diff --git a/lnp_ml/modeling/benchmark.py b/lnp_ml/modeling/benchmark.py
index 2c6be78..42f2ff5 100644
--- a/lnp_ml/modeling/benchmark.py
+++ b/lnp_ml/modeling/benchmark.py
@@ -1,6 +1,7 @@
 """Benchmark 脚本：在 baseline 论文公开的 CV 划分上评估模型（仅 delivery 任务）"""
 
 import json
+import math
 from pathlib import Path
 from typing import Dict, List, Optional
 
@@ -9,6 +10,7 @@ import pandas as pd
 import torch
 import torch.nn as nn
 from torch.utils.data import DataLoader
+from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR, SequentialLR
 from loguru import logger
 from tqdm import tqdm
 from sklearn.metrics import mean_squared_error, r2_score
@@ -158,6 +160,7 @@ def train_fold(
     weight_decay: float = 1e-5,
     epochs: int = 50,
     patience: int = 10,
+    warmup_epochs: int = 3,
     config: Optional[Dict] = None,
 ) -> Dict:
     """训练单个 fold"""
@@ -166,9 +169,19 @@ def train_fold(
     logger.info(f"{'='*60}")
     
     optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
-    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
-        optimizer, mode="min", factor=0.5, patience=5
+    
+    warmup_scheduler = LambdaLR(
+        optimizer, lr_lambda=lambda epoch: (epoch + 1) / warmup_epochs
     )
+    cosine_scheduler = CosineAnnealingLR(
+        optimizer, T_max=epochs - warmup_epochs
+    )
+    scheduler = SequentialLR(
+        optimizer,
+        schedulers=[warmup_scheduler, cosine_scheduler],
+        milestones=[warmup_epochs],
+    )
+    
     early_stopping = EarlyStopping(patience=patience)
     
     best_val_loss = float("inf")
@@ -198,7 +211,7 @@ def train_fold(
             "lr": current_lr,
         })
         
-        scheduler.step(val_metrics["loss"])
+        scheduler.step()
         
         if val_metrics["loss"] < best_val_loss:
             best_val_loss = val_metrics["loss"]
@@ -284,11 +297,11 @@ def main(
     data_dir: Path = PROCESSED_DATA_DIR / "benchmark",
     output_dir: Path = MODELS_DIR / "benchmark",
     # 模型参数
-    d_model: int = 128,
-    num_heads: int = 4,
-    n_attn_layers: int = 2,
+    d_model: int = 256,
+    num_heads: int = 8,
+    n_attn_layers: int = 4,
     fusion_strategy: str = "attention",
-    head_hidden_dim: int = 64,
+    head_hidden_dim: int = 128,
     dropout: float = 0.1,
     # MPNN 参数
     use_mpnn: bool = False,
@@ -416,7 +429,9 @@ def main(
             model.rdkit_encoder._cache = rdkit_cache
             logger.info(f"Reusing RDKit cache with {len(rdkit_cache)} entries")
         
-        logger.info(f"Model parameters: {sum(p.numel() for p in model.parameters()):,}")
+        n_params_total = sum(p.numel() for p in model.parameters())
+        n_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        logger.info(f"Model parameters: {n_params_total:,} total, {n_params_trainable:,} trainable")
         
         # 训练
         result = train_fold(
diff --git a/lnp_ml/modeling/pretrain.py b/lnp_ml/modeling/pretrain.py
index abb702d..b7d9d66 100644
--- a/lnp_ml/modeling/pretrain.py
+++ b/lnp_ml/modeling/pretrain.py
@@ -303,8 +303,9 @@ def main(
             dropout=dropout,
         )
 
-    n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
-    logger.info(f"Model parameters: {n_params:,}")
+    n_params_total = sum(p.numel() for p in model.parameters())
+    n_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    logger.info(f"Model parameters: {n_params_total:,} total, {n_params_trainable:,} trainable")
 
     # 预热 RDKit 缓存（避免训练时阻塞）
     all_smiles = train_df["smiles"].tolist() + val_df["smiles"].tolist()