diff --git a/examples/train/configs/genrl_wan2.1_t2v_1.3B_longcat.yaml b/examples/train/configs/genrl_wan2.1_t2v_1.3B_longcat.yaml
index 73783b023..3965f8a72 100644
--- a/examples/train/configs/genrl_wan2.1_t2v_1.3B_longcat.yaml
+++ b/examples/train/configs/genrl_wan2.1_t2v_1.3B_longcat.yaml
@@ -115,8 +115,7 @@ training:
     enable_gradient_checkpointing_type: full
 
 callbacks:
-  grad_clip:
-    max_grad_norm: 0.0  # Disabled; GenRLMethod clips internally.
+  # Gradnorm call back Disabled; GenRLMethod clips internally.
   ema:
     decay: 0.9
     start_iter: 0
diff --git a/fastvideo/train/methods/rl/genrl.py b/fastvideo/train/methods/rl/genrl.py
index 73eba7c43..24cc80243 100644
--- a/fastvideo/train/methods/rl/genrl.py
+++ b/fastvideo/train/methods/rl/genrl.py
@@ -31,27 +31,27 @@ from fastvideo.train.methods.base import (
     LogScalar,
     TrainingMethod,
 )
-from fastvideo.train.methods.rl.advantages import (
+from fastvideo.train.methods.rl.utils.advantages import (
     compute_advantages,
 )
-from fastvideo.train.methods.rl.data import (
+from fastvideo.train.methods.rl.utils.data import (
     build_prompt_dataloaders,
 )
-from fastvideo.train.methods.rl.diffusion import (
+from fastvideo.train.methods.rl.utils.diffusion import (
     compute_log_prob,
 )
-from fastvideo.train.methods.rl.embeddings import (
+from fastvideo.train.methods.rl.utils.embeddings import (
     compute_text_embeddings,
 )
-from fastvideo.train.methods.rl.rewards import (
+from fastvideo.train.methods.rl.utils.rewards import (
     move_reward_models,
     multi_score,
     reward_models_on_device,
 )
-from fastvideo.train.methods.rl.sampling import (
+from fastvideo.train.methods.rl.utils.sampling import (
     sample_epoch,
 )
-from fastvideo.train.methods.rl.stat_tracking import (
+from fastvideo.train.methods.rl.utils.stat_tracking import (
     PerPromptStatTracker,
 )
 from fastvideo.train.models.base import ModelBase
diff --git a/fastvideo/train/methods/rl/utils/__init__.py b/fastvideo/train/methods/rl/utils/__init__.py
new file mode 100644
index 000000000..e4282003a
--- /dev/null
+++ b/fastvideo/train/methods/rl/utils/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Utility modules for RL training."""
diff --git a/fastvideo/train/methods/rl/advantages.py b/fastvideo/train/methods/rl/utils/advantages.py
similarity index 99%
rename from fastvideo/train/methods/rl/advantages.py
rename to fastvideo/train/methods/rl/utils/advantages.py
index f8fb09c56..b4d935af9 100644
--- a/fastvideo/train/methods/rl/advantages.py
+++ b/fastvideo/train/methods/rl/utils/advantages.py
@@ -8,7 +8,7 @@ from typing import Any
 import numpy as np
 
 from fastvideo.logger import init_logger
-from fastvideo.train.methods.rl.stat_tracking import (
+from fastvideo.train.methods.rl.utils.stat_tracking import (
     EPSILON,
     PerPromptStatTracker,
 )
diff --git a/fastvideo/train/methods/rl/data.py b/fastvideo/train/methods/rl/utils/data.py
similarity index 100%
rename from fastvideo/train/methods/rl/data.py
rename to fastvideo/train/methods/rl/utils/data.py
diff --git a/fastvideo/train/methods/rl/diffusion.py b/fastvideo/train/methods/rl/utils/diffusion.py
similarity index 98%
rename from fastvideo/train/methods/rl/diffusion.py
rename to fastvideo/train/methods/rl/utils/diffusion.py
index 3fc47639c..e43a8e438 100644
--- a/fastvideo/train/methods/rl/diffusion.py
+++ b/fastvideo/train/methods/rl/utils/diffusion.py
@@ -5,7 +5,7 @@ from __future__ import annotations
 
 import torch
 
-from fastvideo.train.methods.rl.sde import (
+from fastvideo.train.methods.rl.utils.sde import (
     sde_step_with_logprob,
 )
 
diff --git a/fastvideo/train/methods/rl/embeddings.py b/fastvideo/train/methods/rl/utils/embeddings.py
similarity index 100%
rename from fastvideo/train/methods/rl/embeddings.py
rename to fastvideo/train/methods/rl/utils/embeddings.py
diff --git a/fastvideo/train/methods/rl/evaluation.py b/fastvideo/train/methods/rl/utils/evaluation.py
similarity index 96%
rename from fastvideo/train/methods/rl/evaluation.py
rename to fastvideo/train/methods/rl/utils/evaluation.py
index d019c591b..1c10bf64b 100644
--- a/fastvideo/train/methods/rl/evaluation.py
+++ b/fastvideo/train/methods/rl/utils/evaluation.py
@@ -9,10 +9,10 @@ from typing import Any
 import torch
 
 from fastvideo.logger import init_logger
-from fastvideo.train.methods.rl.embeddings import (
+from fastvideo.train.methods.rl.utils.embeddings import (
     compute_text_embeddings,
 )
-from fastvideo.train.methods.rl.pipeline import (
+from fastvideo.train.methods.rl.utils.pipeline import (
     wan_denoising_with_logprob,
 )
 
diff --git a/fastvideo/train/methods/rl/pipeline.py b/fastvideo/train/methods/rl/utils/pipeline.py
similarity index 99%
rename from fastvideo/train/methods/rl/pipeline.py
rename to fastvideo/train/methods/rl/utils/pipeline.py
index 59ee42ca7..197ec7989 100644
--- a/fastvideo/train/methods/rl/pipeline.py
+++ b/fastvideo/train/methods/rl/utils/pipeline.py
@@ -19,7 +19,7 @@ from fastvideo.logger import init_logger
 
 _pipeline_logger = init_logger(__name__)
 
-from fastvideo.train.methods.rl.sde import (
+from fastvideo.train.methods.rl.utils.sde import (
     sde_step_with_logprob,
 )
 
diff --git a/fastvideo/train/methods/rl/rewards.py b/fastvideo/train/methods/rl/utils/rewards.py
similarity index 100%
rename from fastvideo/train/methods/rl/rewards.py
rename to fastvideo/train/methods/rl/utils/rewards.py
diff --git a/fastvideo/train/methods/rl/sampling.py b/fastvideo/train/methods/rl/utils/sampling.py
similarity index 98%
rename from fastvideo/train/methods/rl/sampling.py
rename to fastvideo/train/methods/rl/utils/sampling.py
index cc663d0ef..8f68730b2 100644
--- a/fastvideo/train/methods/rl/sampling.py
+++ b/fastvideo/train/methods/rl/utils/sampling.py
@@ -11,10 +11,10 @@ from typing import Any
 import torch
 
 from fastvideo.logger import init_logger
-from fastvideo.train.methods.rl.embeddings import (
+from fastvideo.train.methods.rl.utils.embeddings import (
     compute_text_embeddings,
 )
-from fastvideo.train.methods.rl.pipeline import (
+from fastvideo.train.methods.rl.utils.pipeline import (
     wan_denoising_with_logprob,
 )
 
diff --git a/fastvideo/train/methods/rl/sde.py b/fastvideo/train/methods/rl/utils/sde.py
similarity index 100%
rename from fastvideo/train/methods/rl/sde.py
rename to fastvideo/train/methods/rl/utils/sde.py
diff --git a/fastvideo/train/methods/rl/stat_tracking.py b/fastvideo/train/methods/rl/utils/stat_tracking.py
similarity index 100%
rename from fastvideo/train/methods/rl/stat_tracking.py
rename to fastvideo/train/methods/rl/utils/stat_tracking.py