{"id":"3e0d9fe4-f6c4-469f-bdba-51afe610a05d","arxiv_id":"2607.05859","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"AVA-VLM reduces visual-token usage by 69% while improving PPE-violation F1 by 13 points over direct-QA baselines by training a VLM to adaptively crop high-resolution local regions from a downsampled global image.","lead":"This paper builds a vision-language model for construction sites that first looks at a low-resolution overview image, then selectively zooms into specific regions for detail—like a human inspector. It matters because it cuts visual-token usage by ~70% while improving violation detection, especially for distant or low-resolution camera feeds.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Train/inference resolution gap for crop-box prediction is the load-bearing risk: the model learns to predict crop coordinates from full-resolution images but must do so from downsampled images at test time, and crop prediction quality is never directly measured.","rationale":"The reader correctly identified the most load-bearing concern: crop-region prediction quality is never directly evaluated, and the train/inference resolution gap is an unstated risk. I agree with this assessment. The reader's verdict of CONDITIONAL is appropriate — the paper presents a well-executed contribution with thorough experiments and honest reporting, but the missing direct evaluation of crop prediction quality under the resolution mismatch prevents full confidence in the central claim. The paper's own Table 6 and Figure 5(d) provide indirect evidence that the resolution gap matters (performance degradation and reduced tool-call triggering under downsampling), which strengthens the case that this concern is real rather than hypothetical. However, the paper does show that AVA-VLM outperforms baselines even under downsampling, and the gradual rather than catastrophic degradation suggests the model generalizes to some extent. The CONDITIONAL verdict with MODERATE confidence is well-calibrated. No adjustment needed. The concrete test I propose (direct crop-box IoU evaluation by resolution and distance) would settle whether the concern lands: if crop predictions remain accurate under downsampling, the framework is sound; if they degrade substantially, the token-efficiency claim comes at an unacknowledged accuracy cost that the downstream metrics partially mask.","tokens_in":40772,"tokens_out":2736,"duration_ms":179316,"concrete_test":"Compute IoU between predicted crop boxes and ground-truth crop boxes (from Algorithms 1-2) on the test set, broken down by global image resolution (full, 1/2, 1/4) and camera distance. This directly tests whether the model can accurately predict where to crop from downsampled inputs. If the mean crop IoU drops by more than ~15 percentage points from full-size to 1/2-downsampled inputs, the headline claim's reliance on downsampled inference is undermined, because the model's crop predictions would be too unreliable to recover the evidence that justifies the token savings.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the model's ability to decide where to crop from a low-resolution global image. During training (Section 4.1, Algorithm 3), the model always receives full-size global images and learns to predict crop boxes from them. At inference (Section 4.2, Eq. 14-15), the default configuration uses 1/2-downsampled global images, and the predicted crop box in downsampled coordinates is linearly scaled back to full-resolution (ŝ^G = s · b̂^C). The model has never been trained on downsampled inputs, so its crop-prediction behavior in this regime is uncharacterized. Table 6 shows F1 degrades from 82.7% (full-size) to 75.1% (1/2 downsampled) to 64.5% (1/4 downsampled), and Figure 5(d) shows tool-call ratio drops from 35.8% to 20.0% as resolution decreases — suggesting the model increasingly fails to trigger crops when it should. The paper never directly evaluates crop-box prediction quality (e.g., IoU between predicted crop boxes and ground-truth crop boxes from Algorithms 1-2). The IoU metrics in Table 7 measure violation-localization quality, which conflates crop quality with answer-generation quality and only covers VI positive samples where the model predicts a violation. For no-violation samples that trigger cropping (for detailed inspection) and for all OD samples, crop quality is entirely unassessed. If crop predictions are poor under downsampled inputs, the model either crops irrelevant regions (wasting tokens without recovering useful evidence) or misses critical regions (reducing accuracy). The training design's explicit decoupling of crop prediction from answer generation (using ground-truth crops during training, Section 4.1) means the answer-generation module has also never been trained on imperfect crops, so errors at inference compound without recovery.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes AVA-VLM, a construction-site VLM that combines low-resolution global reasoning with selective high-resolution local cropping via a tool-call mechanism. The model is trained on a region-aware CoT dataset derived from ConstructionSite10K, where crop decisions and regions are generated from ground-truth annotations and a YOLO detector. At inference, the model predicts whether to crop and where, using a downsampled global image and mapping predicted crop coordinates back to full resolution. Experiments compare AVA-VLM against direct-QA and example-image CoT baselines on violation identification (VI), object detection (OD), and image captioning (IC), evaluating across camera distances and reduced resolutions. The central claim is that AVA-VLM achieves better F1 on PPE-violation identification (75.1 vs. 62.0 baseline) while using only 30.6% of the visual-token budget, and maintains advantages under reduced resolution and long-distance scenarios.","tokens_in":41728,"tokens_out":1444,"duration_ms":297105,"significance":"The paper addresses a practically important problem: making VLMs more efficient and robust for construction-site monitoring under realistic deployment constraints. The experimental design is commendable—same backbone (Qwen2.5-VL 7B), same dataset, same hyperparameters across all methods, with multiple evaluation axes (camera distance, resolution, per-class metrics). The region-aware CoT dataset construction is a useful contribution, and the selective cropping mechanism is a reasonable approach to the efficiency-accuracy trade-off. The paper is transparent about performance degradation under aggressive downsampling and about the OD trade-off. However, the central claim rests on the model's ability to predict crop regions from downsampled inputs at inference, and this capability is never directly evaluated, which weakens the evidentiary basis for the headline results.","major_comments":[{"comment":"The paper never directly evaluates crop-region prediction quality (e.g., IoU between predicted crop boxes and ground-truth crop boxes from Algorithms 1–2). The entire framework depends on the model's ability to accurately predict WHERE to crop from a downsampled global image, yet this is only indirectly assessed through downstream task metrics. The IoU metrics in Table 7 measure violation-localization quality, which conflates crop quality with answer-generation quality and only covers VI positive samples where the model predicts a violation. For no-violation samples that trigger cropping and for all OD samples, crop quality is entirely unassessed. Adding a direct crop-box IoU evaluation would substantially strengthen the paper's central claim. (§4.2, Eqs. 14–15; Algorithms 1–2; Table 7)","section":null},{"comment":"Tables 4, 7, and 10 have inconsistent or unclear descriptions of the AVA-VLM configuration. Table 4's caption states 'AVA-VLM uses 1/4 downsampled global images,' but Table 6 shows that 1/4 downsampling yields F1=64.5%, while 1/2 downsampling yields F1=75.1%. The text in §5.2.1 says the 1/2 setting is used as the 'default AVA-VLM configuration in the overall comparison reported in Table 4.' This means Table 4's caption is incorrect—it should say 1/2, not 1/4. This inconsistency affects interpretation of the efficiency claims. (Table 4 caption; Table 6; §5.2.1 final paragraph)","section":null},{"comment":"The train/inference resolution gap is a load-bearing concern. During training (§4.1, Algorithm 3), the model always receives full-size global images and learns to predict crop boxes from them. At inference (§4.2, Eqs. 14–15), the default configuration uses 1/2-downsampled global images, and the predicted crop box in downsampled coordinates is linearly scaled back to full resolution. The model has never been trained on downsampled inputs, so its crop-prediction behavior in this regime is uncharacterized. Table 6 shows F1 degrades from 82.7% (full-size) to 75.1% (1/2) to 64.5% (1/4), and Figure 5(d) shows tool-call ratio drops from 35.8% to 20.0% as resolution decreases, suggesting the model increasingly fails to trigger crops when it should. The paper should discuss this train/test distribution shift explicitly and consider whether training with augmented downsampled inputs could mitigate","section":null}],"minor_comments":[{"comment":"Section 3.2: The free parameters (tau_obj_VI, tau_crop_VI, tau_obj_OD, tau_crop_OD, alpha) are used in Algorithms 1–2 but their values are not reported. Please include them in the implementation details or supplementary material.","section":null},{"comment":"Table 2: The 'Test images' column (1,360 / 1,310 / 334) does not match the test image count in Table 1 (3,004 total). Please clarify whether Table 2 reports a subset or a different grouping.","section":null},{"comment":"Section 5.1: The downsampling factor s is described in §4.2 but its specific values for each Table 6 setting should be stated explicitly in the implementation details for reproducibility.","section":null},{"comment":"Figure 5(d) is referenced in the text as showing tool-call ratio vs. resolution, but the figure label says 'reduced global image resolutions' without specifying the exact scales (1/2, 1/4). Consider adding scale labels directly on the figure.","section":null},{"comment":"Table 10: AVA-VLM's OD performance (60.3% Avg. IoU) is below the baseline (68.1%). The text acknowledges this trade-off, but the abstract's claim of 'substantially reducing visual-token usage' could be read as implying no accuracy cost. Consider qualifying the abstract to note the OD trade-off.","section":null},{"comment":"Reference [12] (Chen and Zou, 2026) and several other references have 2026 dates. Please verify these are correct publication dates and not placeholder dates.","section":null},{"comment":"Section 4.1: The notation m^(·)_i uses a dot placeholder for the task type but this is not formally defined. Consider explicitly stating that the dot denotes the task type (VI or OD).","section":null}],"recommendation":"major_revision","confidential_remarks":"The train/inference resolution gap is the most substantive concern. The paper's own Table 6 data shows significant degradation under downsampling, and the tool-call ratio drop in Figure 5(d) suggests the model's cropping behavior is fragile under the distribution shift. However, the paper is transparent about these limitations and the core contribution (region-aware CoT + selective cropping) is sound. A direct crop-box IoU evaluation and corrected table captions should be feasible within revision. The OD performance trade-off is honestly reported and does not undermine the paper's claims, but the abstract should be more balanced."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee correctly identifies that the central mechanism of AVA-VLM—predicting crop regions from downsampled global images—deserves direct evaluation, and also identifies a caption error and an important train/test distribution shift. We address each point below.","responses":[{"response":"The referee is correct that we do not directly evaluate crop-region prediction quality, and we agree this is a gap. We will add a direct crop-box IoU evaluation in the revision. Specifically, for all samples where the model triggers a crop at inference (covering VI positive, VI negative, and OD samples), we will compute IoU between the predicted crop box and the ground-truth crop box derived from Algorithms 1–2. This will be reported separately from the downstream task metrics to decouple crop quality from answer-generation quality. We note that the current Table 7 IoU metrics do conflate crop quality with answer quality as the referee states, so the direct crop-box IoU evaluation will provide a cleaner assessment of the mechanism the referee rightly identifies as load-bearing.","revision_made":"yes","referee_comment":"The paper never directly evaluates crop-region prediction quality (e.g., IoU between predicted crop boxes and ground-truth crop boxes from Algorithms 1–2). The entire framework depends on the model's ability to accurately predict WHERE to crop from a downsampled global image, yet this is only indirectly assessed through downstream task metrics. The IoU metrics in Table 7 measure violation-localization quality, which conflates crop quality with answer-generation quality and only covers VI positive samples where the model predicts a violation. For no-violation samples that trigger cropping and for all OD samples, crop quality is entirely unassessed. Adding a direct crop-box IoU evaluation would substantially strengthen the paper's central claim."},{"response":"The referee is correct. Table 4's caption states '1/4 downsampled global images,' but the text in §5.2.1 explicitly states that the 1/2-width-and-height setting is used as the default configuration for the overall comparison in Table 4. The caption is erroneous. We will correct the captions of Tables 4, 7, and 10 to state '1/2 downsampled global images' to match the actual configuration used. We will also carefully audit all table captions for consistency. We note that the efficiency claims themselves are unaffected: the 30.6% visual-token figure in Table 4 corresponds to the 1/2 setting (25% from the downsampled global image plus 5.6% from cropped images), as confirmed by Table 6's 1/2 row.","revision_made":"yes","referee_comment":"Tables 4, 7, and 10 have inconsistent or unclear descriptions of the AVA-VLM configuration. Table 4's caption states 'AVA-VLM uses 1/4 downsampled global images,' but Table 6 shows that 1/4 downsampling yields F1=64.5%, while 1/2 downsampling yields F1=75.1%. The text in §5.2.1 says the 1/2 setting is used as the 'default AVA-VLM configuration in the overall comparison reported in Table 4.' This means Table 4's caption is incorrect—it should say 1/2, not 1/4. This inconsistency affects interpretation of the efficiency claims."},{"response":"The referee raises a valid concern. There is indeed a train/test resolution gap: the model is trained on full-size global images but evaluated on downsampled inputs at inference. We agree this distribution shift is not currently discussed in the paper, and we will add an explicit discussion in the revision. The degradation pattern the referee identifies—F1 dropping from 82.7% to 75.1% to 64.5% and tool-call ratio dropping from 35.8% to 20.0%—is consistent with the hypothesis that downsampled inputs make it harder for the model to recognize when cropping is needed, leading to missed crops and increased false negatives. We will acknowledge this limitation transparently and discuss training with resolution augmentation (i.e., randomly downsampling global images during training) as a promising mitigation strategy. We are not able to run new training experiments with resolution augmentation before the revision deadline, so we will frame this as a concrete future direction rather than reporting results. We believe this is an honest accounting of what we can and cannot address at this stage.","revision_made":"partial","referee_comment":"The train/inference resolution gap is a load-bearing concern. During training (§4.1, Algorithm 3), the model always receives full-size global images and learns to predict crop boxes from them. At inference (§4.2, Eqs. 14–15), the default configuration uses 1/2-downsampled global images, and the predicted crop box in downsampled coordinates is linearly scaled back to full resolution. The model has never been trained on downsampled inputs, so its crop-prediction behavior in this regime is uncharacterized. Table 6 shows F1 degrades from 82.7% (full-size) to 75.1% (1/2) to 64.5% (1/4), and Figure 5(d) shows tool-call ratio drops from 35.8% to 20.0% as resolution decreases, suggesting the model increasingly fails to trigger crops when it should. The paper should discuss this train/test distribution shift explicitly and consider whether training with augmented downsampled inputs could mitigate"}],"tokens_in":40781,"tokens_out":1212,"duration_ms":187927,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things to know up front: (1) this paper applies adaptive coarse-to-fine visual attention to construction-site VLMs and gets a genuine win — 70% visual-token reduction with improved PPE-violation F1 (75.1 vs. 62.0 baseline). (2) The central mechanism — the model deciding where to crop from a low-resolution global image — is never directly evaluated, and the train/inference resolution gap makes this a real soft spot, not a hypothetical one. The paper is still worth a serious referee. What's actually new: the region-aware CoT dataset construction is thoughtful. The cropping algorithms (Algorithms 1–2) use camera distance, object area thresholds, and YOLO detections to generate ground-truth crop decisions, and the annotation-level crop ratio of 16.2% shows the dataset teaches selective tool use rather than always-crop. The experimental design is clean — same backbone (Qwen2.5-VL 7B), same data, same hyperparameters across all methods. The paper honestly reports that OD performance drops (60.3 vs. 68.1 IoU) and doesn't hide the trade-off. The distance-stratified and resolution-stratified analyses (Tables 5–6, 8–9, 11–12) are thorough and show AVA-VLM's advantage grows exactly where you'd expect: long distance and reduced resolution. The stress-test concern about the train/inference resolution gap lands. During training, the model always sees full-resolution global images and learns crop boxes from them. At inference, the default config feeds 1/2-downsampled images, and the model must predict crop coordinates from inputs it was never trained on. Table 6 and Figure 5(d) show the consequences: F1 drops from 82.7 to 75.1, and tool-call ratio drops from 35.8% to 30.0%, meaning the model increasingly fails to trigger crops when it should. The paper acknowledges this degradation but never directly measures crop-box prediction quality (e.g., IoU between predicted and ground-truth crop boxes). The IoU metrics in Table 7 conflate crop quality with answer quality and only cover VI positive samples. For no-violation samples that trigger cropping and for all OD samples, crop quality is entirely unassessed. This is the load-bearing gap. That said, the paper still outperforms baselines at every resolution setting, so the mechanism works — it's just uncharacterized how well. The missing crop-quality evaluation is a condition for full acceptance, not a fatal flaw. The broader adaptive-attention idea exists in VisionThink and AdaptVision, so novelty is in the application and dataset, not the mechanism itself. No code or data is shipped, which limits reproducibility. Recommendation: accept for peer review. The paper deserves a serious referee who should ask for (1) direct crop-box IoU evaluation, (2) explicit acknowledgment of the train/inference resolution mismatch, and (3) stated values for the crop annotation thresholds. The core contribution is real and the experiments are honest; the gaps are addressable in revision.","headline":"Solid applied contribution with a real evaluation gap on the load-bearing mechanism","tokens_in":41668,"tokens_out":1040,"would_cite":false,"duration_ms":89952,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"VLM learns to zoom in, cuts visual tokens by 70%","keywords":["vision-language model","construction safety","adaptive visual attention","coarse-to-fine reasoning","visual token efficiency","PPE violation detection","chain-of-thought"],"falsifier":"If crop-region prediction quality (IoU between predicted and ground-truth crop boxes) is poor at inference, the model either wastes tokens on irrelevant regions or misses critical evidence, collapsing the performance gains.","tokens_in":40839,"feed_emoji":"🔍","tokens_out":375,"duration_ms":189255,"temperature":0.7,"pith_summary":"A vision-language model for construction-site monitoring that first looks at a downsampled global image and selectively crops high-resolution regions only when needed, improving PPE-violation detection while using less than a third of the visual tokens of standard approaches.","feed_headline":"VLM cuts visual tokens 70% by learning when to zoom in","feed_subtitle":"Adaptive crop-and-inspect strategy beats direct-QA baselines on construction-site PPE violation detection while using a third of the compute","key_machinery":"Region-aware CoT dataset, adaptive crop tool, decoupled training","core_discovery":"The paper introduces a coarse-to-fine reasoning strategy where a VLM trained on a region-aware chain-of-thought dataset learns to decide when local inspection is needed, where to crop, and how to incorporate cropped evidence, achieving 75.1 F1 on PPE-violation identification versus 62.0 for direct-QA baselines while using only 30.6% of the visual-token budget.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Adaptive crop-and-inspect VLM cuts token use while raising F1 to 75.1","Coarse-to-fine VLM hits 75.1 F1 on construction sites with a third of tokens","VLM trained to zoom in on construction sites boosts F1 and cuts visual tokens","Region-aware VLM outperforms direct-QA baselines using 70% fewer visual tokens","Adaptive zoom boosts VLM reliability at distance while cutting visual tokens 70%"],"cache_read_input_tokens":0,"weakest_assumption_plain":"During training, the model is given ground-truth crop regions rather than having to predict them itself, so the framework's success at inference depends entirely on the model's ability to identify the right regions from a low-resolution global image—a capability never directly measured.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive crop-and-inspect VLM cuts token use while raising F1 to 75.1","Coarse-to-fine VLM hits 75.1 F1 on construction sites with a third of tokens","VLM trained to zoom in on construction sites boosts F1 and cuts visual tokens","Region-aware VLM outperforms direct-QA baselines using 70% fewer visual tokens","Adaptive zoom boosts VLM reliability at distance while cutting visual tokens 70%"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1345,"prompt_tokens":456,"completion_tokens":889,"prompt_tokens_details":null},"tokens_in":456,"tokens_out":889,"duration_ms":58928,"temperature":1.0,"reasoning_tokens":612,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T22:09:28.197885+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If crop-region prediction quality (IoU between predicted and ground-truth crop boxes) is poor at inference, the model either wastes tokens on irrelevant regions or misses critical evidence, collapsing the performance gains.","supporting_citations":[],"review_version":1}