{"id":"cdef73aa-1880-4ca6-b67b-67227d066ea0","arxiv_id":"2506.03926","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MIST adapts CLIP with two Gaussian-sampled prompts per class and improves few-shot accuracy under extreme domain shift on four benchmarks.","lead":"MIST adapts CLIP to extreme domain shifts with only a few labeled examples by giving each class two randomizable prompts. It reports average accuracy gains over strong prompt-tuning baselines on four satellite, medical, and agricultural benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inference-time architecture is underspecified: once text prompts become class-specific (Sec. 4.2), the paper never states whether the visual prompts from Sec. 3.1 (θv=fϕ(θt)) are class-specific or shared; this blocks reproduction of Table 2.","rationale":"The reader's weakest assumption concerned whether two stochastic text prototypes per class suffice when a class has more than two modes. That is a meaningful modeling question, but the paper already probes the number of prompts in Fig. 5 and finds two to be empirically best, so the assumption is at least partially supported by the authors' own ablation. The more load-bearing issue for the central empirical claim is that the method's architecture is not fully specified at inference. The base network couples text and visual prompts through θv = fϕ(θt), while the proposed MIST introduces two class-specific text prompts. The paper does not state whether this coupling makes the visual prompts class-specific as well. If it does, inference is circular because the image encoder would need the class label to produce the image embedding used to classify that same image; if it does not, the paper omits which text prompt is the input to fϕ. This is not a minor implementation detail: it determines whether MIST is a standard shared-visual-encoder few-shot classifier or a much stronger class-conditioned encoder that runs many forward passes per test image. Without code or an explicit protocol, Table 2 cannot be reproduced, so the reported gains remain unverified. This supports the reader's CONDITIONAL verdict rather than changing it; the condition is precisely the release of code or a precise statement of the inference-time architecture. I therefore mark agreement as partial: the reader identified the need for a precise inference protocol, but the specific load-bearing weakness I found is the unresolved text-to-visual prompt coupling, not the number of modes per class.","tokens_in":11661,"tokens_out":17371,"duration_ms":168349,"concrete_test":"Ask the authors for the official code and inspect the inference function. If unavailable, implement both readings: (A) shared visual prompt θv = fϕ(θ_global) for all classes; (B) class-specific visual prompts, using for each test image the visual prompt of each candidate class (C forward passes) and then max logit. Run 1-shot and 8-shot on EuroSAT and ISIC under both readings. If (B) is the only setting that reproduces the reported 77.90 and 34.40 1-shot numbers, the method uses class information in the image encoder and the comparison is not a standard few-shot classifier; if (A) reproduces them, the paper must specify which text prompt generates θv. Either outcome settles whether Table 2 is a valid, reproducible result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sections 3.1 and 4.2 together leave the image-encoding path undefined. Sec. 3.1 defines the base network with one text prompt θt and one visual prompt θv = fϕ(θt). Sec. 4.2 then replaces the text prompt by two class-specific stochastic prompts θ1_t, θ2_t per class (Eq. 2), used to build 2C text embeddings. The paper never states what happens to θv. If θv is also class-specific because fϕ maps the class-specific θ_i_t, then a test image's visual embedding z_v in Eq. 3 cannot be computed without knowing its class: the method would need C or 2C image-encoder forward passes per test image, or it would use the class label at inference, leaking information. If θv remains shared, the paper must specify which of the two class-specific prompt sets (or some other global prompt) is passed through fϕ; no such prompt is mentioned. The figure and text are ambiguous, and no code is provided. Since the headline result is the accuracy table, this ambiguity is load-bearing: a reimplementation cannot decide what to run, and a class-conditioned image encoder would be a different, much stronger method than the baselines compared in Table 2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MIST, a few-shot CLIP adaptation method for target datasets with large visual appearance and label-semantic shifts, in a non-episodic setting where all classes are present. MIST augments a multimodal deep-prompting base network (MaPLe-style) with two stochastic text prompts per class: one sampled from a Gaussian with fixed mean 'A photo' and learnable variance, and one sampled from a fully learnable Gaussian. Visual prompts are obtained from text prompts through a projection layer. Training uses a per-image loss that assigns the image to the closest of its two class prompts and a centroid-similarity regularizer. Experiments on the BSCDFSL benchmark (EuroSAT, ISIC, Plant Disease, ChestX) at 1, 2, 4, 8, and 16 shots report average accuracy gains over CoOp, TaskRes, MaPLe, PromptSRC, CLAP, and TCP, plus a class-imbalance study, variance analysis, and ablations.","tokens_in":11984,"tokens_out":11514,"duration_ms":102644,"significance":"If the claims are correct, MIST addresses a practical gap: adapting CLIP to non-episodic few-shot settings with extreme domain and semantic shift while all classes are handled at once. The combination of multiple prompts per class and stochastic prompt sampling is a sensible way to regularize scarce data, and the paper provides useful ablations (Table 5), a class-imbalance experiment (Table 3), and variance analysis (Table 4). The reported average gains in Table 2 are internally consistent with the stated margins. However, two load-bearing points are unresolved: the class-specific visual-prompt path is underspecified, and the regularizer claimed to enforce prompt diversity does not appear to implement that goal. These issues must be resolved before the empirical claims can be accepted.","major_comments":[{"comment":"The visual-prompt pathway is undefined once text prompts become class-specific. Section 3.1 defines the visual prompts as θv = fφ(θt) for a single text prompt. In Section 4.2, the text prompt is replaced by two class-specific prompts θi_t per class in Eq. (2), but the paper never states whether θv is class-specific or shared. If θv is class-specific, the image embedding z̃v used in Eqs. (3)-(4) cannot be computed for a test image without knowing its class; the method would need 2C image-encoder forward passes per test image or it would leak the label at inference. If θv is shared, the paper does not say which prompt is passed through fφ. The text and Figure 3 are ambiguous and no code is provided. This blocks reproduction of Table 2 and must be resolved by specifying the exact training and inference image-encoding path.","section":"Sec. 3.1 and Sec. 4.2, Eq. (2)"},{"comment":"The regularizer Lreg does not enforce diversity between the two prompts. Minimizing -sim(z̃v, (z̃1_t+z̃2_t)/2) encourages the image embedding to be close to the centroid of the two class text embeddings, but it contains no term that penalizes similarity between z̃1_t and z̃2_t. If the two prompts collapse to the same embedding, the centroid is that embedding and the regularizer simply rewards closeness to it; it imposes no penalty on the collapse. For an image close to one prompt, moving the other prompt toward that image tends to increase sim(z̃v, centroid), so Eq. (5) can actively encourage collapse. The stated purpose 'prevents the two prompts from collapsing' is therefore not implemented by the given objective, and the ablation in Table 5 does not isolate a diversity effect. Please add an explicit repulsion/diversity term or reframe what Lreg actually does.","section":"Sec. 4.2, Eq. (5)"},{"comment":"The claim of superiority is stated too broadly. The 'consistent average gains' in Section 5.1 are average gains, not per-dataset wins: Table 2 shows MIST below the best baseline in several cells, e.g., 1-shot PDisease (50.27 vs MaPLe 51.53), 1-shot ChestX (17.10 vs CoOp 22.83), 2-shot ISIC (36.37 vs TCP 36.87), 4-shot ChestX (18.67 vs MaPLe 19.73), and 16-shot ChestX (14.77 vs CoOp 18.80). The Limitations section states 'MIST outperforms state-of-the-art methods across all datasets' and then immediately says its performance 'slightly drops on the grayscale ChestX dataset,' which is internally inconsistent. Please qualify the claims to refer to average performance and acknowledge the per-dataset exceptions.","section":"Sec. 5.1 and Limitations"},{"comment":"The inference-time sampling protocol is underspecified. The paper says that at inference 'we can sample weights' for the two text prompts and 'take the maximum logit among the two text prompts,' but it does not state how many samples are drawn per class or whether predictions are averaged over multiple samples. Because the test-time text embeddings are random draws from learned Gaussians, the reported accuracies in Table 2 are not reproducible without fixing a seed, a sample count, or an averaging rule. Please specify the exact inference protocol and, ideally, report the variance due to inference sampling.","section":"Sec. 4.2, Inference"}],"minor_comments":[{"comment":"Equation (7) writes the expectation over D_tgt, but only the support set is available for training; please write the empirical expectation over the support set S.","section":"Sec. 3, Eq. (7)"},{"comment":"There is a typo: 'descrie' should be 'describe.' Also, 'upto' in the implementation details should be 'up to.'","section":"Sec. 4, introductory paragraph"},{"comment":"The phrase 'take the maximum logit among the two text prompts' is ambiguous because each text prompt is stochastic; please define the class score explicitly in terms of the sampled text embeddings.","section":"Sec. 4.2, Inference"},{"comment":"Figure 5 would be easier to interpret with numeric x-axis values and error bars; the statement that accuracy 'decreases after a point' is not quantitative.","section":"Sec. 5.2 and Figure 5"},{"comment":"Variance across seeds is reported only for EuroSAT and ISIC; please also report it for PDisease and ChestX, or justify why those datasets are omitted.","section":"Sec. 5.2, Table 4"},{"comment":"The related-work discussion mentions Bayesian prompt learning [5] but does not clarify how MIST differs from it in the stochastic-prompt design; a brief comparison would help position the contribution.","section":"Sec. 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful incremental contribution to few-shot CLIP adaptation under large domain shifts, and the results are plausible, but the paper leaves a load-bearing ambiguity in the inference architecture that needs to be fixed before the numbers can be fully trusted.\n\nWhat's actually new: MIST combines two existing ideas — per-class multiple prompts (from mixture-prototype literature) and stochastic Gaussian prompt sampling (Bayesian prompt learning) — in a non-episodic, all-class setting, and reports consistent average gains over MaPLe/PromptSRC/TCP on the BSCDFSL benchmark. The Table 5 ablations support each component, and the lower variance in Table 4 suggests the gains are not just lucky seeds. That is honest progress.\n\nThe soft spots, in order of severity.\n\nFirst, the visual-prompt path is underspecified. Section 3.1 defines θv = fϕ(θt) with a single text prompt. Section 4.2 then gives each class two text prompts (θ1_t, θ2_t), but never says what happens to θv. If visual prompts are class-specific, computing the image embedding z~v in Eq. 3 requires knowing the class at inference, which leaks the label and multiplies encoder forward passes by C. If they are shared, the paper does not say which text prompt is passed through fϕ. The figure caption does not resolve this. This is not a nitpick: a reimplementation literally does not know what model to run. The stress-test concern holds up.\n\nSecond, the 'outperforms across all datasets' claim contradicts Table 2. On ChestX at 16 shots, MIST (14.77) is below CoOp (18.80), CLAP (17.47), TCP (16.53), and PromptSRC (14.83). The Limitations paragraph acknowledges a drop on ChestX, but the phrasing is still wrong.\n\nThird, the inference protocol for stochastic prompts is underspecified: do you sample once or average over multiple samples at test time? That affects variance and reproducibility.\n\nFourth, the closest baselines — Bayesian Prompt Learning [5] and Prompt Distribution Learning [15] — are cited but never compared directly, which would have clarified the novelty.\n\nThe two-prototype-per-class assumption is a reasonable design choice, and the attempt to justify it with t-SNE is okay, though it is not deeply tested.\n\nWho this is for: people working on parameter-efficient CLIP adaptation in specialized domains like medical or satellite imagery. It does not open a new paradigm, but it is a credible increment.\n\nRecommendation: yes, send it to peer review. Require the authors to either release code or precisely specify the visual-prompt generation at inference. Without that, the central table is not independently reproducible, and that is a blocker for acceptance as-is.","headline":"A plausible incremental advance in few-shot CLIP adaptation, but an underspecified visual-prompt inference path currently blocks independent reproduction of the headline results.","tokens_in":12464,"tokens_out":4689,"would_cite":true,"duration_ms":40754,"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":"MIST claims that two stochastic text prompts per class—one anchored to the frozen 'A photo' prompt and one fully learned—let a frozen CLIP adapt to extreme visual and semantic shifts with only a few labeled examples from all classes at…","keywords":["few-shot learning","prompt tuning","vision-language models","CLIP adaptation","extreme domain shift","stochastic prompts","multimodal prompts","cross-domain few-shot learning"],"falsifier":"Construct or find a target class whose embedding space splits into three well-separated visual subgroups; if MIST with its fixed two prompts per class cannot match a three-prompt variant on that class, the two-prototype capacity assumption is falsified.","tokens_in":11473,"feed_emoji":"🖼️","tokens_out":8912,"duration_ms":80583,"temperature":0.7,"pith_summary":"Most CLIP adaptation methods assume that a class's images form one cluster in feature space, but satellite, medical, and plant-disease images fragment each class into several modes and push different classes together. MIST replaces the single text prompt per class with two prompts sampled from Gaussians—one anchored to the frozen 'A photo' text prompt and one fully learnable—and trains them end-to-end while CLIP's encoders stay frozen. The paper's claim is that this lets one model handle the full target label set at once with one to sixteen labeled examples per class, a setting where episodic cross-domain few-shot methods do not apply. If the claim holds, a large pretrained image-text model can be specialized to extreme domain shifts with a handful of labels, with the stochastic sampling acting as the regularization that prevents overfitting.","feed_headline":"Two stochastic prompts per class beat CLIP few-shot baselines","feed_subtitle":"Frozen CLIP adapts to satellite, medical, and plant-disease images using just 1-16 labels per class.","key_machinery":"The machinery is a pair of stochastic text prompts per class inside a deep multimodal prompt-tuning adapter. For each class, prompt one is sampled as $\\theta^1_t \\sim \\mathcal{N}(\\bar{\\mu}_0, \\sigma_0)$ with $\\bar{\\mu}_0$ fixed to the 'A photo' embedding and $\\sigma_0$ learnable, while prompt two is sampled as $\\theta^2_t \\sim \\mathcal{N}(\\mu_1, \\sigma_1)$ with both parameters learnable. The text prompts pass through a projection $f_\\phi$ to produce visual prompts, and the reparameterization trick $\\theta = \\mu + \\mathcal{N}(0, I) \\odot \\sigma$ lets gradients reach the distribution parameters. Training minimizes $\\mathcal{L}_{mp}$ (assigning each image to its closest text prototype) plus $\\mathcal{L}_{reg}$ (pulling the image toward the centroid of its class's two prompts), and inference takes the maximum logit over the two prompts for each class.","core_discovery":"The paper claims that the degraded few-shot performance of prompt-tuned CLIP under extreme domain and semantic shift comes mainly from modeling each class with a single text prompt, and that two stochastic text prototypes per class remove that limitation. MIST keeps the CLIP encoders frozen and optimizes only the Gaussian prompt parameters plus a projection layer, with the first prompt sampled from a Gaussian whose mean is fixed to the pretrained 'A photo' embedding and the second sampled from a fully learnable Gaussian. Each image is assigned to the nearer of its class's two text embeddings during training, and a regularization term pulls the image toward the centroid of the two, stopping the prompts from collapsing onto one mode. On the BSCDFSL benchmark the authors report average gains of 3.19%, 2.01%, 1.79%, 3.19%, and 1.60% over the best prior methods at 1, 2, 4, 8, and 16 shots, and lower variance across seeds.","pith_inferences":["Editorial inference: if two prompts per class are the reason for the gains, a data-dependent number of prompts chosen by clustering the support-set embeddings could push accuracy further; the paper itself shows performance falls after three prompts, so the fixed choice of two may leave modes uncaptured in some classes.","Editorial inference: the combination of one anchored prompt and one fully learned prompt is a general regularization recipe that could transfer to adapter tuning or to other frozen foundation models, not just CLIP text prompts.","Editorial inference: since MIST trains without a validation set and without episodes, it could be paired with validation-free model selection to address the separate problem of choosing hyperparameters in few-shot adaptation, a direction this paper does not explore."],"forward_implications":["The largest reported gain comes at 1-shot, supporting the paper's claim that sampling around a fixed pretrained prompt anchors CLIP to its prior knowledge and suppresses overfitting when each class has a single image.","Because all classes are trained together rather than in sampled episodes, the method applies directly to deployment settings where the complete target label set is known at once.","The class-imbalanced experiment and the worst-class-bin analysis imply that the two-stochastic-prompt design improves coverage of difficult and visually similar classes, not just average accuracy.","The acknowledged weaker result on grayscale chest X-rays indicates that visual prompts can hurt when the shift is largely a change in low-level image statistics, so the method is not uniformly superior across all domain shifts."],"supporting_citations":[{"why":"Supplies the frozen CLIP model whose text and image encoders MIST adapts.","marker":"[19]"},{"why":"Defines the BSCDFSL benchmark with its four domain-shifted datasets used for all evaluations.","marker":"[7]"},{"why":"Provides the deep multimodal prompt-tuning base network that MIST builds on and the strongest baseline to beat.","marker":"[11]"},{"why":"Gives the reparameterization trick that lets gradients flow through sampled stochastic prompt weights.","marker":"[13]"},{"why":"Motivates the multiple-prototypes-per-class idea that MIST transfers into prompt space.","marker":"[2]"},{"why":"Supplies the mixture-prototype few-shot formulation and the collapse problem MIST's regularization term addresses.","marker":"[1]"},{"why":"Documents prompt-tuning overfitting in low-shot settings and supplies the self-regulating baseline compared throughout.","marker":"[12]"}],"fun_headline_variants":["Two stochastic prompts per class fix CLIP under domain shift","Stochastic prompt pairs adapt CLIP to extreme shifts","MIST: Gaussian prompts per class beat CLIP few-shot","Double stochastic prompts improve CLIP few-shot adaptation","Per-class Gaussian prompts win under extreme domain shift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Two stochastic text prototypes per class are enough to cover the visual modes a class can have under extreme domain shift, and the few labeled support images reveal those modes.","fun_headline_variants_meta":{"raw":{"variants":["Two stochastic prompts per class fix CLIP under domain shift","Stochastic prompt pairs adapt CLIP to extreme shifts","MIST: Gaussian prompts per class beat CLIP few-shot","Double stochastic prompts improve CLIP few-shot adaptation","Per-class Gaussian prompts win under extreme domain shift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1457,"prompt_tokens":972,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":588,"tokens_out":485,"duration_ms":5128,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:52:41.737606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or find a target class whose embedding space splits into three well-separated visual subgroups; if MIST with its fixed two prompts per class cannot match a three-prompt variant on that class, the two-prototype capacity assumption is falsified.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural lan- guage supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP model whose text and image encoders MIST adapts."},{"cited_title":"A broader study of cross-domain few- shot learning","cited_arxiv_id":null,"evidence_quote":"Defines the BSCDFSL benchmark with its four domain-shifted datasets used for all evaluations."},{"cited_title":"Maple: Multi-modal prompt learning","cited_arxiv_id":null,"evidence_quote":"Provides the deep multimodal prompt-tuning base network that MIST builds on and the strongest baseline to beat."},{"cited_title":"Infinite mixture prototypes for few-shot learning","cited_arxiv_id":null,"evidence_quote":"Motivates the multiple-prototypes-per-class idea that MIST transfers into prompt space."},{"cited_title":"Mixture-based feature space learning for few-shot image classification","cited_arxiv_id":null,"evidence_quote":"Supplies the mixture-prototype few-shot formulation and the collapse problem MIST's regularization term addresses."},{"cited_title":"Self-regulating prompts: Founda- tional model adaptation without forgetting","cited_arxiv_id":null,"evidence_quote":"Documents prompt-tuning overfitting in low-shot settings and supplies the self-regulating baseline compared throughout."}],"review_version":1}