Cancer Indication Prediction on MSK cohort oncology
0.778AUC-ROC BladderHybrid (M+I)
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Hybrid (M+I)Backbone=Qwen3 8B, Learning Objective=SFT + JEPA, Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.778 | 0.686 | 0.818 | 0.766 | 0.698 | 0.76 | 0.751 | 0.698 | 0.75 | |
| Curr. (M+I)Backbone=Qwen3 8B, Learning Objective=Curriculum (SFT then JEPA), Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.773 | 0.689 | 0.817 | 0.766 | 0.694 | 0.748 | 0.749 | 0.694 | 0.748 | |
| SFT (M)Backbone=Qwen3 8B, Learning Objective=Supervised Fine-Tuning, Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.771 | 0.682 | 0.816 | 0.756 | 0.691 | 0.759 | 0.744 | 0.686 | 0.744 | |
| SFT (M+I)Backbone=Qwen3 8B, Learning Objective=Supervised Fine-Tuning, Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.77 | 0.683 | 0.815 | 0.756 | 0.692 | 0.754 | 0.745 | 0.688 | 0.745 | |
| Curr. (M)Backbone=LLaMA3.1 8B, Learning Objective=Curriculum (SFT then JEPA), Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.769 | 0.675 | 0.818 | 0.758 | 0.702 | 0.746 | 0.743 | 0.681 | 0.743 | |
| Hybrid (M+I)Backbone=LLaMA3.1 8B, Learning Objective=SFT + JEPA, Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.764 | 0.674 | 0.819 | 0.757 | 0.695 | 0.749 | 0.742 | 0.693 | 0.744 | |
| Hybrid (M)Backbone=LLaMA3.1 8B, Learning Objective=SFT + JEPA, Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.763 | 0.691 | 0.805 | 0.746 | 0.667 | 0.748 | 0.737 | 0.667 | 0.735 | |
| Curr. (M+I)Backbone=LLaMA3.1 8B, Learning Objective=Curriculum (SFT then JEPA), Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.763 | 0.683 | 0.816 | 0.755 | 0.695 | 0.748 | 0.742 | 0.691 | 0.739 | |
| SFT (M+I)Backbone=LLaMA3.1 8B, Learning Objective=Supervised Fine-Tuning, Training Data=MSK + INSPECT, Evaluation Protocol=Linear Probing2026.01 | 0.759 | 0.667 | 0.805 | 0.745 | 0.66 | 0.737 | 0.732 | 0.693 | 0.734 | |
| Curr. (M)Backbone=Qwen3 8B, Learning Objective=Curriculum (SFT then JEPA), Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.759 | 0.692 | 0.803 | 0.743 | 0.669 | 0.754 | 0.726 | 0.672 | 0.722 | |
| Hybrid (M)Backbone=Qwen3 8B, Learning Objective=SFT + JEPA, Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.752 | 0.687 | 0.801 | 0.744 | 0.671 | 0.756 | 0.717 | 0.666 | 0.723 | |
| SFT (M)Backbone=LLaMA3.1 8B, Learning Objective=Supervised Fine-Tuning, Training Data=MSK, Evaluation Protocol=Linear Probing2026.01 | 0.751 | 0.665 | 0.803 | 0.744 | 0.656 | 0.752 | 0.729 | 0.678 | 0.73 | |
| Random Foresttype=Baseline, Evaluation Protocol=Linear Probing2026.01 | 0.707 | 0.582 | 0.664 | 0.745 | 0.696 | 0.637 | 0.735 | 0.628 | 0.681 | |
| XGBoosttype=Baseline, Evaluation Protocol=Linear Probing2026.01 | 0.701 | 0.641 | 0.671 | 0.735 | 0.695 | 0.659 | 0.731 | 0.643 | 0.689 | |
| Logistic Regressiontype=Baseline, Evaluation Protocol=Linear Probing2026.01 | 0.676 | 0.561 | 0.672 | 0.67 | 0.679 | 0.579 | 0.652 | 0.63 | 0.653 |