ALICE: A Single Foundation Model Absorbs Eight Pathology AI Systems Into One Backbone

A research team has introduced ALICE, a pathology foundation model built not from scratch but from the distilled knowledge of eight separate teacher models. Rather than training a single network on raw slide data alone, the researchers behind ALICE used a technique they call multi-stage agglomerative distillation to merge vision-only, vision-language, and slide-level models into dedicated modules within one unified backbone.

The scale of the underlying pretraining data is substantial. ALICE was exposed to nearly 25 million tile-level pathology images alongside more than 155,000 high-resolution whole-slide images, giving it broad visual coverage of tissue patterns before any distillation from teacher models even began. The paper, submitted to arXiv on July 10, 2026, is filed under both Computer Vision and Pattern Recognition and Artificial Intelligence categories, reflecting its dual identity as both a vision model and a multimodal reasoning system.

Why Consolidation Matters in Digital Pathology

Pathology AI has historically fragmented into narrow specialists: some models excel at classifying tissue regions of interest, others handle vision-language tasks like generating captions or answering questions about a slide, and still others operate at the whole-slide level needed for clinical diagnostics. Running separate models for each of these jobs is computationally expensive and operationally messy for hospitals or labs trying to deploy AI at scale.

ALICE's agglomerative distillation approach addresses that fragmentation directly. By folding eight distinct teacher models into dedicated modules of a single architecture, the system aims to preserve each teacher's specialized strengths while offering a single point of deployment. That structural choice is what differentiates ALICE from typical foundation models, which are usually trained end-to-end on one objective rather than assembled from multiple pre-existing specialists.

Benchmark Results Across Nearly 100 Tasks

The evaluation methodology behind ALICE is notably broad. The researchers tested the model across 21 task scenarios spanning 96 downstream tasks, drawing on 48 distinct data sources. This breadth covers the three major categories of pathology AI work: region-of-interest tissue analysis, vision-language multimodal evaluation, and whole-slide clinical assessment.

Across all three categories, ALICE recorded the best average rank among foundation models that were matched to the same tasks, according to the paper. That result suggests the distillation strategy did not merely average out the capabilities of its teacher models but produced a system that performs competitively, and in some cases better, than models purpose-built for a single evaluation category.

For institutions weighing whether to adopt digital pathology AI, the appeal of a consolidated model is straightforward: fewer systems to validate, deploy, and maintain, without sacrificing performance across the range of tasks pathologists and researchers actually need. Whether ALICE's approach generalizes beyond the 48 data sources tested, and how it performs in real clinical workflows outside benchmark conditions, remains an open question the paper itself does not resolve.