{"id":"319b23b9-d30f-4496-a8b2-520feffe6227","arxiv_id":"2502.08391","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-scale vision-language MIL framework with LLM-generated descriptive prompts and prototype-guided feature aggregation improves few-shot whole slide image classification.","lead":"ViLa-MIL adds LLM-generated, dual-scale descriptive text prompts to a frozen CLIP model for classifying whole slide images from limited labels. It reports gains of 1.7 to 7.2 percent AUC over MIL baselines on three cancer subtyping datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing TOP baseline leaves the SOTA claim unsubstantiated; a head-to-head comparison is needed.","rationale":"The reader's verdict (CONDITIONAL) is appropriate, but their stated weakest assumption (LLM text accuracy) is not the most load-bearing. The paper already partially addresses text accuracy: three pathologists examined the prompts for the two cancer types used, and Table 6 shows robustness across GPT-3.5/GPT-4/PaLM-2/LLaMA-2. The omitted TOP baseline is a direct threat to the comparative claim and is easily testable. It also explains why the reader's rationale (which mentions TOP) is stronger than their formal weakest_assumption. Therefore I agree with the CONDITIONAL verdict but identify a different load-bearing concern.","tokens_in":18252,"tokens_out":6856,"duration_ms":71663,"concrete_test":"Run TOP [30] under the exact 16-shot protocol of Table 1: same training/validation/test splits, same frozen CLIP ResNet-50 image encoder, same 5x/10x patches (or TOP's native single scale), and same metrics (AUC/F1/ACC) across TIHD-RCC, TCGA-RCC, and TCGA-Lung, using the authors' released code. Compare mean and std over five seeds with a paired t-test. If TOP's mean AUC/F1 are within the reported 1.7–7.2% gap or better, the SOTA claim in the abstract is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that viLa-MIL outperforms state-of-the-art methods by 1.7–7.2% AUC and 2.1–7.3% F1 under 16-shot. Table 1 compares against nine MIL baselines, but the most directly related VLM-MIL method, TOP [30], is discussed in Related Works ('two-level prompt learning MIL framework') and never evaluated. TOP is specifically designed for few-shot WSI classification with language priors, so it is the natural SOTA comparator for viLa-MIL. Without it, the 'state-of-the-art' claim is unsubstantiated: the observed gains may be due to comparing against methods that do not use language priors at all. The paper also omits MI-Zero and other VLM-based WSI methods from the main table, though PLIP/QuiltNet appear only as backbone replacements in Supplementary L. This is a load-bearing omission because the headline claim is explicitly comparative and the missing baseline is the closest in spirit. If TOP matches or exceeds viLa-MIL under the same protocol, the central claim collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ViLa-MIL, a multiple instance learning framework for few-shot whole slide image classification that combines a frozen CLIP backbone with (i) dual-scale visual descriptive text prompts generated by a frozen LLM (GPT-3.5), (ii) a prototype-guided patch decoder that aggregates large sets of patch features into learnable prototypes, and (iii) a context-guided text decoder that refines text features using patch and prototype features. The model is trained end-to-end with a cross-entropy loss on slide labels. Experiments on three WSI subtyping datasets (TIHD-RCC, TCGA-RCC, TCGA-Lung) under a 16-shot setting report AUC/F1/ACC improvements over several MIL baselines, plus ablation studies, cross-dataset generalization, and interpretability visualizations.","tokens_in":18567,"tokens_out":5414,"duration_ms":51551,"significance":"If the reported gains hold, the paper makes a useful contribution by showing that LLM-generated, scale-conditioned descriptive prompts can transfer CLIP to pathology in a parameter-efficient way, reducing the need for large collections of pathology image-text pairs. The prototype-guided aggregation is a sensible way to adapt CLIP to gigapixel images, and the cross-center evaluation is a strength. The paper ships code and reports multiple runs with standard deviations. However, the central claim of state-of-the-art performance is presently supported only against MIL baselines that do not use language priors, and the most directly related VLM-MIL method (TOP) is not evaluated, which tempers the significance until that comparison is added.","major_comments":[{"comment":"The main comparison omits TOP, a two-level prompt learning MIL framework for few-shot weakly-supervised WSI classification that is the most closely related VLM-MIL method. Without a head-to-head comparison under the same protocol, the claim that ViLa-MIL outperforms state-of-the-art by 1.7–7.2% AUC is not substantiated for VLM-based methods. Please add TOP (and, if feasible, MI-Zero [24]) to Table 1 and the cross-dataset table, or revise the claim to refer only to MIL baselines.","section":"Section 4.2 and Table 1; Section 2 (TOP [30])"},{"comment":"Hyperparameters Np and M are selected on the TIHD-RCC dataset, which is also one of the three datasets used for the main evaluation. The manuscript does not state whether this selection used the validation split only. To avoid optimistic bias on TIHD-RCC, please clarify the model-selection protocol and, if the test split was used, report results retrained with hyperparameters fixed on a validation set for all datasets.","section":"Supplementary K and Section 4.1"},{"comment":"The gain from adding the context-guided text decoder to the dual-scale model is 83.6±2.7 to 84.3±4.6 AUC on TIHD-RCC; with five runs and overlapping standard deviations, this difference is not established as significant. Please report paired significance tests (or a confidence interval) for the ablation steps, and temper the corresponding sentence in Section 4.4.","section":"Table 3"},{"comment":"The 'Large Language Model' baseline uses a single-scale GPT-3.5 prompt, while ViLa-MIL uses dual-scale prompts plus two new decoders. This confounds the contribution of the dual-scale text prompt with the architectural changes. Please add an ablation that uses the dual-scale prompt with ABMIL (no prototype or text decoders) to isolate the effect of the text prompt design.","section":"Table 5 and Section 4.4"}],"minor_comments":[{"comment":"Please define Ql, Kl, Vl explicitly before the equation; currently the notation is introduced only in the surrounding prose.","section":"Section 3.3, Eq. (4)"},{"comment":"Please clarify whether α_l and α_h are learned or fixed (the implementation says both are 1); if fixed, note it in the text.","section":"Section 3.5, Eq. (9)"},{"comment":"The superscript * is described as denoting comparable performance based on a paired t-test, but it is not stated to which reference the asterisked methods are compared; please clarify.","section":"Table 1 caption"},{"comment":"Please specify the exact GPT-3.5 model version used for reproducibility (e.g., text-davinci-003 or gpt-3.5-turbo).","section":"Section 4.1"},{"comment":"The phrase 'to obtain the prediction result' should read 'to obtain the visualization result' to match the context of threshold-based visualization.","section":"Supplementary F"},{"comment":"The claim that GPT-4 achieves superior performances over GPT-3.5 is not true for F1 (68.0 vs 68.7); please report precisely.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The main gap is the missing TOP comparison, which is directly relevant to the SOTA claim and should be fixable in a revision. Also, the use of TIHD-RCC for hyperparameter tuning and for reporting a headline result should be clarified. The core method seems sound and the paper is otherwise well organized."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid, well-written few-shot WSI classification paper, and the genuinely new bit is using an LLM to generate separate low- and high-scale visual descriptions as text prompts. That is a simple and practical way to inject pathological prior without extra pretraining. The prototype-guided patch decoder and context-guided text decoder are also sensible, parameter-light choices, and the experiments cover three datasets with cross-center evaluation plus a code release. The training protocol is standard supervised few-shot, so there is no circularity issue.\n\nThe soft spots are real but not fatal. The biggest one is the missing TOP baseline. TOP is the closest VLM-MIL method in spirit, and it is discussed in Related Works but never compared. Without that head-to-head, the headline claim of 1.7–7.2% AUC gains could just be the advantage of using language priors at all. That is easily fixable, and it needs to be fixed before the SOTA claim is credible. Second, the main ablations run only on the in-house TIHD-RCC set, and some incremental gains (e.g., 83.6 → 84.3 AUC) are within one standard deviation, so the added value of the context-guided text decoder is not strongly evidenced. The prompt validation is also thin: three pathologists, two cancer types, no quantitative check on whether the descriptions are actually discriminative.\n\nNone of this undermines the core idea. The method is coherent, the writing is honest, and the stress-test worry that the whole paper collapses without TOP is overstated — the framework would still be a useful contribution even if TOP matches it. But the paper as-is overclaims, and a serious revision should add TOP, run the key ablations on at least one TCGA dataset, and report effect sizes with confidence intervals.\n\nI'd send this to peer review rather than desk-reject. It deserves referee time, and the fixes are concrete. My own verdict would be conditional acceptance after those additions.","headline":"A credible few-shot WSI method with a neat dual-scale LLM-prompt idea, but the missing TOP baseline keeps the state-of-the-art claim unsubstantiated.","tokens_in":18969,"tokens_out":1853,"would_cite":true,"duration_ms":22261,"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":"LLM-written dual-scale pathology descriptions let a frozen vision-language model beat few-shot whole-slide classifiers.","keywords":["whole slide image classification","multiple instance learning","vision-language model","CLIP","few-shot learning","text prompt","prototype-guided aggregation","digital pathology"],"falsifier":"Retrain ViLa-MIL with the generated descriptive prompts replaced by class-name-replacement templates while keeping all decoders, data, and training identical; if the AUC margin over the best MIL baseline does not shrink substantially, the descriptive prompt is not the operative component. A complementary check is to reuse the same frozen prompts on slides from a hospital not in the training set and watch whether the cross-center AUC advantage persists.","tokens_in":1842,"feed_emoji":"🔬","tokens_out":2206,"duration_ms":97179,"temperature":0.7,"pith_summary":"ViLa-MIL proposes that a vision-language model can classify whole-slide images under extreme data scarcity if the text prompt carries the diagnostic reasoning pathologists actually use. The paper constructs dual-scale visual descriptive prompts that describe each cancer subtype at low resolution in terms of tissue architecture and at high resolution in terms of cellular detail, generated by a frozen large language model and prefixed with learnable vectors. These prompts guide a frozen CLIP backbone, while a prototype-guided patch decoder aggregates gigapixel patches into slide-level features and a context-guided text decoder refines the text features with image context. On three multi-cancer, multi-center subtyping datasets with only 16 labeled slides per class, the paper reports gains of 1.7-7.2% in AUC and 2.1-7.3% in F1 over state-of-the-art MIL baselines. If correct, this would mean pathology language priors can be injected without collecting millions of image-text pairs or retraining a foundation model.","feed_headline":"LLM-written pathology descriptions beat few-shot slide classifiers","feed_subtitle":"Dual-scale text prompts carry diagnostic prior, outrunning MIL baselines by up to 7.2 AUC points.","key_machinery":"The mechanism has three interacting pieces. First, the dual-scale visual descriptive text prompt: a frozen large language model is asked to describe each class at low and high resolution, yielding two text strings per class, one about architecture-level morphology for low magnification and one about cytologic detail for high magnification, with $M$ learnable context vectors prepended to each string. Second, the prototype-guided patch decoder: a set of learnable prototype vectors serves as queries in a cross-attention layer over patch features, $\\mathrm{Softmax}(QK^\\top/\\sqrt{d})V$ with a residual connection and layer normalization, so patches sharing semantics are grouped into the same prototype, and an attention pooling over the prototypes produces the slide-level feature. Third, the context-guided text decoder: the text features are used as queries in another cross-attention layer whose keys and values are the concatenation of local patch features and global prototype features, pulling image context into the language features. Training uses cross-entropy on the sum of per-scale similarity logits, and inference is the same forward pass. The decoders are lightweight and trainable while the vision-language encoders and the language model stay frozen, which is what makes the transfer parameter-efficient.","core_discovery":"The paper's central claim is that the weak link in vision-language whole-slide classification is the text prompt, not the encoder: replacing class-name templates with dual-scale descriptive texts that encode routine pathological diagnostic reasoning unlocks the few-shot regime. On three multi-cancer, multi-center subtyping tasks with 16 labeled slides per class, the proposed framework reports state-of-the-art results, improving AUC by 1.7-7.2% and F1 by 2.1-7.3% over existing MIL methods, and holding a 5.5% AUC advantage in cross-center transfer between two renal cell carcinoma datasets. The discovery, stated on the paper's own terms, is that a frozen LLM can write the pathology knowledge into the prompt, and two lightweight decoders can adapt a frozen vision-language model to gigapixel slides without any image-text pre-training on pathology data.","pith_inferences":["If the descriptive prompts are the main driver, then prompt quality should be directly tunable: editing the generated texts with a pathologist or using a stronger LLM should yield further few-shot gains without touching the decoders, a cheap experiment on the same three datasets.","The same dual-scale design could transfer to other gigapixel-image domains, such as satellite or whole-brain imaging, whenever a domain expert can specify what morphology appears at low versus high resolution, since no image-text pre-training pairs are needed.","Because the method computes per-scale similarities and sums their logits, the scale weights could be made learnable per task, so a task whose diagnostic features live at one magnification would automatically down-weight the other scale.","A failure mode worth testing externally is the rare-disease case: if the LLM has never seen the cancer subtype, its generated descriptions may be generic, and a test on a rare subtype would reveal how much of the gain depends on the LLM's prior knowledge."],"forward_implications":["With 16 labeled slides per class, the method outperforms the best MIL baselines by 1.7-7.2% in AUC and 2.1-7.3% in F1 across three multi-cancer, multi-center subtyping datasets.","In cross-center transfer between two renal cell carcinoma datasets, the model keeps a 5.5% AUC advantage over the best baseline, indicating that the language prior transfers to unseen data distributions.","Ablations show each component contributes: the prototype-guided decoder improves over attention pooling, the dual-scale prompt adds over either single scale, and the context-guided text decoder adds further gains.","The prompt generation is robust to the choice of LLM: several popular large language models all produce prompts that beat class-name-replacement templates.","Swapping the frozen CLIP backbone for pathology-specific vision-language encoders raises AUC by about one point, so the method improves further as better vision-language encoders become available."],"supporting_citations":[{"why":"Supplies the frozen large language model that generates the dual-scale visual descriptive text prompts.","marker":"[1]"},{"why":"Provides the attention-based MIL aggregation baseline that the prototype-guided patch decoder replaces and outperforms.","marker":"[14]"},{"why":"Provides a dual-stream multi-scale MIL baseline that motivates the use of patches from two magnifications.","marker":"[16]"},{"why":"Serves as the strongest weakly-supervised MIL comparison baseline in the few-shot evaluation.","marker":"[25]"},{"why":"Is the previous prompt-learning MIL framework that ViLa-MIL extends by adding dual-scale descriptive prompts.","marker":"[30]"},{"why":"Supplies the frozen CLIP backbone whose image and text encoders define the shared embedding space.","marker":"[31]"},{"why":"Provides a transformer-based MIL baseline that models inter-patch correlations and is used as a comparison method.","marker":"[34]"},{"why":"Contributes the learnable context-vector prompt technique that ViLa-MIL prepends to the descriptive texts.","marker":"[49]"}],"fun_headline_variants":["LLM-written pathology descriptions boost slide classification","Few-shot slide AI gains from frozen LLM text prompts","Dual-scale descriptive text prompts outperform MIL baselines","Pathology-aware prompts unlock gigapixel slide diagnosis","No pathology pretraining: LLM text prompts suffice"],"cache_read_input_tokens":21120,"weakest_assumption_plain":"The entire improvement depends on the generated text descriptions being accurate, class-discriminative, and transferable across institutions, a premise the paper checks with only three pathologists reviewing two cancer types.","fun_headline_variants_meta":{"raw":{"variants":["LLM-written pathology descriptions boost slide classification","Few-shot slide AI gains from frozen LLM text prompts","Dual-scale descriptive text prompts outperform MIL baselines","Pathology-aware prompts unlock gigapixel slide diagnosis","No pathology pretraining: LLM text prompts suffice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3158,"prompt_tokens":981,"completion_tokens":2177,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2102}},"tokens_in":597,"tokens_out":2177,"duration_ms":19351,"temperature":1.0,"reasoning_tokens":2102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T05:14:59.619661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ViLa-MIL with the generated descriptive prompts replaced by class-name-replacement templates while keeping all decoders, data, and training identical; if the AUC margin over the best MIL baseline does not shrink substantially, the descriptive prompt is not the operative component. A complementary check is to reuse the same frozen prompts on slides from a hospital not in the training set and watch whether the cross-center AUC advantage persists.","supporting_citations":[{"cited_title":"Lan- guage models are few-shot learners","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen large language model that generates the dual-scale visual descriptive text prompts."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Provides the attention-based MIL aggregation baseline that the prototype-guided patch decoder replaces and outperforms."},{"cited_title":"Dual-stream multiple instance learning network for whole slide image classifica- tion with self-supervised contrastive learning","cited_arxiv_id":null,"evidence_quote":"Provides a dual-stream multi-scale MIL baseline that motivates the use of patches from two magnifications."},{"cited_title":"Data-efficient and weakly supervised computational pathology on whole- slide images","cited_arxiv_id":null,"evidence_quote":"Serves as the strongest weakly-supervised MIL comparison baseline in the few-shot evaluation."},{"cited_title":"The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Is the previous prompt-learning MIL framework that ViLa-MIL extends by adding dual-scale descriptive prompts."},{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP backbone whose image and text encoders define the shared embedding space."},{"cited_title":"TransMIL: Transformer based correlated multiple instance learning for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Provides a transformer-based MIL baseline that models inter-patch correlations and is used as a comparison method."}],"review_version":1}