EXAONE Path 2.0

EXAONE Path 2.0 is a computational pathology foundation model from LG AI Research that learns patch-level representations of whole-slide images (WSIs) under direct slide-level supervision. Most pathology foundation models follow a two-stage recipe: pretrain a patch encoder with self-supervised learning (DINO or DINOv2) on millions of small tiles, then aggregate the frozen patch embeddings with multiple instance learning or a slide encoder. That recipe scales, but the supervision signal never reaches the patch encoder, so the features it learns are shaped by generic image augmentations rather than by the molecular endpoints clinicians actually care about — mutation status, microsatellite instability, hormone receptor subtype.

EXAONE Path 2.0 closes that loop. Slide-level labels propagate as gradients through every stage of a hierarchical vision transformer, all the way down to the patch encoder, so the encoder is optimized for clinically relevant signal from the start. The payoff is data efficiency: trained on 37,195 WSIs — a fraction of what image-only self-supervised models consume — it reaches the highest average AUROC across a 10-task biomarker benchmark, ahead of TITAN, PRISM, CHIEF, Prov-GigaPath, and UNI.

Released as an arXiv technical report in July 2025, it sits between the patch-level EXAONEPath 1.0 and the later multi-omics EXAONE Path 2.5 in LG AI Research's pathology lineage.

#Key Features

• End-to-end slide-level supervision: Cross-entropy loss computed on slide labels backpropagates through all three transformer stages, letting the patch encoder learn features tied to biomarker endpoints instead of relying only on self-supervised proxies.
• Hierarchical three-stage transformer: A Hierarchical Image Pyramid Transformer design in which a first-stage ViT encodes individual patches, a second-stage ViT builds region representations, and a third-stage ViT integrates regions into a slide.
• Curriculum learning with progressive resolution: Training first applies DINO losses at 256×256 and 1024×1024, then scales the region stage to 4096×4096 and introduces the slide-level supervised loss, avoiding full-resolution three-stage computation at every step.
• Multi-task supervision: Joint optimization over cancer subtyping across 33 cancer types, tissue classification across 12 organ systems, and molecular biomarker prediction, which discourages overfitting to any single endpoint.
• Early-exit downstream adaptation: Rather than fine-tuning the full hierarchy, downstream tasks use first-stage patch features aggregated by a CLAM attention module, which limits overfitting in the small-cohort regime typical of pathology.

#Technical Details

Training used 37,195 formalin-fixed, paraffin-embedded H&E-stained WSIs, which yielded 144,450 image-label pairs across 16 training tasks — each slide carries several labels spanning subtyping, tissue type, and biomarker status. Gigapixel end-to-end training is made tractable through activation checkpointing and CPU offloading, computing and transferring activations on demand rather than holding all patch embeddings in GPU memory. Evaluation covers 10 slide-level classification tasks: four built from private cohorts at one Korean and two US hospitals (lung adenocarcinoma TMB, EGFR, and KRAS; colorectal MSI) and six from public CPTAC breast, renal, and colon collections (TP53, PIK3CA, PBRM1, BAP1, KRAS, TP53). With the backbone frozen and only the aggregator trained, averaged over four seeds, EXAONE Path 2.0 reaches 0.784 mean AUROC versus 0.765 for PRISM, 0.755 for UNI2-h, 0.748 for TITAN, 0.731 for EXAONEPath 1.0, 0.707 for Prov-GigaPath, and 0.696 for CHIEF. Task-level highlights include 0.853 on LUAD-EGFR, 0.912 on COAD-KRAS, and 0.807 on RCC-BAP1.

#Applications

The model targets molecular biomarker inference from routine H&E histology — predicting mutation status, microsatellite instability, and tumor mutation burden without a sequencing assay. Practical uses include pre-screening patients for confirmatory molecular testing, enriching retrospective cohorts where sequencing was never performed, and serving as a frozen feature extractor for slide-level classifiers built by translational research groups. LG AI Research has also published task-specific derivatives on HuggingFace, including EGFR- and MSI-focused checkpoints built on the same backbone.

#Impact

EXAONE Path 2.0 is a useful counterweight to the prevailing assumption that pathology foundation models improve mainly by scaling self-supervised pretraining: it matches or beats models trained on an order of magnitude more slides and with substantially more parameters, by spending its supervision budget differently. Several caveats bound the result. The work is an arXiv technical report rather than a peer-reviewed paper; the weights are released under the EXAONEPath AI Model License Agreement 1.0 - NC, which restricts use to research and forbids commercial deployment; four of the ten benchmarks rest on private hospital data that cannot be independently reproduced; and the CPTAC test splits are small enough — RCC-BAP1 has 50 test slides, COAD-KRAS 19 — that individual task AUROCs carry wide uncertainty. The training corpus itself is not released.