acceptodds
Under review as a conference paper at ICLR 2027

FACET: Rethinking Whole-Slide Foundation Models Through Task-Conditioned Views

Abstract

Foundation models (FMs) have demonstrated great success in computational pathology, first at the patch level and more recently at the whole-slide level. Despite being trained on increasingly large WSI corpora and, in some cases, incorporating additional molecular or textual modalities, existing slide-level FMs still produce a single static representation per slide, agnostic to the downstream task. This design is inherently limiting: a whole-slide image contains diverse morphological signals, only a subset of which may be relevant for any given clinical or biological question. We propose FACET, a whole-slide foundation model that conditions slide representations on task descriptions, enabling dynamic inference-time aggregation tailored to each prediction objective. FACET is pretrained through supervised multitask learning across 29 carefully curated prediction tasks spanning six cancer types, aligning task-conditioned slide representations with text-derived class prototypes. We evaluate FACET on an external benchmark comprising 40 distinct tasks, including tasks and cancer types not observed during pretraining. FACET achieves state-of-the-art performance against leading slide-level foundation models and maintains high performance on both unseen tasks and unseen cancer types, demonstrating robust generalization. Together, these results demonstrate task conditioning as a powerful approach for learning generalizable whole-slide representations. We release code, label files, task manifests, and evaluation resources at https://github.com/anonymous-95/FACET.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Foundation models, Whole-slide images, Representation learning, Computational pathology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KxZZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FACET is a slide-level pathology model that uses a text description of the task to guide WSI aggregation. Instead of using one fixed slide representation for all tasks, the same slide can produce different representations for tasks such as MSI prediction, grading, or mutation prediction. The task text is used as a query to decide which patches should receive more attention.

Critical strengths
  1. The motivation is clear. Different pathology tasks may depend on different tissue regions. The paper provides experiments showing that attention maps change across tasks, which supports the use of task-specific pooling.
  2. It provides a shared way to handle different tasks. A standard approach is to train a separate MIL head for each task. FACET instead trains one shared aggregation model and uses task text to control how the slide is pooled. This makes the model more flexible when adding new tasks. Previous work has already used language or pathology concepts for task-specific WSI analysis, but FACET applies this idea directly to a shared slide-level representation model.
  3. The evaluation is broad. The model is tested on external CPTAC cohorts, including cancer types and prediction tasks not used in FACET pretraining. The paper also evaluates linear probing, KNN, finetuning, and survival prediction.
Critical weaknesses
  1. The benefit over task-specific MIL is not fully demonstrated. A simple MIL head trained for each task can already learn task-specific attention and focus on different tissue regions. FACET instead uses one shared aggregator controlled by task text. However, the paper does not directly compare these two settings under the same patch features and downstream data. Since FACET still requires labeled data for new tasks, it is unclear how much extra value the text-conditioned shared aggregator provides over simply training a lightweight MIL head for each task.

  2. The comparison with other foundation models is not fully equal. FACET receives extra task information. Its text description includes the cancer type, target classes, and possible H&E features related to the task. Most baseline models only provide a fixed slide representation. Therefore, part of the improvement may come from the extra task-specific prior, not only from a better visual representation. A useful comparison would be a strong task-specific MIL head or a baseline given the same text information.

  3. Unseen task generalization is still supervised transfer. For new tasks, FACET is evaluated with linear probing, KNN, or finetuning using labeled downstream data. It does not directly solve a new task from the text description alone. Therefore, the result mainly shows that its representation transfers well to new tasks, rather than true zero-shot task prediction.

  4. The claim that a fixed slide representation is inherently limited is stronger than the evidence shows. Different tasks clearly use different attention regions, but this does not mean that one rich slide representation cannot contain information useful for several tasks. The experiments support task-conditioned pooling as a useful design, but they do not prove that fixed slide representations are fundamentally unsuitable.

Questions for authors

See weakness

Reviewer_veQhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FACET, a task-conditioned foundation model for computational pathology whole-slide images (WSIs). The authors argue that existing slide-level foundation models typically compress each WSI into a fixed, task-agnostic representation, which may not sufficiently adapt to diverse downstream objectives. To address this limitation, FACET uses natural-language task descriptions to obtain task embeddings and employs a cross-attention mechanism to guide WSI patch aggregation, enabling task-specific slide representations to be generated at inference time. Experimental results show that FACET outperforms existing slide-level foundation models across multiple evaluation settings. Overall, this work introduces a clinically meaningful research direction. However, the current methodological novelty mainly lies in task-conditioned aggregation, and its fundamental distinction from existing vision-language adaptation and prompt-based approaches requires further clarification.

Critical strengths

1.The paper identifies a potential limitation of current slide-level foundation models: representing an entire WSI with a fixed embedding implicitly assumes that all tasks require the same evidence. This issue is highly relevant in computational pathology. The proposed task-conditioned representation learning framework is intuitive and clinically reasonable. 2.The paper conducts systematic evaluation on 40 external CPTAC tasks. Compared with approaches validated only on a single task or dataset, the scale and comprehensiveness of the experimental evaluation are strong.

Critical weaknesses
  1. The methodological novelty is limited and appears more like introducing task-conditioned attention into an existing framework. The core technical contribution of FACET is mainly adding text-conditioned cross-attention pooling to an existing WSI foundation model pipeline. Although task conditioning is a valuable idea, the overall method is still closely related to text-guided adaptation, prompt-based MIL, and vision-language conditioned aggregation. The paper does not sufficiently clarify the fundamental innovation of FACET compared with existing text-conditioned pathology models or prompt learning methods.

  2. The semantic information contained in task descriptions may be an important source of performance improvement. FACET uses LLM-generated task descriptions and class descriptions as model inputs. However, these texts contain not only morphological information but also class semantics and existing medical knowledge. Therefore, it remains unclear whether the model truly learns better visual representations or simply leverages external medical knowledge contained in the text.

  3. The validation of true open-ended task generalization is insufficient. The paper emphasizes that FACET can generalize to unseen tasks and unseen cancer types. However, the evaluated tasks are still from similar pathology prediction scenarios, and their task definitions are relatively close to the pretraining tasks. It remains unclear whether FACET can support completely new label spaces and morphological concepts unseen during training. For example, whether a newly introduced pathology task after training can obtain an effective representation using only a natural-language task description has not been validated.

  4. The paper lacks sufficient controlled experiments investigating the role of task description information. Although the paper compares GPT-5.4 and other text generation methods, more important control experiments are still missing, such as randomized task descriptions, using only task names/morphology descriptions, and descriptions without class semantic information. Therefore, it is difficult to accurately determine the source of performance improvements.

  5. FACET lacks sufficient comparison with simple task-conditioning approaches. The paper mainly compares with existing foundation models but lacks more direct baselines, such as simple fusion of slide embeddings and task embeddings, using task embeddings as classifier prompts, and task-conditioned MIL pooling. If these simple approaches can achieve similar performance, the additional design value of FACET needs to be reconsidered.

  6. Unlike traditional slide-level foundation models, FACET needs to generate task-conditioned representations for each task. Therefore, when one WSI is used for multiple tasks, whether repeated computation is required, how much additional inference cost is introduced, and whether this affects clinical applications are not sufficiently discussed.

Questions for authors

1.What is the fundamental difference between FACET and existing text-guided adaptation, prompt learning, and vision-language MIL methods? Does the main contribution come from task-conditioned aggregation, multitask pretraining, or text semantic alignment?

2.How much of the performance improvement comes from the medical semantics contained in task descriptions rather than visual representation learning?

3.Can FACET generate effective slide representations for completely unseen tasks after training by only providing task descriptions without additional training?

4.Could the authors include comparisons with simple task-conditioning approaches?

5.Does FACET require re-aggregating WSIs for each task during inference? Compared with traditional slide-level foundation models, what is the additional computational cost?

Reviewer_UwDhOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FACET, a task-conditioned whole-slide foundation model that generates task-specific slide representations using natural-language descriptions. The framework employs a frozen CONCH v1.5 patch encoder and a shared Transformer to extract contextualized patch features, followed by text-guided cross-attention pooling to generate task-dependent slide embeddings. FACET is pretrained through supervised multitask learning on 29 prediction tasks across six TCGA cancer cohorts, aligning slide representations with text-derived class prototypes.

The authors evaluate the model on 40 downstream tasks across nine external CPTAC cohorts, including unseen tasks and cancer types. FACET achieves strong average performance against several slide-level foundation models under linear probing, KNN, fine-tuning, and survival prediction.

Critical strengths
  1. The paper provides a well-motivated perspective on task-adaptive WSI representation learning. The attention overlap and cross-task transfer analyses in Figure 1 provide empirical evidence that different prediction tasks require different morphological information. Task-conditioned aggregation offers a practical way to dynamically generate task-specific representations without retraining the slide encoder for each downstream task.

  2. The evaluation covers a broad range of external datasets and prediction tasks, including unseen cancer types. FACET achieves competitive performance against several established slide-level foundation models despite using a relatively small pretraining dataset. The ablation studies and attention visualizations further support the potential benefits of task-conditioned representation learning.

Critical weaknesses
  1. The baseline comparisons do not sufficiently isolate the contribution of task-conditioned aggregation. Existing slide-level foundation models use different patch encoders, pretraining datasets, and supervision strategies. Moreover, comparisons with conventional patch-level foundation models combined with MIL aggregators are missing. Controlled experiments using the same CONCH v1.5 features and training data, comparing FACET against task-agnostic MIL methods such as ABMIL and TransMIL, would better establish the effectiveness of the proposed approach.

  2. The contribution of semantic task descriptions remains insufficiently validated. Although the learned-query ablation demonstrates the usefulness of text conditioning, it does not establish whether morphology-rich descriptions provide meaningful task-specific guidance beyond task identity. Comparisons with task-name-only descriptions, shuffled descriptions, or simpler task embeddings would strengthen this claim.

Questions for authors
  1. Have the authors compared FACET against conventional MIL models, such as ABMIL or TransMIL, using the same frozen CONCH v1.5 patch features and the same downstream evaluation splits? If such results are available, how do they compare with FACET?

  2. Have the authors evaluated a task-agnostic slide encoder pretrained on the same 29 tasks under an identical training protocol? Such a comparison would help determine how much of FACET's performance improvement is attributable specifically to task-conditioned aggregation rather than supervised multitask pretraining.

  3. Are there results comparing morphology-rich task descriptions with task-name-only or mismatched descriptions? This would help clarify whether FACET learns meaningful task-dependent visual selection beyond encoding task identity.

  4. Given that FACET utilizes class prototypes for text alignment during the pre-training phase, why does the paper not fully demonstrate its zero-shot classification performance based directly on the class prototypes of new tasks?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.