{"id":"23fcfc10-e20b-4917-97bd-0a71fda398a0","arxiv_id":"2508.04101","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"NEARL-CLIP reports top accuracy on three medical image classification benchmarks by adding bidirectional cross-modal attention and orthogonalized adapters to CLIP with only 1.46M parameters.","lead":"A new CLIP adaptation method, NEARL-CLIP, uses tiny cross-modal attention modules plus an orthogonal projection step to improve medical image classification with only 1.46 million trainable parameters. It reports top accuracy on pneumonia, Alzheimer, and retina binary tasks, beating prompt learning and one-way interaction baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is measured on binary subsets with the hardest classes discarded; the reported generalization to medical vision-language understanding is not supported.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: the empirical contribution is confined to selectively binarized versions of public benchmarks. The paper's own Sec. IV-A states the class-dropping and balancing procedures, so this is not an external critique but an internal scope limitation. The central claim, as written in the abstract and conclusion, is that NEARL-CLIP achieves state-of-the-art medical vision-language understanding across three modalities; that claim requires either full multi-class evaluation or an explicit delimitation of the claim to binary triage tasks. The internal ablations and parameter-efficiency evidence are useful and keep the method plausible, but they do not repair the external-validity gap. A concrete multi-class re-run with paired significance testing would settle whether the concern actually lands. I therefore agree with the reader's conditional verdict: the paper should be accepted with the condition that the SOTA claim is either demonstrated on the full tasks or explicitly narrowed.","tokens_in":10751,"tokens_out":3899,"duration_ms":52157,"concrete_test":"Re-run the full comparison, including NEARL-CLIP and all Table I baselines, on the original multi-class splits: 4-class Alzheimer (Non-demented / Very mild / Mild / Moderate) and 4-class Retina (CNV / DME / DRUSEN / NORMAL), using macro-F1 as the primary metric, with the same training protocol and at least 5 seeds. If NEARL-CLIP remains best after paired significance testing, the overreach concern is resolved; if the margin shrinks below one standard deviation or reverses, the SOTA claim should be restricted to the binary subsets.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim rests on Sec. IV-A's dataset construction. Alzheimer is reduced to a binary task by discarding 'Very mild dementia' and randomly subsampling Non-demented to balance; Retina is reduced to DME vs NORMAL, dropping CNV and DRUSEN. These are not arbitrary omissions: the dropped classes are the diagnostically adjacent gradations where a medical VLM's text-guided distinctions are most stressed. Table I therefore demonstrates state-of-the-art on two curated binary subsets, not on the medical diagnosis problems the abstract and conclusion claim ('state-of-the-art on three medical imaging modalities', 'medical vision-language understanding'). Because all comparison methods are re-run on the same subsets, the internal ranking may be fair, but the headline claim overreaches. The paper also lacks a multi-class experiment or a justification for why the discarded classes are out of scope, and the quoted 'up to 2.1% improvement' on Pneumonia comes with a baseline std of ±3.2 (MaPLe), so even the headline gain is not shown to be statistically reliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NEARL-CLIP, a parameter-efficient adaptation of CLIP for medical image classification. The method introduces two modules: USEformer, which performs bidirectional cross-modal interaction via learnable queries and cross-attention, and OCA, an adapter that applies Gram-Schmidt orthogonalization to project adaptation features onto the orthogonal complement of pre-trained features, claimed to separate 'truly novel' from 'incremental' knowledge. The model adds 1.46M learnable parameters and is evaluated on three medical datasets (Pneumonia, Alzheimer, Retina), reporting state-of-the-art accuracy and F1 against prompt-learning and unidirectional-interaction baselines. Ablations on Pneumonia show contributions of each module and sensitivity to hyperparameters M and r.","tokens_in":11115,"tokens_out":2965,"duration_ms":35807,"significance":"If the empirical claims hold, NEARL-CLIP offers a compact and effective recipe for adapting CLIP to medical domains, with a plausible mechanism (bidirectional interaction and orthogonal regularization) and strong parameter efficiency. The paper's strengths include a clean architecture, consistent internal ablations, and a clear attempt to preserve pre-trained generalization while enabling adaptation. However, the external validity of the headline SOTA claim is substantially weakened by the evaluation protocol: two of the three datasets are reduced to binary subsets by discarding diagnostically relevant classes, and no statistical significance testing is reported. The claimed conceptual distinction between 'truly novel' and 'incremental' knowledge is also not directly evidenced. These issues are fixable within the manuscript's scope, but they currently limit the strength of the conclusions.","major_comments":[{"comment":"The evaluation protocol reduces multi-class medical datasets to binary subsets: Alzheimer discards 'Very mild dementia' and randomly subsamples Non-demented to balance; Retina keeps only DME vs NORMAL, dropping CNV and DRUSEN. These are not arbitrary omissions: the discarded classes are diagnostic gradations where text-guided distinctions are most stressed. Table I therefore supports SOTA on curated binary subsets, not on the medical diagnosis tasks named in the abstract and conclusion. Please either evaluate on the full multi-class tasks, or provide a principled justification for excluding classes and add multi-class results to substantiate the 'medical vision-language understanding' claim.","section":"Sec. IV-A, Table I"},{"comment":"The headline improvement on Pneumonia is 2.1% over MaPLe (94.7±0.2 vs 92.6±3.2), but MaPLe's standard deviation is 3.2 and only three seeds are averaged. No significance tests are reported for any comparison or ablation. The claim of state-of-the-art performance is therefore not statistically established. Report pairwise significance tests (e.g., corrected t-test or bootstrap) or confidence intervals/effect sizes for the main comparisons and for the ablations in Tables II-IV.","section":"Sec. IV-B, Table I"},{"comment":"The implementation details list architecture hyperparameters (M, Nq, Dq, r, tau, epoch) but omit the optimizer, learning rate, weight decay, batch size, input resolution, augmentations, and the exact prompt template. No code is released. Since all comparison methods were re-implemented (as stated in Table I), these details are necessary for reproducibility and for assessing whether the comparison is fair. Please provide the full training protocol.","section":"Sec. IV-A, Implementation Details"},{"comment":"Equation (8) is a fixed Gram-Schmidt projection of Δf onto the orthogonal complement of f. The paper's conceptual claim that this operation 'decouples new knowledge into truly novel and incremental components' is an interpretation, not something demonstrated by the experiments. The ablation shows that orthogonal regularization helps empirically, but it does not validate that the projected component corresponds to semantically 'truly novel' information. Please soften the interpretation or provide a direct analysis of the projected components (e.g., norms, similarity to pre-trained features, or downstream effect of each component) to support the claim.","section":"Sec. III-C, Eq. (8)"}],"minor_comments":[{"comment":"Typo: 'NROMAL' should be 'NORMAL'.","section":"Sec. IV-A, Retina"},{"comment":"The abstract says 'three medical imaging modalities (X-ray, MRI)' but the Retina dataset is OCT; the conclusion says '(X-ray, MRI, etc.)'. Please correct the modality list for consistency.","section":"Abstract / Sec. V"},{"comment":"The softmax expressions appear to have unmatched parentheses; clarify the numerator and denominator.","section":"Sec. III-B, Eqs. (4)-(5)"},{"comment":"The notation 'The text encoder T processes prompts T' uses T for both the encoder and the prompt set, which is confusing. Consider different symbols.","section":"Sec. III-A"},{"comment":"Seeds are averaged but no seed values or random split generation details are given; please specify how the three seeds were used.","section":"Sec. IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically coherent and the internal ablations are consistent, but the evaluation protocol's selective dataset construction and lack of statistical rigor currently prevent the stated SOTA and 'medical vision-language understanding' claims from being fully supported. The missing training details and absence of code further hinder reproducibility. These issues are addressable, so major revision rather than rejection seems appropriate. The conceptual framing of 'truly novel' knowledge is somewhat marketing-like; tightening it would improve credibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2508.04101.\n\nThe paper's real contribution is a parameter-efficient adapter for CLIP that combines bidirectional cross-modal attention (USEformer) with a fixed Gram-Schmidt orthogonalization step (OCA). Both ideas are known in isolation, but the combination is new and the ablations are internally consistent. The authors report 1.46M learnable parameters and show sensible gains over prompt-learning and one-way interaction baselines on their chosen benchmarks. The ablations support both components, and the layer-wise interaction analysis is a nice touch.\n\nThat said, the central empirical claim overreaches. The SOTA results are obtained on binary subsets constructed by discarding the diagnostically interesting classes. Alzheimer drops 'Very mild dementia' and balances 'Non-demented'; Retina keeps only DME vs. NORMAL, dropping CNV and DRUSEN. These are not arbitrary omissions; they are the gradations that stress a medical VLM. So Table I shows SOTA on two curated binary tasks, not on the medical diagnosis problems the abstract and conclusion invoke. The internal ranking may be fair, but the headline 'state-of-the-art on three medical imaging modalities' is not supported. Also, the Pneumonia gain is 2.1% with a MaPLe baseline std of ±3.2, so the gain is not shown to be reliable. The paper also omits optimizer, learning rate, augmentation, batch size, and code, which makes reproduction hard. And a small thing: the abstract says 'X-ray, MRI' but Retina is OCT.\n\nOn the positive side, the architecture is plausible, the math is simple and correct (Eq. 8 is a fixed projection), and the parameter efficiency is real. The 'truly novel vs incremental' framing is post hoc and somewhat marketing-flavored, but the orthogonalization itself is a reasonable inductive bias.\n\nWho's this for? Researchers working on parameter-efficient adaptation of CLIP to medical imaging. They'll find the combination useful, but they should be cautious about the claimed generalization. I would not cite it as evidence of SOTA on multi-class medical diagnosis.\n\nI'd send it to review, with the expectation that the authors either add multi-class evaluations or tone down the claims, release code and training details, and address the statistical reliability. A good referee could turn this into a solid contribution.","headline":"Sensible adapter combination with consistent ablations, but the SOTA claim rests on cherry-picked binary subsets and missing training details.","tokens_in":11486,"tokens_out":2176,"would_cite":false,"duration_ms":23451,"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":"NEARL-CLIP claims that bidirectionally querying both modalities — image querying text and text querying image — plus orthogonal decoupling of adaptation increments, adapts CLIP to medical images with state-of-the-art accuracy on Pneumonia,","keywords":["medical vision-language understanding","CLIP adaptation","parameter-efficient fine-tuning","cross-modal interaction","orthogonal regularization","prompt learning","medical image classification","modality alignment"],"falsifier":"Run NEARL-CLIP and the best unidirectional baselines on the full multiclass versions of the same datasets — all four Alzheimer classes (including Very mild dementia) and all four OCT classes (CNV, DME, DRUSEN, NORMAL). If the accuracy gain over MaPLe or XCoOp shrinks or reverses once the discarded, harder classes are included, the central claim that bidirectional interaction resolves modality misalignment would be weakened.","tokens_in":10707,"feed_emoji":"🩻","tokens_out":5643,"duration_ms":54768,"temperature":0.7,"pith_summary":"NEARL-CLIP aims to show that CLIP's failure on medical images is mainly a modality-misalignment problem, not a capacity problem: adapting only the text branch or only one direction of interaction leaves the frozen image encoder unable to extract meaningful medical features. The paper proposes a bidirectional scheme in which learnable queries from each modality attend to the other's features at every layer (USEformer), and a second module (OCA) that decouples the new knowledge from the pretrained feature space via Gram-Schmidt orthogonalization to prevent interference. The reported results — 94.7% accuracy on chest X-ray Pneumonia, 92.6% on Alzheimer MRI, 98.5% on Retina OCT — beat prompt-learning and unidirectional baselines with only 1.46M trainable parameters, suggesting that mutual enrichment and feature decoupling together close the domain gap.","feed_headline":"Bidirectional CLIP tuning beats one-way prompts on medical scans","feed_subtitle":"Cross-modal queries plus orthogonal decoupling let CLIP adapt to X-ray, MRI, and OCT with 1.46M trainable weights.","key_machinery":"Two coupled modules. USEformer (Unified Synergy Embedding Transformer): M stacked cross-attention blocks; learnable image queries attend to text features and text queries attend to image features (image-to-text and text-to-image attention), producing compact mutual enrichments that are fused back into both branches. OCA (Orthogonal Cross-Attention Adapter): for each layer, fuses the pretrained layer output with the USEformer output via cross-attention, then applies Gram-Schmidt orthogonalization, projecting the increment onto the orthogonal complement of the pretrained feature subspace: $\\Delta f_\\perp = \\Delta f - \\frac{\\langle \\Delta f, f\\rangle}{\\langle f, f\\rangle} f$. This enforces geom","core_discovery":"On its own terms, the paper establishes that dual-modality interaction with orthogonal regularization is a sufficient recipe for parameter-efficient CLIP adaptation to medical imaging. Each layer's frozen features are augmented by a compact cross-modal query pool: image queries summarize text features, text queries summarize image features, and the resulting increments are projected onto the orthogonal complement of the frozen feature subspace before being added. The paper reports best-in-table figures across three datasets of different modalities (X-ray, MRI, OCT), with accuracies of 94.7%, 92.6%, and 98.5% and F1-scores of 94.2%, 92.6%, and 98.2%, and attributes the gains to preventing the","pith_inferences":["Because the benchmarks are binary tasks created by discarding classes (Alzheimer drops 'Very mild dementia'; Retina keeps only DME vs NORMAL), the stated gains may not transfer to multi-class differential diagnosis, where the excluded classes are often the confusing ones.","A direct testable extension: apply the same orthogonal adapter to LoRA-style or prompt-only baselines to see whether the improvement comes specifically from bidirectional querying or from the decoupling regularizer alone.","The orthogonalization step is reminiscent of gradient-projection methods for continual learning; one could test whether NEARL-CLIP also reduces forgetting when adapting sequentially to multiple medical datasets.","Since the paper's layer-group experiment favours mid-level interaction, an adaptive-depth variant (learning where to insert USEformer blocks) might achieve similar accuracy with fewer parameters."],"forward_implications":["If the reported gains hold, CLIP-based medical classifiers can be built by freezing the backbone and training only ~1.5M parameters, which suits low-annotation medical settings.","The superiority over unidirectional baselines on all three datasets implies that one-way adaptation (e.g., image-to-text or text-to-image only) leaves systematic alignment errors that a symmetric querying mechanism can recover.","Orthogonal decoupling is presented as a generalisable component: any PEFT adapter can be post-processed with Gram-Schmidt projection to decorrelate new knowledge from pretrained features.","The framework's layer-wise analysis suggests mid-level layers carry the most useful cross-modal signal, implying future methods might concentrate interaction where it matters rather than across all layers."],"supporting_citations":[{"why":"Supplies the frozen CLIP backbone that NEARL-CLIP adapts.","marker":"[10]"},{"why":"CoOp, prompt-learning baseline adapting only text prompts; NEARL aims to beat it.","marker":"[1]"},{"why":"FATE, unidirectional modality-interaction baseline that NEARL compares against and argues is insufficient.","marker":"[2]"},{"why":"XCoOp, explainable medical prompt-learning baseline used for comparison.","marker":"[11]"},{"why":"ViP, medical prompt-learning baseline using LLM-generated visual descriptors; NEARL compares against it.","marker":"[12]"},{"why":"CoCoOp, unidirectional baseline conditioning prompts on image features; NEARL compares against it.","marker":"[14]"},{"why":"MaPLe, unidirectional baseline transferring text prompts to image prompts; NEARL compares against it.","marker":"[15]"},{"why":"Kermany dataset source for Pneumonia and Retina; used in evaluations.","marker":"[17]"},{"why":"OASIS dataset source for Alzheimer; used in evaluations.","marker":"[33]"},{"why":"LoRA baseline used in the ablation study; NEARL-CLIP is compared against it when both modules are removed.","marker":"[34]"}],"fun_headline_variants":["Two-way query tuning helps CLIP on X-ray, MRI, OCT","Orthogonal cross-attention adapts CLIP to medical scans","CLIP learns medical imaging via bidirectional queries","Parameter-efficient CLIP with orthogonal decoupling for scans","Bidirectional CLIP adaptation beats one-way prompts"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The evaluation rests on binary classification datasets created by discarding classes (e.g., Alzheimer's 'Very mild dementia' and Retina's CNV/DRUSEN classes), so the claim that bidirectional interaction closes the medical domain gap is only tested on the easiest split of each task.","fun_headline_variants_meta":{"raw":{"variants":["Two-way query tuning helps CLIP on X-ray, MRI, OCT","Orthogonal cross-attention adapts CLIP to medical scans","CLIP learns medical imaging via bidirectional queries","Parameter-efficient CLIP with orthogonal decoupling for scans","Bidirectional CLIP adaptation beats one-way prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":2912,"prompt_tokens":819,"completion_tokens":2093,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2014}},"tokens_in":563,"tokens_out":2093,"duration_ms":17982,"temperature":1.0,"reasoning_tokens":2014,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:52:19.600030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run NEARL-CLIP and the best unidirectional baselines on the full multiclass versions of the same datasets — all four Alzheimer classes (including Very mild dementia) and all four OCT classes (CNV, DME, DRUSEN, NORMAL). If the accuracy gain over MaPLe or XCoOp shrinks or reverses once the discarded, harder classes are included, the central claim that bidirectional interaction resolves modality misalignment would be weakened.","supporting_citations":[{"cited_title":"Learning to prompt for vision- language models,","cited_arxiv_id":null,"evidence_quote":"CoOp, prompt-learning baseline adapting only text prompts; NEARL aims to beat it."},{"cited_title":"Fate: Feature-adapted parameter tuning for vision-language models,","cited_arxiv_id":null,"evidence_quote":"FATE, unidirectional modality-interaction baseline that NEARL compares against and argues is insufficient."},{"cited_title":"Xcoop: Explainable prompt learning for computer-aided diagnosis via concept-guided context op- timization,","cited_arxiv_id":null,"evidence_quote":"XCoOp, explainable medical prompt-learning baseline used for comparison."},{"cited_title":"Aligning medical images with general knowledge from large language models,","cited_arxiv_id":null,"evidence_quote":"ViP, medical prompt-learning baseline using LLM-generated visual descriptors; NEARL compares against it."},{"cited_title":"Conditional prompt learning for vision-language models,","cited_arxiv_id":null,"evidence_quote":"CoCoOp, unidirectional baseline conditioning prompts on image features; NEARL compares against it."},{"cited_title":"Maple: Multi-modal prompt learning,","cited_arxiv_id":null,"evidence_quote":"MaPLe, unidirectional baseline transferring text prompts to image prompts; NEARL compares against it."},{"cited_title":"Identifying medical diagnoses and treatable diseases by image-based deep learning,","cited_arxiv_id":null,"evidence_quote":"Kermany dataset source for Pneumonia and Retina; used in evaluations."},{"cited_title":"Open access series of imaging studies (oasis): Cross- sectional mri data in young, middle aged, nondemented, and demented older adults,","cited_arxiv_id":null,"evidence_quote":"OASIS dataset source for Alzheimer; used in evaluations."}],"review_version":1}