docs: update README file

Author: yousefkotp

CITATION.cff

@@ -41,4 +41,5 @@ preferred-citation:
     - family-names: Hosseini
       given-names: Mahdi S.
       orcid: "https://orcid.org/0000-0002-9147-0731"
+  url: "https://arxiv.org/abs/2603.27048"
   year: 2026

MODEL_CARD.md

@@ -1,5 +1,7 @@
 # Model Card: MOOZY
 
+**Paper:** [arXiv:2603.27048](https://arxiv.org/abs/2603.27048)
+
 ## Model Description
 
 MOOZY is a patient-first foundation model for computational pathology. It produces 768-dimensional embeddings at the slide and case (patient) level from whole-slide image feature grids.
@@ -31,24 +33,7 @@ Anatomical sites: adrenal gland, bladder, brain, breast, cervix, colon/rectum, e
 ## Training Procedure
 
 - **Stage 1**: 200 epochs (14,400 steps, ~436 GPU-hours on 8 GPUs). Effective batch size 1,024. AdamW with cosine LR schedule. EMA teacher with cosine momentum (0.996 to 1.0). Multi-crop: 2 global (20x20) + 4 local (12x12) with block masking (ratio 0.1-0.5).
-- **Stage 2**: 12 epochs (600 steps, ~307 GPU-hours on 8 GPUs). Effective batch size 1,024. AdamW with base LR 5e-5, cosine schedule, gradient clipping at 0.3. Classification: cross-entropy with label smoothing (0.03) and inverse-frequency weighting. Survival: discrete-time hazard NLL with adaptive binning.
-
-## Evaluation
-
-Evaluated on 8 held-out tasks with 5-fold frozen-feature MLP probe, patient-level fold grouping. All evaluation tasks were excluded from training.
-
-| Task | Dataset | Weighted F1 | Weighted ROC-AUC | Balanced Accuracy |
-|------|---------|-------------|------------------|-------------------|
-| Residual Cancer Burden | BC Therapy | 0.56 | 0.74 | 0.51 |
-| TP53 Mutation | CPTAC-BRCA | 0.87 | 0.86 | 0.86 |
-| BAP1 Mutation | CPTAC-CCRCC | 0.89 | 0.79 | 0.78 |
-| ACVR2A Mutation | CPTAC-COAD | 0.91 | 0.91 | 0.90 |
-| Histologic Grade | CPTAC-LSCC | 0.78 | 0.75 | 0.77 |
-| KRAS Mutation | CPTAC-LUAD | 0.85 | 0.80 | 0.79 |
-| IDH Status | EBRAINS | 0.97 | 0.99 | 0.97 |
-| Treatment Response | MBC | 0.58 | 0.68 | 0.48 |
-
-Macro-average improvements over TITAN: +7.37% F1, +5.50% AUC, +7.83% balanced accuracy. Over PRISM: +8.83% F1, +10.70% AUC, +9.78% balanced accuracy.
+- **Stage 2**: 20 epochs (1,000 steps, ~512 GPU-hours on 8 GPUs). Effective batch size 1,024 cases (micro batch 1, 128 accumulation steps). AdamW with base LR 5e-5, cosine schedule, gradient clipping at 0.3. Classification: cross-entropy with label smoothing (0.03) and inverse-frequency weighting. Survival: discrete-time hazard NLL with adaptive binning.
 
 ## Intended Use