implications":["The full model should beat its own global-only backbone by roughly 3.5 points on GQA and 3.7 IoU on RefCOCO, matching the reported ablation.","Removing the Semantic Consistency Validator should reduce accuracy and raise hallucination in attribute-heavy queries such as color and size judgments.","Adaptive fusion should outperform simple concatenation when global and local evidence conflict, because low-consistency local regions can be down-weighted.","The reported latency increase over MiniGPT-4 (145 ms versus 110 ms per query) is the price of the accuracy and grounding gains, under the paper's efficiency measurements."],"supporting_citations":[{"why":"Supplies the Flamingo baseline whose reported scores the paper's gains are measured against on all three benchmarks.","marker":"[9]"},{"why":"Supplies the BLIP-2 baseline numbers for GQA, RefCOCO, and A-OKVQA that anchor the claimed improvements.","marker":"[10]"},{"why":"Supplies the MiniGPT-4 baseline for both accuracy comparison and the human hallucination evaluation.","marker":"[11]"},{"why":"Defines the GQA benchmark that provides the fine-grained compositional VQA accuracy numbers.","marker":"[12]"},{"why":"Defines the A-OKVQA benchmark that provides the knowledge-based VQA accuracy numbers.","marker":"[13]"},{"why":"Defines the RefCOCO/+/g benchmark that provides the referring expression comprehension IoU numbers.","marker":"[14]"}],"fun_headline_variants":["Coarse-to-fine vision beats one-pass on fine-grained tasks","Two-step zoom grounds vision, cuts hallucination by half","Global pass, local verify: hierarchical LVLM improves fine-grained grounding","HCG-LVLM: hierarchical grounding lifts fine-grained VQA to 61.3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison rests on the reported baseline scores for Flamingo, BLIP-2, and MiniGPT-4; if those published numbers are not measured under the same protocol or are inaccurate, the claimed gains lose their support.","fun_headline_variants_meta":{"raw":{"variants":["Coarse-to-fine vision beats one-pass on fine-grained tasks","Two-step zoom grounds vision, cuts hallucination by half","Global pass, local verify: hierarchical LVLM improves fine-grained grounding","HCG-LVLM: hierarchical grounding lifts fine-grained VQA to 61.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000797,"raw_usage":{"total_tokens":3525,"prompt_tokens":984,"completion_tokens":2541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2462}},"tokens_in":600,"tokens_out":2541,"duration_ms":21488,"temperature":1.0,"reasoning_tokens":2462,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:08:16.286175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact GQA, A-OKVQA, and RefCOCO evaluations using the public checkpoints of Flamingo, BLIP-2, and MiniGPT-4 with the same prompts, decoding settings, and preprocessing, and recompute Table I; if any baseline meets or exceeds 61.3% on GQA, 68.2% IoU on RefCOCO, or 35.0% on A-OKVQA, the claimed state-of-the-art result is not supported.","supporting_citations":[{"cited_title":"Flamingo: a visual language model for few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Flamingo baseline whose reported scores the paper's gains are measured against on all three benchmarks."},{"cited_title":"BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models,","cited_arxiv_id":null,"evidence_quote":"Supplies the BLIP-2 baseline numbers for GQA, RefCOCO, and A-OKVQA that anchor the claimed improvements."},{"cited_title":"GQA: A new dataset for real- world visual reasoning and compositional question answering,","cited_arxiv_id":null,"evidence_quote":"Defines the GQA benchmark that provides the fine-grained compositional VQA accuracy numbers."},{"cited_title":"A-OKVQA: A benchmark for visual question answering using world knowledge,","cited_arxiv_id":null,"evidence_quote":"Defines the A-OKVQA benchmark that provides the knowledge-based VQA accuracy numbers."},{"cited_title":"Revisiting referring expression comprehension evaluation in the era of large multimodal models,","cited_arxiv_id":null,"evidence_quote":"Defines the RefCOCO/+/g benchmark that provides the referring expression comprehension IoU numbers."}],"review_version":2}