diff --git a/benchmark/reports/gsm8k.md b/benchmark/reports/gsm8k.md
new file mode 100644
index 0000000000..3ebfeb57f8
--- /dev/null
+++ b/benchmark/reports/gsm8k.md
@@ -0,0 +1,899 @@
+# GSM8K Benchmark Results
+
+## 1. Task Overview
+
+[GSM8K](https://huggingface.co/datasets/openai/gsm8k) is a dataset comprising over 8,500 high-school-level mathematics word problems. Each problem is accompanied by a detailed, step-by-step solution that includes intermediate calculations. The task requires the model to generate such a complete reasoning trace from the problem statement alone.
+
+---
+
+## 2. Experimental Setup
+
+Due to inherent stochasticity in both vLLM-based rollouts and the training process itself, we designed four controlled experiments to isolate sources of divergence and ensure a fair comparison between **veRL** and **Trinity**:
+
+- **Experiment 1 (Exact Reproduction)**: veRL saves both the input batch and the full model/optimizer state at every `update_actor` step. Trinity loads both artifacts to exactly reproduce veRL’s training trajectory.
+- **Experiment 2 (State-Aligned Training)**: veRL saves only the model and optimizer states after each step. Trinity loads these states but generates its own rollout data (i.e., does not reuse veRL’s input batches).
+- **Experiment 3 (Independent Training, Zero Regularization)**: Both frameworks train independently with identical hyperparameters and regularization coefficients set to zero (kl_loss_coef = entropy_coef = 0), allowing divergence from both rollouts and training dynamics.
+- **Experiment 4 (Independent Training, Non-zero Regularization)**: Identical to Experiment 3 but with active regularization (kl_loss_coef = 0.001, entropy_coef = 0.003).
+
+To enable these comparisons, targeted modifications were made to both codebases as detailed below.
+
+---
+
+### Code Modifications
+
+#### veRL Modifications
+
+- **`protocol.py`**: Fixed a bug in `DataProto.concat()` (present in `verl < 0.6.0`) related to aggregating `meta_info`. The fix aligns with the official `verl==0.6.0` implementation.
+- **`ray_trainer.py`**: Added logic to save the actor’s input batch at every training step.
+- **`fsdp_workers.py`**: Adjusted implementation to match Trinity (specifically, commented out a type conversion to avoid potential numerical discrepancies).
+
+> ⚠️ **Note**: All `batch_file` paths are relative local paths. Users must adjust them according to their runtime environment.
+
+<details><summary>Click to expand patch</summary>
+
+```patch
+diff --git a/verl/protocol.py b/verl/protocol.py
+index 39979f84..8ea485b5 100644
+--- a/verl/protocol.py
++++ b/verl/protocol.py
+@@ -758,8 +758,32 @@ class DataProto:
+         for key, val in non_tensor_batch.items():
+             non_tensor_batch[key] = np.concatenate(val, axis=0)
+
++        # Merge meta_info with special handling for metrics
++        merged_meta_info = {}
++        if data:
++            # Merge non-metric meta_info and aggregate metrics from all workers.
++            all_metrics = []
++            for d in data:
++                for k, v in d.meta_info.items():
++                    if k == "metrics":
++                        if v is not None:
++                            if isinstance(v, list):
++                                all_metrics.extend(v)
++                            else:
++                                all_metrics.append(v)
++                    else:
++                        if k in merged_meta_info:
++                            # Ensure consistency for overlapping non-metric keys
++                            assert merged_meta_info[k] == v, f"Conflicting values for meta_info key '{k}'"
++                        else:
++                            merged_meta_info[k] = v
++
++            # Flatten list of dicts to dict of lists for consistent metrics structure
++            if all_metrics:
++                merged_meta_info["metrics"] = list_of_dict_to_dict_of_list(all_metrics)
++
+         cls = type(data[0]) if len(data) > 0 else DataProto
+-        return cls(batch=new_batch, non_tensor_batch=non_tensor_batch, meta_info=data[0].meta_info)
++        return cls(batch=new_batch, non_tensor_batch=non_tensor_batch, meta_info=merged_meta_info)
+
+     def reorder(self, indices):
+         """
+diff --git a/verl/trainer/ppo/ray_trainer.py b/verl/trainer/ppo/ray_trainer.py
+index 45ed80a4..6458de18 100644
+--- a/verl/trainer/ppo/ray_trainer.py
++++ b/verl/trainer/ppo/ray_trainer.py
+@@ -1316,6 +1316,16 @@ class RayPPOTrainer:
+                         critic_output_metrics = reduce_metrics(critic_output.meta_info["metrics"])
+                         metrics.update(critic_output_metrics)
+
++                    import os, pickle, random
++                    batch_file = f'./data/step_{self.global_steps}.pkl'
++                    if os.path.exists(batch_file):
++                        with open(batch_file, 'rb') as f:
++                            batch = pickle.load(f)
++                    else:
++                        os.makedirs('./data/', exist_ok=True)
++                        with open(f'./data/step_{self.global_steps}.pkl', 'wb') as f:
++                            pickle.dump(batch, f)
++
+                     # implement critic warmup
+                     if self.config.trainer.critic_warmup <= self.global_steps:
+                         # update actor
+diff --git a/verl/workers/fsdp_workers.py b/verl/workers/fsdp_workers.py
+index 4141d986..2895fe16 100644
+--- a/verl/workers/fsdp_workers.py
++++ b/verl/workers/fsdp_workers.py
+@@ -309,7 +309,7 @@ class ActorRolloutRefWorker(Worker, DistProfilerExtension):
+             )
+
+             # some parameters may not in torch_dtype. TODO(zhangchi.usc1992) remove this after we switch to fsdp2
+-            actor_module.to(torch_dtype)
++            # actor_module.to(torch_dtype)
+
+             if enable_gradient_checkpointing:
+                 actor_module.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
+```
+
+</details>
+
+---
+
+#### Trinity Modifications
+
+The primary changes enable Trinity to load checkpoints and batch data saved by veRL:
+
+- In `verl_trainer.py`, added logic to load veRL-saved input batches (used in Experiment 1).
+- Implemented checkpoint loading for the actor model and optimizer state at corresponding training steps.
+
+> ⚠️ **Note**: Both `batch_file` and `verl_checkpoint_folder` currently use hardcoded local paths for demonstration.
+> - **Experiment 1** uses batch loading and checkpoint restoration as shown.
+> - **Experiment 2** does not require batch loading; the relevant code should be disabled.
+> - **Experiments 3 and 4** were run using the *original, unmodified* version of `verl_trainer.py` (i.e., without the changes in this patch). To reproduce those experiments, these modifications must be reverted or disabled.
+
+<details><summary>Click to expand patch</summary>
+
+```patch
+diff --git a/trinity/trainer/verl_trainer.py b/trinity/trainer/verl_trainer.py
+index 603f54c98e..805fb91683 100644
+--- a/trinity/trainer/verl_trainer.py
++++ b/trinity/trainer/verl_trainer.py
+@@ -464,6 +464,12 @@ class VerlPPOTrainerWrapper(RayPPOTrainer, TrainEngineWrapper):
+                 critic_output_metrics = reduce_metrics(critic_output.meta_info["metrics"])
+                 metrics.update(critic_output_metrics)
+
++            import pickle
++            batch_file = f'./data/step_{self.global_steps}.pkl'
++            if os.path.exists(batch_file):
++                self.logger.info(f'Load batch from {batch_file}')
++                with open(batch_file, 'rb') as f:
++                    batch = pickle.load(f)
+             # implement critic warmup
+             if (
+                 not self.algorithm.use_critic
+@@ -485,6 +491,16 @@ class VerlPPOTrainerWrapper(RayPPOTrainer, TrainEngineWrapper):
+             compute_throughout_metrics(batch=batch, timing_raw=timing_raw, n_gpus=n_gpus)
+         )
+
++        verl_checkpoint_folder = 'checkpoints/Trinity-RFT-cys-gsm8k/verl-unshuffle-0kl-0entropy-no_metric_bug-6gpu'
++        if not os.path.isabs(verl_checkpoint_folder):
++            working_dir = os.getcwd()
++            verl_checkpoint_folder = os.path.join(working_dir, verl_checkpoint_folder)
++        suffix = f'global_step_{self.global_steps}'
++        actor_checkpoint_path = os.path.join(verl_checkpoint_folder, suffix, 'actor')
++        self.logger.info(f"loading checkpoint from {actor_checkpoint_path}")
++        self.actor_rollout_wg.load_checkpoint(actor_checkpoint_path)
++        self.logger.info(f"loaded checkpoint from {actor_checkpoint_path}")
++
+         return metrics
+
+     def save_checkpoint(self, block_until_saved: bool = False, save_as_hf: bool = False) -> None:
+```
+
+</details>
+
+Additionally, to ensure reward computation matches veRL exactly, we implemented a dedicated workflow `VerlGSM8kWorkflow` that reuses veRL’s GSM8K scoring logic:
+
+<details><summary>Click to expand implementation</summary>
+
+```python
+from dataclasses import asdict
+from typing import List, Optional
+
+import openai
+from trinity.common.experience import Experience
+from trinity.common.models.model import ModelWrapper
+from trinity.common.workflows.workflow import WORKFLOWS, Task, Workflow
+
+from verl.utils.reward_score import gsm8k
+
+
+@WORKFLOWS.register_module("verl_gsm8k_workflow")
+class VerlGSM8kWorkflow(Workflow):
+    can_reset: bool = True
+    can_repeat: bool = True
+
+    def __init__(
+        self,
+        *,
+        task: Task,
+        model: ModelWrapper,
+        auxiliary_models: Optional[List[openai.OpenAI]] = None,
+    ):
+        self.reset(task)
+        super().__init__(
+            task=task,
+            model=model,
+            auxiliary_models=auxiliary_models,
+        )
+
+    def reset(self, task: Task):
+        self.task = task
+        self.raw_task = task.raw_task
+        self.task_message = self.raw_task['prompt']
+        self.truth = self.raw_task['reward_model']['ground_truth']
+
+    def set_repeat_times(self, repeat_times, run_id_base):
+        self.repeat_times = repeat_times
+        self.task.rollout_args.n = repeat_times
+        self.run_id_base = run_id_base
+
+    @property
+    def rollout_args(self):
+        return asdict(self.task.rollout_args)
+
+    def run(self) -> List[Experience]:
+        messages = self.task_message
+
+        self.logger.debug("start chat")
+        responses = self.model.chat(messages, **self.rollout_args)
+        for i, response in enumerate(responses):
+            reward_dict = {
+                "accuracy": gsm8k.compute_score(solution_str=response.response_text, ground_truth=self.truth),
+            }
+
+            if response.metrics is None:
+                response.metrics = {}
+            response.metrics.update(reward_dict)
+            reward = sum(reward_dict.values())
+            response.reward = reward
+            response.eid.run = i + self.run_id_base
+
+            self.logger.debug(
+                f"self.task_message: {self.task_message}, messages: {messages}, response: {response.response_text}, reward: {reward}"
+            )
+        return responses
+```
+
+</details>
+
+---
+
+### Hardware Configuration
+
+To ensure checkpoint compatibility, **both veRL and Trinity trainers were configured to use 6 GPUs**, despite Trinity’s YAML specifying `gpu_per_node=8`. Only 6 GPUs were actively used during training.
+
+---
+
+#### Experiment 1: Exact Training Trajectory Reproduction
+
+<details><summary>veRL Launch Script</summary>
+
+```bash
+export MODEL_PATH=Qwen/Qwen2.5-1.5B-Instruct
+project_name=Trinity-RFT-cys-gsm8k
+exp_name=verl-unshuffle-0kl-0entropy-no_metric_bug-6gpu
+PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
+ algorithm.adv_estimator=grpo \
+ algorithm.use_kl_in_reward=false \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ data.train_files=$HOME/data/gsm8k/train.parquet \
+ data.val_files=$HOME/data/gsm8k/test.parquet \
+ data.train_batch_size=96 \
+ data.max_prompt_length=1024 \
+ data.max_response_length=2048 \
+ data.shuffle=false \
+ actor_rollout_ref.model.use_remove_padding=true \
+ actor_rollout_ref.actor.use_kl_loss=true \
+ actor_rollout_ref.actor.kl_loss_coef=0.0 \
+ actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+ actor_rollout_ref.actor.entropy_coeff=0.0 \
+ actor_rollout_ref.model.path=$MODEL_PATH \
+ actor_rollout_ref.actor.strategy=fsdp \
+ actor_rollout_ref.actor.optim.lr=1e-5 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=96 \
+ actor_rollout_ref.actor.use_dynamic_bsz=true \
+ actor_rollout_ref.actor.ppo_max_token_len_per_gpu=16384 \
+ actor_rollout_ref.actor.clip_ratio_low=0.2 \
+ actor_rollout_ref.actor.clip_ratio_high=0.28 \
+ actor_rollout_ref.actor.loss_agg_mode="token-mean" \
+ actor_rollout_ref.rollout.name=vllm \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.rollout.n=8 \
+ actor_rollout_ref.ref.log_prob_use_dynamic_bsz=true \
+ actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=16384 \
+ trainer.logger=wandb \
+ trainer.project_name=$project_name \
+ trainer.experiment_name=$exp_name \
+ trainer.val_before_train=false \
+ trainer.n_gpus_per_node=6 \
+ trainer.nnodes=1 \
+ trainer.save_freq=1 \
+ trainer.test_freq=1000 \
+ trainer.total_epochs=1 2>&1 | tee verl_gsm8k.log
+```
+
+</details>
+
+<details><summary>Trinity Configuration (YAML)</summary>
+
+```yaml
+mode: both
+project: Trinity-RFT-cys-gsm8k
+group: alignment
+name: 0-kl-0-entropy-fixed_input-no_metric_bug-load_checkpoint
+checkpoint_root_dir: ./checkpoints/
+algorithm:
+  algorithm_type: grpo
+  repeat_times: 8
+  sample_strategy: default
+  kl_loss_fn: low_var_kl
+  kl_loss_fn_args:
+    kl_coef: 0  # 0.001
+  policy_loss_fn_args:
+    clip_range_low: 0.2
+    clip_range_high: 0.28
+  entropy_loss_fn: default
+  entropy_loss_fn_args:
+    entropy_coef: 0  # 0.003
+  optimizer:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    warmup_style: constant
+data_processor: {}
+model:
+  model_path: Qwen/Qwen2.5-1.5B-Instruct
+  max_prompt_tokens: 1024
+  max_response_tokens: 2048
+cluster:
+  node_num: 1
+  gpu_per_node: 8
+buffer:
+  batch_size: 96
+  total_epochs: 1
+  explorer_input:
+    taskset:
+      name: taskset
+      storage_type: file
+      path: /root/data/gsm8k
+      split: train
+      format:
+        prompt_key: question
+        response_key: answer
+      rollout_args:
+        temperature: 1.0
+        logprobs: 0
+    eval_tasksets: []
+    default_workflow_type: verl_gsm8k_workflow
+  trainer_input:
+    experience_buffer:
+      name: experience_buffer
+      storage_type: queue
+      replay_buffer:
+        enable: false
+        priority_fn: linear_decay
+        reuse_cooldown_time: null
+        priority_fn_args:
+          decay: 2.0
+explorer:
+  runner_per_model: 8
+  rollout_model:
+    engine_num: 2
+    tensor_parallel_size: 1
+    enforce_eager: true
+    enable_prefix_caching: false
+    enable_chunked_prefill: false
+    gpu_memory_utilization: 0.9
+    dtype: bfloat16
+    seed: 42
+    enable_thinking: false
+    enable_history: false
+    enable_openai_api: false
+    enable_auto_tool_choice: false
+    tool_call_parser: null
+    reasoning_parser: null
+  auxiliary_models: []
+  eval_interval: 1000
+trainer:
+  trainer_type: verl
+  save_interval: 100
+  enable_preview: true
+  grad_clip: 1.0
+  max_token_len_per_gpu: 16384
+monitor:
+  monitor_type: wandb
+synchronizer:
+  sync_method: nccl
+  sync_style: fixed
+  sync_interval: 1
+  sync_timeout: 1200
+log:
+  level: INFO
+```
+
+</details>
+
+---
+
+#### Experiment 2: Model State Alignment with Independent Rollouts
+
+<details><summary>veRL Launch Script</summary>
+
+```bash
+export MODEL_PATH=Qwen/Qwen2.5-1.5B-Instruct
+project_name=Trinity-RFT-cys-gsm8k
+exp_name=verl-unshuffle-6gpu
+PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
+ algorithm.adv_estimator=grpo \
+ algorithm.use_kl_in_reward=false \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ data.train_files=$HOME/data/gsm8k/train.parquet \
+ data.val_files=$HOME/data/gsm8k/test.parquet \
+ data.train_batch_size=96 \
+ data.max_prompt_length=1024 \
+ data.max_response_length=2048 \
+ data.shuffle=false \
+ actor_rollout_ref.model.use_remove_padding=true \
+ actor_rollout_ref.actor.use_kl_loss=true \
+ actor_rollout_ref.actor.kl_loss_coef=0.001 \
+ actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+ actor_rollout_ref.actor.entropy_coeff=0.003 \
+ actor_rollout_ref.model.path=$MODEL_PATH \
+ actor_rollout_ref.actor.strategy=fsdp \
+ actor_rollout_ref.actor.optim.lr=1e-5 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=96 \
+ actor_rollout_ref.actor.use_dynamic_bsz=true \
+ actor_rollout_ref.actor.ppo_max_token_len_per_gpu=16384 \
+ actor_rollout_ref.actor.clip_ratio_low=0.2 \
+ actor_rollout_ref.actor.clip_ratio_high=0.28 \
+ actor_rollout_ref.actor.loss_agg_mode="token-mean" \
+ actor_rollout_ref.rollout.name=vllm \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.rollout.n=8 \
+ actor_rollout_ref.ref.log_prob_use_dynamic_bsz=true \
+ actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=16384 \
+ trainer.logger=wandb \
+ trainer.project_name=$project_name \
+ trainer.experiment_name=$exp_name \
+ trainer.val_before_train=false \
+ trainer.n_gpus_per_node=6 \
+ trainer.nnodes=1 \
+ trainer.save_freq=1 \
+ trainer.test_freq=1000 \
+ trainer.total_epochs=1 2>&1 | tee verl_gsm8k.log
+```
+
+</details>
+
+<details><summary>Trinity Configuration (YAML)</summary>
+
+```yaml
+mode: both
+project: Trinity-RFT-cys-gsm8k
+group: alignment
+name: load_checkpoint
+checkpoint_root_dir: ./checkpoints/
+algorithm:
+  algorithm_type: grpo
+  repeat_times: 8
+  sample_strategy: default
+  kl_loss_fn: low_var_kl
+  kl_loss_fn_args:
+    kl_coef: 0.001
+  policy_loss_fn_args:
+    clip_range_low: 0.2
+    clip_range_high: 0.28
+  entropy_loss_fn: default
+  entropy_loss_fn_args:
+    entropy_coef: 0.003
+  optimizer:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    warmup_style: constant
+data_processor: {}
+model:
+  model_path: Qwen/Qwen2.5-1.5B-Instruct
+  max_prompt_tokens: 1024
+  max_response_tokens: 2048
+cluster:
+  node_num: 1
+  gpu_per_node: 8
+buffer:
+  batch_size: 96
+  total_epochs: 1
+  explorer_input:
+    taskset:
+      name: taskset
+      storage_type: file
+      path: /root/data/gsm8k
+      split: train
+      format:
+        prompt_key: question
+        response_key: answer
+      rollout_args:
+        temperature: 1.0
+        logprobs: 0
+    eval_tasksets: []
+    default_workflow_type: verl_gsm8k_workflow
+  trainer_input:
+    experience_buffer:
+      name: experience_buffer
+      storage_type: queue
+      replay_buffer:
+        enable: false
+        priority_fn: linear_decay
+        reuse_cooldown_time: null
+        priority_fn_args:
+          decay: 2.0
+explorer:
+  runner_per_model: 8
+  rollout_model:
+    engine_num: 2
+    tensor_parallel_size: 1
+    enforce_eager: true
+    enable_prefix_caching: false
+    enable_chunked_prefill: false
+    gpu_memory_utilization: 0.9
+    dtype: bfloat16
+    seed: 42
+    enable_thinking: false
+    enable_history: false
+    enable_openai_api: false
+    enable_auto_tool_choice: false
+    tool_call_parser: null
+    reasoning_parser: null
+  auxiliary_models: []
+  eval_interval: 1000
+trainer:
+  trainer_type: verl
+  save_interval: 100
+  enable_preview: true
+  grad_clip: 1.0
+  max_token_len_per_gpu: 16384
+monitor:
+  monitor_type: wandb
+synchronizer:
+  sync_method: nccl
+  sync_style: fixed
+  sync_interval: 1
+  sync_timeout: 1200
+log:
+  level: INFO
+```
+
+</details>
+
+---
+
+#### Experiment 3: Independent Training, Zero Regularization
+
+<details><summary>veRL Launch Script</summary>
+
+```bash
+export MODEL_PATH=Qwen/Qwen2.5-1.5B-Instruct
+project_name=Trinity-RFT-cys-gsm8k
+exp_name=verl-unshuffle-0kl-0entropy
+PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
+ algorithm.adv_estimator=grpo \
+ algorithm.use_kl_in_reward=false \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ data.train_files=$HOME/data/gsm8k/train.parquet \
+ data.val_files=$HOME/data/gsm8k/test.parquet \
+ data.train_batch_size=96 \
+ data.max_prompt_length=1024 \
+ data.max_response_length=2048 \
+ data.shuffle=false \
+ actor_rollout_ref.model.use_remove_padding=true \
+ actor_rollout_ref.actor.use_kl_loss=true \
+ actor_rollout_ref.actor.kl_loss_coef=0.0 \
+ actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+ actor_rollout_ref.actor.entropy_coeff=0.0 \
+ actor_rollout_ref.model.path=$MODEL_PATH \
+ actor_rollout_ref.actor.strategy=fsdp \
+ actor_rollout_ref.actor.optim.lr=1e-5 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=96 \
+ actor_rollout_ref.actor.use_dynamic_bsz=true \
+ actor_rollout_ref.actor.ppo_max_token_len_per_gpu=16384 \
+ actor_rollout_ref.actor.clip_ratio_low=0.2 \
+ actor_rollout_ref.actor.clip_ratio_high=0.28 \
+ actor_rollout_ref.actor.loss_agg_mode="token-mean" \
+ actor_rollout_ref.rollout.name=vllm \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.rollout.n=8 \
+ actor_rollout_ref.ref.log_prob_use_dynamic_bsz=true \
+ actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=16384 \
+ trainer.logger=wandb \
+ trainer.project_name=$project_name \
+ trainer.experiment_name=$exp_name \
+ trainer.val_before_train=false \
+ trainer.n_gpus_per_node=6 \
+ trainer.nnodes=1 \
+ trainer.save_freq=1 \
+ trainer.test_freq=1000 \
+ trainer.total_epochs=1 2>&1 | tee verl_gsm8k.log
+```
+
+</details>
+
+<details><summary>Trinity Configuration (YAML)</summary>
+
+```yaml
+mode: both
+project: Trinity-RFT-cys-gsm8k
+group: alignment
+name: vanilla-0kl-0entropy
+checkpoint_root_dir: ./checkpoints/
+algorithm:
+  algorithm_type: grpo
+  repeat_times: 8
+  sample_strategy: default
+  kl_loss_fn: low_var_kl
+  kl_loss_fn_args:
+    kl_coef: 0.0
+  policy_loss_fn_args:
+    clip_range_low: 0.2
+    clip_range_high: 0.28
+  entropy_loss_fn: default
+  entropy_loss_fn_args:
+    entropy_coef: 0.0
+  optimizer:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    warmup_style: constant
+data_processor: {}
+model:
+  model_path: Qwen/Qwen2.5-1.5B-Instruct
+  max_prompt_tokens: 1024
+  max_response_tokens: 2048
+cluster:
+  node_num: 1
+  gpu_per_node: 8
+buffer:
+  batch_size: 96
+  total_epochs: 1
+  explorer_input:
+    taskset:
+      name: taskset
+      storage_type: file
+      path: /root/data/gsm8k
+      split: train
+      format:
+        prompt_key: question
+        response_key: answer
+      rollout_args:
+        temperature: 1.0
+        logprobs: 0
+    eval_tasksets: []
+    default_workflow_type: verl_gsm8k_workflow
+  trainer_input:
+    experience_buffer:
+      name: experience_buffer
+      storage_type: queue
+      replay_buffer:
+        enable: false
+        priority_fn: linear_decay
+        reuse_cooldown_time: null
+        priority_fn_args:
+          decay: 2.0
+explorer:
+  runner_per_model: 8
+  rollout_model:
+    engine_num: 2
+    tensor_parallel_size: 1
+    enforce_eager: true
+    enable_prefix_caching: false
+    enable_chunked_prefill: false
+    gpu_memory_utilization: 0.9
+    dtype: bfloat16
+    seed: 42
+    enable_thinking: false
+    enable_history: false
+    enable_openai_api: false
+    enable_auto_tool_choice: false
+    tool_call_parser: null
+    reasoning_parser: null
+  auxiliary_models: []
+  eval_interval: 1000
+trainer:
+  trainer_type: verl
+  save_interval: 100
+  enable_preview: true
+  grad_clip: 1.0
+  max_token_len_per_gpu: 16384
+monitor:
+  monitor_type: wandb
+synchronizer:
+  sync_method: nccl
+  sync_style: fixed
+  sync_interval: 1
+  sync_timeout: 1200
+log:
+  level: INFO
+```
+
+</details>
+
+
+#### Experiment 4: Independent Training, Non-zero Regularization
+
+<details><summary>veRL Launch Script</summary>
+
+```bash
+export MODEL_PATH=Qwen/Qwen2.5-1.5B-Instruct
+project_name=Trinity-RFT-cys-gsm8k
+exp_name=verl-unshuffle
+PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
+ algorithm.adv_estimator=grpo \
+ algorithm.use_kl_in_reward=false \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ data.train_files=$HOME/data/gsm8k/train.parquet \
+ data.val_files=$HOME/data/gsm8k/test.parquet \
+ data.train_batch_size=96 \
+ data.max_prompt_length=1024 \
+ data.max_response_length=2048 \
+ data.shuffle=false \
+ actor_rollout_ref.model.use_remove_padding=true \
+ actor_rollout_ref.actor.use_kl_loss=true \
+ actor_rollout_ref.actor.kl_loss_coef=0.001 \
+ actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+ actor_rollout_ref.actor.entropy_coeff=0.003 \
+ actor_rollout_ref.model.path=$MODEL_PATH \
+ actor_rollout_ref.actor.strategy=fsdp \
+ actor_rollout_ref.actor.optim.lr=1e-5 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=96 \
+ actor_rollout_ref.actor.use_dynamic_bsz=true \
+ actor_rollout_ref.actor.ppo_max_token_len_per_gpu=16384 \
+ actor_rollout_ref.actor.clip_ratio_low=0.2 \
+ actor_rollout_ref.actor.clip_ratio_high=0.28 \
+ actor_rollout_ref.actor.loss_agg_mode="token-mean" \
+ actor_rollout_ref.rollout.name=vllm \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.rollout.n=8 \
+ actor_rollout_ref.ref.log_prob_use_dynamic_bsz=true \
+ actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=16384 \
+ trainer.logger=wandb \
+ trainer.project_name=$project_name \
+ trainer.experiment_name=$exp_name \
+ trainer.val_before_train=false \
+ trainer.n_gpus_per_node=6 \
+ trainer.nnodes=1 \
+ trainer.save_freq=1 \
+ trainer.test_freq=1000 \
+ trainer.total_epochs=1 2>&1 | tee verl_gsm8k.log
+```
+
+</details>
+
+<details><summary>Trinity Configuration (YAML)</summary>
+
+```yaml
+mode: both
+project: Trinity-RFT-cys-gsm8k
+group: alignment
+name: vanilla
+checkpoint_root_dir: ./checkpoints/
+algorithm:
+  algorithm_type: grpo
+  repeat_times: 8
+  sample_strategy: default
+  kl_loss_fn: low_var_kl
+  kl_loss_fn_args:
+    kl_coef: 0.001
+  policy_loss_fn_args:
+    clip_range_low: 0.2
+    clip_range_high: 0.28
+  entropy_loss_fn: default
+  entropy_loss_fn_args:
+    entropy_coef: 0.003
+  optimizer:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    warmup_style: constant
+data_processor: {}
+model:
+  model_path: Qwen/Qwen2.5-1.5B-Instruct
+  max_prompt_tokens: 1024
+  max_response_tokens: 2048
+cluster:
+  node_num: 1
+  gpu_per_node: 8
+buffer:
+  batch_size: 96
+  total_epochs: 1
+  explorer_input:
+    taskset:
+      name: taskset
+      storage_type: file
+      path: /root/data/gsm8k
+      split: train
+      format:
+        prompt_key: question
+        response_key: answer
+      rollout_args:
+        temperature: 1.0
+        logprobs: 0
+    eval_tasksets: []
+    default_workflow_type: verl_gsm8k_workflow
+  trainer_input:
+    experience_buffer:
+      name: experience_buffer
+      storage_type: queue
+      replay_buffer:
+        enable: false
+        priority_fn: linear_decay
+        reuse_cooldown_time: null
+        priority_fn_args:
+          decay: 2.0
+explorer:
+  runner_per_model: 8
+  rollout_model:
+    engine_num: 2
+    tensor_parallel_size: 1
+    enforce_eager: true
+    enable_prefix_caching: false
+    enable_chunked_prefill: false
+    gpu_memory_utilization: 0.9
+    dtype: bfloat16
+    seed: 42
+    enable_thinking: false
+    enable_history: false
+    enable_openai_api: false
+    enable_auto_tool_choice: false
+    tool_call_parser: null
+    reasoning_parser: null
+  auxiliary_models: []
+  eval_interval: 1000
+trainer:
+  trainer_type: verl
+  save_interval: 100
+  enable_preview: true
+  grad_clip: 1.0
+  max_token_len_per_gpu: 16384
+monitor:
+  monitor_type: wandb
+synchronizer:
+  sync_method: nccl
+  sync_style: fixed
+  sync_interval: 1
+  sync_timeout: 1200
+log:
+  level: INFO
+```
+
+</details>
+
+---
+
+### Results
+
+The following figures compare the performance of Trinity (blue) and veRL (red).
+
+#### Experiment 1: Identical Inputs and Model States
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-1-trainer.png)
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-1-explorer.png)
+
+> ✅ The curves are nearly indistinguishable, confirming that under strictly controlled conditions—identical inputs and model states—Trinity and veRL exhibit virtually identical behavior, with only negligible floating-point discrepancies.
+
+#### Experiment 2: Aligned Model States with Independent Rollouts (vLLM Stochasticity)
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-2-trainer.png)
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-2-explorer.png)
+
+> 🔍 When only model states are synchronized, metrics such as `Reward`, `Grad Norm`, `Entropy`, and `KL loss` remain closely aligned. However, `PG loss` and `PPO KL` show slightly larger deviations—consistent with expectations, as these quantities are more sensitive to variations in rollout data.
+
+#### Experiment 3: Independent Training, Zero Regularization
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-3-trainer.png)
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-3-explorer.png)
+
+> Primary performance indicators (`Reward`, `Grad Norm`) maintain reasonable alignment, while secondary metrics (`Entropy`, `KL loss`, `PG loss`, `PPO KL`) demonstrate increased divergence due to compounded stochastic effects from both independent rollouts and optimization trajectories.
+
+
+#### Experiment 4: Independent Training, Non-zero Regularization
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-4-trainer.png)
+
+![](../../docs/sphinx_doc/assets/gsm8k-setup-4-explorer.png)
+
+> Similar to Experiment 3, fundamental metrics (`Reward`, `Grad Norm`) remain reasonably consistent between frameworks, while secondary metrics (`Entropy`, `KL loss`, `PG loss`, `PPO KL`) exhibit pronounced divergence attributable to accumulated stochastic variations in both data generation and parameter updates.
+
+---
+
+### Conclusion
+
+Our comprehensive experimental evaluation demonstrates that Trinity and veRL produce functionally equivalent results on the GSM8K benchmark when configured identically. The observed behavioral consistency under controlled conditions (Experiment 1) validates the correctness of both implementations.
+
+Divergences observed in less constrained experimental settings (Experiments 2-4) are attributable to expected sources of randomness—including vLLM sampling stochasticity and independent optimization trajectories—rather than fundamental implementation discrepancies.
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-1-explorer.png b/docs/sphinx_doc/assets/gsm8k-setup-1-explorer.png
new file mode 100644
index 0000000000..18d1e32909
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-1-explorer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-1-trainer.png b/docs/sphinx_doc/assets/gsm8k-setup-1-trainer.png
new file mode 100644
index 0000000000..dfc160ddab
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-1-trainer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-2-explorer.png b/docs/sphinx_doc/assets/gsm8k-setup-2-explorer.png
new file mode 100644
index 0000000000..a4abcbb82c
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-2-explorer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-2-trainer.png b/docs/sphinx_doc/assets/gsm8k-setup-2-trainer.png
new file mode 100644
index 0000000000..f74db2e271
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-2-trainer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-3-explorer.png b/docs/sphinx_doc/assets/gsm8k-setup-3-explorer.png
new file mode 100644
index 0000000000..4302ef2f86
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-3-explorer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-3-trainer.png b/docs/sphinx_doc/assets/gsm8k-setup-3-trainer.png
new file mode 100644
index 0000000000..8bc3dd7cd5
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-3-trainer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-4-explorer.png b/docs/sphinx_doc/assets/gsm8k-setup-4-explorer.png
new file mode 100644
index 0000000000..10c464a9e0
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-4-explorer.png differ
diff --git a/docs/sphinx_doc/assets/gsm8k-setup-4-trainer.png b/docs/sphinx_doc/assets/gsm8k-setup-4-trainer.png
new file mode 100644
index 0000000000..999d6dce4f
Binary files /dev/null and b/docs/sphinx_doc/assets/gsm8k-setup-4-trainer.png differ