{"id":"08f92ea3-0491-4428-b529-19b9ac324f3f","arxiv_id":"2607.06909","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"LePaX enables high-resolution chest X-ray report generation by learning to allocate resolution to diagnostically relevant regions and fusing high-res patches back into global features without increasing token count.","lead":"The paper introduces LePaX, a system that lets AI read chest X-rays at full resolution by selectively zooming into suspicious areas, mimicking how radiologists work. This matters because current AI systems shrink X-rays so much they miss subtle signs of disease, and this approach fixes that without requiring more computing power.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The ablation does not isolate whether gains come from high-resolution patch content or from the GRF cross-attention mechanism itself; a low-res patch control is missing.","rationale":"The reader identified Grad-CAM prior quality as the weakest assumption, but the ablation shows the Grad-CAM prior contributes only 0.007 B-4 (Table 5: report-only 0.246 vs report+policy 0.253), making it a secondary concern. The more load-bearing issue is whether the performance gains actually come from high-resolution information or from the GRF mechanism's added model capacity. The paper's Table 6 provides partial evidence that higher source resolution helps (0.245 at 1024 vs 0.253 at 1920), but this comparison is confounded by patch content differences and does not isolate the GRF mechanism's contribution. The proposed low-res patch control is a clean, inexpensive experiment that would directly test the central claim. The reader's CONDITIONAL verdict is appropriate: the framework is well-motivated and the gains are likely real, but this specific gap in the ablation prevents full confidence in the 'high-res perception' claim specifically. The concern does not warrant rejection — the existing evidence is suggestive — but addressing it would move the paper from defensible to convincing. I also note the reader's concern about Table 1 comparing heterogeneous base models (4B vs 7B vs 14.2B) is valid: the ablation (Table 5/6) shows LePaX-specific gains of ~0.016 B-4 over the same base model, which is modest compared to the large gaps in Table 1 that are mostly attributable to base model choice. This framing issue reinforces the need for the proposed control to clarify what LePaX specifically contributes.","tokens_in":17534,"tokens_out":6663,"duration_ms":234831,"concrete_test":"On MIMIC-CXR, run the full LePaX pipeline (GRF+LSRA, Top-K=5) but replace high-res patches (384×384 crops from 1920×1920) with low-res patches: crop the same spatial regions from the 384×384 downsampled global image and bilinearly upscale to 384×384. If B-4 stays near 0.253, the high-resolution information is not load-bearing. If B-4 drops substantially (e.g., below 0.245, approaching the Uniform-384 baseline of 0.237), the high-res content is genuinely contributing and the central claim is strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LePaX 'enables efficient high-res CXR perception.' Table 5 shows that GRF with random high-res patches already improves B-4 from 0.237 to 0.243 (+0.006) over the no-GRF baseline. However, this ablation confounds two factors: (1) GRF adds a Transformer cross-attention block (Eq. 11) that increases model capacity, and (2) the high-res patches (384×384 crops from 1920×1920) provide finer-grained diagnostic information. If the same GRF mechanism were applied to low-res patches (384×384 crops from the already-downsampled 384×384 global image, i.e., bilinearly upscaled crops of the same spatial regions), and performance remained similar, then the gain would be attributable to the additional attention mechanism rather than to high-resolution perception — directly undermining the paper's core claim. Table 6 partially addresses this by showing Ours-1024 (0.245) vs Ours-1920 (0.253), suggesting higher source resolution helps. But both settings use GRF+LSRA, so the resolution effect is still entangled with patch content differences (a 384×384 crop from 1920 covers a smaller anatomical region than from 1024). The critical control — GRF with low-res patches vs GRF with high-res patches under identical selection — is absent. Without it, the paper cannot distinguish 'high-resolution perception helps' from 'an extra attention block helps.'","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript proposes LePaX, a radiology report generation (RRG) framework that enables high-resolution (up to 1920x1920) chest X-ray perception without increasing the visual token budget. It introduces two components: Learnable Spatial Resolution Allocation (LSRA), which learns to select diagnostically relevant regions for high-resolution patch extraction, and Global-Regional Fusion (GRF), which writes high-resolution regional features back into the global feature grid via cross-attention. The method is evaluated on MIMIC-CXR, IU-Xray, and CheXpertPlus, showing improvements in both NLG and clinical metrics over prior baselines.","tokens_in":17801,"tokens_out":1385,"duration_ms":121616,"significance":"The problem addressed is well-motivated: standard RRG pipelines downsample CXRs to low resolutions (e.g., 256x256), potentially suppressing subtle lesions. The formulation of high-resolution perception as a constrained spatial resolution allocation problem is a principled and novel contribution to the RRG field. The framework is largely self-contained, and the use of a Grad-CAM prior from an external classifier to regularize the allocation policy during training, while dropping it at inference, is a practical design choice. The reported efficiency gains (maintaining a fixed 128-token budget while processing 1920x1920 images) are significant for the deployment of MLLMs in medical imaging.","major_comments":[{"comment":"The central claim that gains are attributable to 'high-resolution perception' is not fully isolated by the ablation study. In Table 5, introducing GRF with high-resolution patches improves B-4 from 0.237 to 0.243. However, GRF inherently adds a Transformer cross-attention block (Eq. 11), increasing model capacity. The manuscript lacks a critical control: applying the GRF mechanism to low-resolution patches (e.g., upscaled crops from the low-res global image) to determine if the performance gain comes from the high-resolution patch content or simply from the additional attention mechanism. Without this control, the paper cannot definitively distinguish 'high-resolution perception helps' from 'an extra attention block helps.' The authors should add this ablation to substantiate the core claim.","section":null},{"comment":"Table 1 compares LePaX (4B parameters) against baselines with varying model sizes, including 7B and 14.2B models. While LePaX achieves superior performance, the comparison is confounded by parameter count. The manuscript does not discuss whether the baselines were re-implemented or evaluated under the same parameter budget, nor does it normalize for model size. The authors should clarify if the 4B model size is a deliberate architectural constraint or a limitation, and ideally provide a comparison against a baseline of equivalent size to ensure fair assessment.","section":null},{"comment":"The '10x fewer visual tokens' claim in the abstract requires explicit validation in the experiments. Table 6 shows that Uniform-1024 uses 1280 tokens and Ours-1920 uses 128 tokens, which is indeed a 10x reduction. However, the abstract frames this as a comparison against 'naive high-res tiling.' The manuscript should explicitly state the resolution and tiling configuration of the naive baseline used for this 10x calculation in the main text (e.g., in Section 4.4) rather than leaving it to be inferred from Table 6.","section":null},{"comment":"The weakest assumption identified is that the Grad-CAM spatial prior from a ResNet-34 classifier provides meaningful guidance for the allocation policy. Table 5 shows that report-only supervision yields B-4 of 0.246, while adding policy supervision (including Grad-CAM) yields 0.253. This suggests the Grad-CAM prior is load-bearing. However, the manuscript does not evaluate the quality of these Grad-CAM maps or their alignment with the subtle lesions that high-resolution processing is designed to capture. If the classifier's saliency maps are poorly aligned with report-relevant regions, the policy could be misguided. The authors should provide a qualitative or quantitative analysis of the Grad-CAM prior quality, or at least discuss this risk in Section 3.3.","section":null}],"minor_comments":[{"comment":"Section 3.1, Eq. (1): The notation uses non-standard characters (e.g., 'F“ϕ imgpXq PR NvˆDv'). This appears to be a rendering issue, but it should be corrected to standard LaTeX/math notation for clarity.","section":null},{"comment":"Section 3.3, Eq. (6): The Grad-CAM formula uses 'Ca' and 'H0W0' which are not defined in the surrounding text. Please define these variables (e.g., number of channels, spatial dimensions).","section":null},{"comment":"Table 2: The 'Ours' row reports BLEU-4 of 0.138, which is significantly lower than the BLEU-4 reported in Table 1 for MIMIC-CXR (0.253). While this is likely due to the different dataset (CheXpertPlus) and benchmark setup (CXPMRG-Bench), a brief footnote or note in the table caption explaining the discrepancy would aid the reader.","section":null},{"comment":"Figure 2: The text in the figure is quite small and difficult to read. Consider enlarging the font sizes or simplifying the diagram for better legibility in the final version.","section":null},{"comment":"Section 4.4, Table 5: The row 'GRF + LSRA (report-only)' reports B-4 of 0.246, which is lower than 'GRF + Grad-CAM' (0.251). This suggests that LSRA with report-only supervision performs worse than a simple Grad-CAM-based selection. The authors should clarify whether LSRA without policy supervision is detrimental or simply unhelpful compared to Grad-CAM.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern regarding the missing low-res patch control is valid and is the most significant issue with the paper. However, given that the paper's primary contribution is the efficient allocation framework (LSRA + GRF) and the strong empirical results across multiple benchmarks, I believe this can be addressed with an additional ablation and discussion rather than requiring a full revision. The parameter size confound in Table 1 is also worth noting to the authors, as it affects the strength of the claimed improvements over larger models."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The core idea here is genuinely useful: instead of uniform tiling for high-res chest X-rays, learn where to allocate high-resolution capacity and write those features back into the global grid without inflating token count. The LSRA + GRF combination is a clean formulation of a real bottleneck, and the token-preserving write-back is the right design choice. Results across MIMIC-CXR, IU-Xray, and CheXpertPlus are strong, and the ablation in Table 5 does a decent job isolating component contributions. The framework is self-contained with no circular dependencies — Grad-CAM comes from an external ResNet-34, benchmarks are standard. Credit where due: the formulation as a constrained spatial resolution allocation problem is novel for RRG and the engineering is clean. The stress-test concern about a missing low-res patch control is the real issue. Table 5 shows GRF with random high-res patches gives +0.006 B-4, but this confounds two factors: the cross-attention block adds capacity, and the high-res patches add information. Without running GRF on low-res patches from the same selected regions, the paper cannot distinguish 'high-resolution perception helps' from 'an extra attention block helps.' Table 6 partially addresses this (Ours-1024 at 0.245 vs Ours-1920 at 0.253), but both use GRF+LSRA so the resolution effect is still entangled with patch content differences. This is the one experiment that would solidify the central claim. Secondary issues are minor: Table 1 mixes model sizes (4B vs 7B vs 14.2B) without normalization, the '10x fewer tokens' claim is only against naive tiling, and lambda/d_min values are unspecified. None of these are dealbreakers. The Grad-CAM prior quality concern is reasonable but Table 5 shows report-only supervision (0.246) vs report+policy (0.253), so the prior helps but isn't load-bearing — the model learns something useful even without it. This paper is for researchers working on medical vision-language models who care about the resolution-efficiency tradeoff. It deserves a serious referee who can ask for the missing control experiment. If the authors add that ablation and the high-res patches still win, the paper is solid.","headline":"Solid high-res RRG framework with one missing ablation control","tokens_in":18390,"tokens_out":533,"would_cite":true,"duration_ms":73894,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"AI Sees X-ray Details Without Paying the Token Tax","keywords":["radiology report generation","high-resolution perception","visual token budget","spatial resolution allocation","chest X-ray","multimodal large language model"],"falsifier":"If replacing Grad-CAM priors with random spatial priors during training yields equivalent final performance, the contribution of the learnable allocation policy is indistinguishable from the fusion mechanism alone.","tokens_in":17618,"feed_emoji":"","tokens_out":929,"duration_ms":130724,"temperature":0.7,"pith_summary":"Radiology report generation models typically downsample chest X-rays to low resolutions (e.g., 256×256), losing subtle pathological cues like faint nodules or fine vascular markings. Naively increasing resolution by tiling the image into patches inflates the number of visual tokens the language model must process, creating prohibitive computational cost. This paper proposes LePaX, a framework that mimics how radiologists work: first surveying the whole image globally to find suspicious regions, then zooming into those specific areas at high resolution. LePaX learns a spatial utility map that predicts which regions deserve high-resolution processing, extracts a small number of high-res patches (up to 1920×1920) from those locations, and then fuses the detailed patch features back into the global feature grid at their original spatial coordinates. This write-back mechanism enriches the global representation with fine-grained diagnostic evidence without adding any new tokens to the visual budget. The result is a system that processes native-resolution details using the same fixed token count as a low-resolution baseline, achieving consistent improvements across linguistic and clinical metrics on three benchmarks.","feed_headline":"","feed_subtitle":"","key_machinery":"Learnable Spatial Resolution Allocation (LSRA) and Global-Regional Fusion (GRF)","core_discovery":"The central mechanism is a two-stage process where high-resolution perception is formulated as a constrained spatial resolution allocation problem. A lightweight policy head (LSRA) learns to predict which grid locations on a low-resolution global feature map carry the most diagnostic utility, then extracts high-resolution patches from those locations in the native image. A fusion module (GRF) performs spatially grounded resolution write-back, replacing the corresponding tokens in the global feature grid with enriched versions that incorporate high-resolution regional evidence via cross-attention. This allows the model to see fine-grained details in diagnostically relevant regions while the L","pith_inferences":["The Grad-CAM prior serves as a training stabilizer for the allocation policy, but the ablation showing report-only supervision still yields B-4 of 0.246 (vs. 0.253 with the prior) suggests the policy can learn meaningful allocation from the generation objective alone, raising the question of whether stronger generation signals could eventually replace external classifier priors entirely.","The fixed token budget constraint is a design choice tied to pretrained encoder architectures; if future vision encoders natively support variable-length or hierarchical token sequences, the allocation formulation could shift from selecting discrete patches to learning continuous resolution density functions over the image.","The 10× token reduction claim is relative to naive tiling at the same resolution, which is a weak baseline; a stronger comparison would be against dynamic-resolution approaches that use learned token merging or adaptive tiling, though the paper does not benchmark against those."],"forward_implications":["If the constrained allocation formulation is sound, the same principle could extend to other medical imaging modalities where abnormalities are spatially sparse, such as pathology slides or retinal scans, avoiding the need to process gigapixel images uniformly.","The token-preserving write-back mechanism could be applied to general vision-language tasks where fine-grained detail in specific regions matters (e.g., document understanding, satellite imagery), allowing high-resolution reasoning without quadratic attention costs.","The framework's reliance on weak localization priors (Grad-CAM) during training but not inference suggests a path toward self-supervised region discovery, where the report generation objective alone learns to identify diagnostically relevant areas."],"fun_headline_variants":["High-resolution X-ray report generation without token inflation via lesion-aware patch fus","Native-resolution chest X-ray perception at one-tenth the visual tokens","Lesion-aware patch allocation enables 1920x1920 X-ray reporting under fixed token budgets","Targeted high-resolution patches improve chest X-ray report quality without extra tokens","Spatial utility maps route high-resolution capacity to diagnostically relevant X-ray regio"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The framework assumes that Grad-CAM heatmaps from a pretrained disease classifier provide meaningful spatial guidance for where high-resolution patches should be extracted during training. If the classifier's saliency maps are poorly aligned with the subtle lesions that high-resolution processing is designed to capture, the allocation policy could be guided toward classifier-salient but report-irrelevant regions.","fun_headline_variants_meta":{"raw":{"variants":["High-resolution X-ray report generation without token inflation via lesion-aware patch fusion","Native-resolution chest X-ray perception at one-tenth the visual tokens","Lesion-aware patch allocation enables 1920x1920 X-ray reporting under fixed token budgets","Targeted high-resolution patches improve chest X-ray report quality without extra tokens","Spatial utility maps route high-resolution capacity to diagnostically relevant X-ray regions"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":709,"prompt_tokens":609,"completion_tokens":100,"prompt_tokens_details":null},"tokens_in":609,"tokens_out":100,"duration_ms":27753,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T23:07:41.385424+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If replacing Grad-CAM priors with random spatial priors during training yields equivalent final performance, the contribution of the learnable allocation policy is indistinguishable from the fusion mechanism alone.","supporting_citations":[],"review_version":1}