{"id":"86032b4e-f98c-4ac5-84e4-ec357ee985ee","arxiv_id":"2509.09397","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DRiFt explicitly decouples clinical from class-agnostic features in medical vision-language models and reports improved few-shot accuracy, but robustness under domain shift is not consistently supported.","lead":"This paper introduces DRiFt, a method that splits medical image and text features into clinical and spurious parts using LoRA and prompts, and tests it on 8 medical imaging tasks. The authors report higher few-shot accuracy than two prompt-based baselines, but the robustness claim is weakened by near-random AUC on many tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The decoupling mechanism is unverifiable: phi_M is omitted from the parameter set, so the claimed cause of the +11.4% gain is not established.","rationale":"The reader's primary weakest assumption — that the phi_M projections are not included in trainable parameters and their architecture is unspecified — is exactly the load-bearing concern I identify. Section 2.1 defines only z_{M i,u} = phi_M(z_{M i}), and the formulas for spurious components are absent. All three losses in Section 2.2 operate on these projected embeddings, so the decoupling mechanism depends entirely on phi_M. Yet the trainable parameter set Θ=θ∪γ excludes phi_M. If phi_M is fixed, the split cannot adapt to the few-shot task; if it is trainable, it is an undocumented part of the method. Either way, the paper does not establish that the reported +11.4% accuracy gain comes from the claimed decoupling rather than from LoRA, prompt tuning, or caption curation. I also note the lack of multi-seed evaluation, but the missing phi_M is more fundamental because it blocks verification of the method's central mechanism. I agree with the reader's REJECT verdict: the current text does not allow a reader to reproduce or validate the stated approach. A targeted code inspection and ablation could settle the question, and a revised submission with explicit phi_M and error bars would merit reconsideration.","tokens_in":8673,"tokens_out":5292,"duration_ms":67255,"concrete_test":"Obtain the released code (github.com/rumaima/DRiFt) and locate the definition of phi_M for vision and text. Then run the Table 1 protocol under three conditions: (1) as released; (2) with phi_M replaced by a fixed random orthogonal projection (or identity if dimensions match); (3) with the split removed entirely (L_sp^r and L_con dropped). If conditions (2)/(3) differ from (1) by less than ~2 points average Top-1, the decoupling losses are not the driver of the reported 11.4-point gain; if they differ substantially, the paper must specify phi_M before the claim can be assessed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DRiFt's gains come from structured decoupling is not supported because the projection functions phi_M that define the invariant/spurious split are never specified. Section 2.1 gives z_{M i,u} = phi_M(z_{M i}) and calls it \"two projection functions\", but no formula for the spurious component z_{M i,s}, no architecture (linear/MLP/random), and no trainable parameters for phi_M are given. The overall trainable parameter set is Θ=θ∪γ (LoRA adapters plus prompt tokens), so if phi_M is frozen the invariant/spurious split is fixed before training and the losses L_ce^u, L_sp^r, L_con in Section 2.2 cannot adaptively enforce decoupling; if phi_M is learned, it is missing from Θ and from the implementation details. Either way, the mechanism advertised as the cause of the +11.4% in-distribution improvement is unverifiable. The single 16-sample draw without error bars compounds this, but the missing projection is the more fundamental gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DRiFt is a parameter-efficient few-shot adaptation framework for medical vision-language models under distribution shift. It decomposes image and text embeddings into invariant (clinically relevant) and spurious components via unspecified projection functions, fine-tunes only LoRA adapters and prompt tokens, and trains with a combination of invariant contrastive alignment, KL-based spurious neutralization, and conditional-independence (HSIC) losses. The authors add InstructBLIP-generated captions to MedIMeta and evaluate on eight in-distribution tasks and cross-dataset transfer against MaPLe-IC and CoOp-OOD-IC. The abstract reports +11.4% Top-1 and +3.3% Macro-F1 over prior prompt-based methods while maintaining robustness across unseen datasets.","tokens_in":8980,"tokens_out":5575,"duration_ms":62099,"significance":"The intended contribution—explicit cross-modal decoupling of clinical and class-agnostic features with parameter-efficient tuning—is timely and, if valid, would be useful for few-shot medical VLM adaptation. The paper has concrete strengths: code is released, two prompt-based baselines are adapted to the image-caption setting, and ablations cover shots, prompt depth, prompt tokens, and backbone. However, the central decoupling mechanism is not actually specified, the HSIC estimator and loss weights are absent, and the reported in-distribution AUC is at chance for most tasks. As a result, the paper's core claims—both the mechanism and the robustness of the gains—are not currently established.","major_comments":[{"comment":"The decomposition is undefined. The paper states only z_{M i,u}=phi_M(zMi) and calls these 'two projection functions'; it neither defines the spurious component z_{M i,s}, nor specifies the architecture of phi_M (linear/MLP/random), nor includes phi_M in the trainable set Theta=theta∪gamma. If phi_M is frozen, the invariant/spurious split is fixed before training and Eqs. (1)-(3) cannot adaptively enforce decoupling; if phi_M is learned, it is omitted from the implementation. Either way, the advertised cause of the +11.4% gain is unverifiable.","section":"Section 2.1"},{"comment":"The conditional-independence losses L_conv and L_cont in Eq. (2) are written as HSIC terms, but no estimator, kernel, or conditioning procedure is given. The paper does not report how I(z_v;z_s|Y) is computed in practice, so Eq. (2) and the overall loss L in Eq. (3) are not reproducible. This is a load-bearing component of the claimed decoupling.","section":"Section 2.2"},{"comment":"The loss weights alpha and beta in Eq. (3) are never reported, and the implementation details do not describe a validation split for choosing them, the LoRA rank r, or the caption-generation thresholds. Ablations vary prompt depth and token count without indicating where selection occurs, raising a risk of test-set-based selection. Reproducibility and the interpretation of the reported gains are therefore limited.","section":"Section 2.2 and 3.1"},{"comment":"AUC is at or below chance for most tasks (bus 34.9, skinl_derm 49.5, derm 46.5, glaucoma 51.2, pneumonia 52.4, mammo_mass 45.9; average 49.1 with chance 50), directly contradicting the abstract's 'strong robustness' and 'reliable' claims. In addition, all results come from a single 16-sample-per-class draw with no error bars or significance tests. With few-shot sample sizes, the Top-1 improvements may be within run-to-run noise.","section":"Table 1"},{"comment":"The cross-dataset generalization claim is not supported. DRiFt is best only for skinl_derm (10.4 vs 6.4-6.5) and is worse on derm (17.8 vs 18.8-18.9) while roughly tying on glaucoma and fundus. No average or statistical comparison is reported, so 'maintaining strong robustness across unseen datasets' is unsubstantiated. Moreover, the method extends prior work [9], but [9] is not among the baselines, leaving the incremental contribution unclear.","section":"Table 2 and Section 3.2"}],"minor_comments":[{"comment":"Typo: 'howvere' should be 'however'. Also 'spuriousrelationshipswhichcanbeverysensitive' is missing spaces.","section":"Section 3.1"},{"comment":"Notation is inconsistent: Eq. (2) defines L_conv and L_cont, but Eq. (3) introduces L_con = avg(L_conv, L_cont); the names should be unified.","section":"Section 2.2"},{"comment":"The reference to 'Figure 4(left)' appears in Section 2.2, before the experiments are described; this should be moved to the results discussion.","section":"Section 2.2"},{"comment":"The pneumonia row for DRiFt (37.5% Acc, 18.2% Macro-F1) is identical to CoOp-OOD-IC, yet the text says 'pneumonia accuracy improves'; the comparison baseline should be stated precisely.","section":"Table 1 / Section 3.2"},{"comment":"The shot-ablation text says bus 'increases' from 50% to 42.3% (decrease); the sentence is internally contradictory and should be corrected.","section":"Section 4a"}],"recommendation":"reject","confidential_remarks":"The core issues are not merely presentational: the decoupling mechanism is undefined, key implementation details (HSIC estimator, alpha/beta, validation procedure) are missing, and the reported AUC is at chance for most tasks. These problems cannot be fixed by small edits; a revised manuscript would need to re-specify the method and supply new experiments with multiple seeds and validation-based hyperparameter selection. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper applies a decoupling idea from the authors' own DIMPLE to medical images with InstructBLIP captions. That is new, and the evaluation on 8 MedIMeta tasks plus four external sets is useful. But the central mechanism is under-specified: the projection functions phi_M that split embeddings into invariant and spurious parts are never defined as trainable or fixed, no architecture is given, and the overall parameter set Theta = theta U gamma excludes them. So the advertised cause of the +11.4% in-distribution gain is not verifiable. This is the main problem.\n\nWhat the paper does well: the problem is real—medical VLMs pack spurious correlations—and the caption curation step is a sensible ingredient. The tables show consistent average gains over two adapted prompt baselines (MaPLe-IC and CoOp-OOD-IC), and the ablations on prompt tokens and depth are informative. There is a code link.\n\nWhere it falls short, in proportion:\n\n1. Missing specification. The paper says 'two projection functions are defined by z_{M i,u} = phi_M(z_{M i})' but gives no formula, no trainable flag, no layer type, no input dimension. If phi_M is frozen, the split is arbitrary and the KL and HSIC losses cannot enforce the decoupling story; if learned, it is missing from Theta and from the implementation details. Either way the paper does not support its own causal claim. Related: alpha and beta are never reported, and the HSIC estimators are not described.\n\n2. No error bars. The experiments use 16 random samples per class, one draw, and no standard deviation. In a low-shot regime that is not enough to make +11.4% credible, especially when Table 1 shows several AUC values near chance.\n\n3. Robustness claim overstates. Table 2 shows DRiFt better on one skin-lesion transfer, roughly equal on fundus and TB, and slightly worse on derm. That is not 'strong robustness across unseen datasets.'\n\n4. No comparison to DIMPLE [9]. Since the framework is an extension of the authors' own prior work, omitting it makes the incremental contribution hard to judge.\n\nThese are fixable, not fatal. The idea is plausible and the domain shift setup is relevant. My own verdict is skeptical, but I would not desk-reject this: the authors need to add multi-seed runs, report all hyperparameters, specify phi_M, and compare against DIMPLE. As it stands, the paper is a recipe with an unverified mechanism.","headline":"A plausible few-shot medical VLM adaptation recipe whose core decoupling mechanism is underspecified and whose robustness claim rests on a single seed; worth refereeing, but only with heavy revisions.","tokens_in":9431,"tokens_out":2968,"would_cite":false,"duration_ms":32887,"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":"DRiFt splits medical image and text features into clinical and spurious components, and this explicit decoupling lifts few-shot accuracy and robustness under distribution shift.","keywords":["medical vision-language models","few-shot learning","distribution shift","feature decoupling","spurious correlations","low-rank adaptation","prompt learning","out-of-distribution generalization"],"falsifier":"Take the same pipeline but replace the projection functions with a fixed random linear map that does not separate clinical content. If average Top-1 accuracy remains near 45.8%, the decoupling losses are not doing the claimed work; if it falls sharply, the projection choice is the critical ingredient. Alternatively, rerun the 16-shot setting with ten different random seeds and check whether the +11.4 average advantage over the baseline persists.","tokens_in":8586,"feed_emoji":"🩺","tokens_out":7150,"duration_ms":69032,"temperature":0.7,"pith_summary":"The paper is trying to establish that medical vision-language models can be made more reliable in few-shot settings by explicitly separating clinically relevant, invariant information from task-agnostic spurious cues in both image and text. It proposes DRiFt, which combines low-rank adapters with learnable prompt tokens and uses three objectives: aligning invariant image-text features, neutralizing spurious features by pushing their class predictions toward uniform, and reducing statistical dependence between invariant and spurious features given the label. The headline numbers are a +11.4% Top-1 accuracy and +3.3% Macro-F1 improvement over prior prompt-based methods on in-distribution medical tasks, with competitive cross-dataset transfer. A sympathetic reader would care because this is a parameter-efficient path (no full fine-tuning) toward suppressing the spurious correlations that make clinical VLMs unsafe under domain shift.","feed_headline":"Splitting clinical from spurious cues lifts few-shot accuracy 11.4%","feed_subtitle":"DRiFt keeps clinical image-text cues separate from domain noise, improving generalization to new imaging settings.","key_machinery":"The feature-decoupling step is the machinery. A projection function phi_M acts on each raw embedding z_M (M for vision or text) and splits it into z_{M,u} (invariant, clinically relevant) and z_{M,s} (spurious). The three losses are then targeted: the contrastive loss aligns only invariant embeddings across modalities; the KL-divergence loss drives predictions from spurious embeddings toward a uniform class distribution so they carry no classification signal; and conditional independence losses reduce the mutual information between invariant and spurious embedding parts given the label. This separation is what lets the model learn clinical cues without being misled by domain-specific artifac","core_discovery":"The central claim is that spurious associations, such as a clothing shadow or incidental report wording, can be suppressed if each embedding is decomposed into invariant and spurious parts and the training losses act on those parts separately. DRiFt projects image and text embeddings into two subspaces, aligns only the invariant subspace with a contrastive loss, pushes the spurious subspace's predictions toward a uniform distribution, and penalizes conditional dependence between the two subspaces given the label. Trained with 16 samples per class, low-rank adapters, and learnable prompts, the framework reports an average Top-1 accuracy of 45.8% across eight medical tasks, compared with 34.4%","pith_inferences":["Because the projection functions that split embeddings are not trained and their architecture is not specified, the decoupling effect may be sensitive to how those projections are chosen; a direct test is to make them learnable or compare several fixed choices.","The few-shot numbers come from a single 16-sample-per-class split with no seeds or error bars, so the +11.4 point gap might not survive resampling; reporting multiple splits would separate mechanism from sampling luck.","The caption-generation step relies on an instruction-tuned captioning model; if caption quality degrades for rare diseases, the invariant alignment could reinforce wrong text-image associations, making caption validation a hidden dependency.","The same loss design could transfer to other high-stakes domains with known spurious cues, such as pathology or retinal OCT, where domain shift follows similar patterns, but this extrapolation is not tested in the paper."],"forward_implications":["If DRiFt's claim holds, then few-shot adaptation of medical vision-language models can be done without full fine-tuning, since only low-rank adapters and prompt tokens are updated.","Explicit spurious-feature neutralization can be added to any prompt-based method as a loss term, potentially improving out-of-distribution behavior.","The decoupling works on both modalities, so degradation from text-side spurious cues (e.g., incidental findings in reports) is also addressed.","Cross-dataset transfer suggests that models trained on one imaging protocol can preserve accuracy on another, which matters for multi-site deployment.","Using two prompt tokens rather than one nearly doubled average accuracy in the experiments, so prompt capacity itself appears to be part of the mechanism."],"fun_headline_variants":["DRiFt decouples invariant from spurious features, lifts few-shot accuracy","Separating clinical cues from spurious noise boosts few-shot Top-1","DRiFt splits subspaces to hit 45.8% few-shot accuracy, up from 34.4%","Decoupling spurious signals yields 11.4-point few-shot accuracy gain","DRiFt aligns invariant cues, disrupts spurious ones for better few-shot"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fixed projection functions used to split embeddings into invariant and spurious parts are meaningful; if those projections are arbitrary or poorly matched to the data, the three losses cannot actually enforce decoupling and the reported gains would come from another mechanism.","fun_headline_variants_meta":{"raw":{"variants":["DRiFt decouples invariant from spurious features, lifts few-shot accuracy","Separating clinical cues from spurious noise boosts few-shot Top-1","DRiFt splits subspaces to hit 45.8% few-shot accuracy, up from 34.4%","Decoupling spurious signals yields 11.4-point few-shot accuracy gain","DRiFt aligns invariant cues, disrupts spurious ones for better few-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3056,"prompt_tokens":734,"completion_tokens":2322,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2221}},"tokens_in":478,"tokens_out":2322,"duration_ms":18556,"temperature":1.0,"reasoning_tokens":2221,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:08:12.783331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same pipeline but replace the projection functions with a fixed random linear map that does not separate clinical content. If average Top-1 accuracy remains near 45.8%, the decoupling losses are not doing the claimed work; if it falls sharply, the projection choice is the critical ingredient. Alternatively, rerun the 16-shot setting with ten different random seeds and check whether the +11.4 average advantage over the baseline persists.","supporting_citations":[],"review_version":1}