README.md

---
license: apache-2.0
pipeline_tag: text-generation
---
# Model Card for Qwen3-4B-ECHO-Sokoban-GRPO

<!-- Provide a quick summary of what the model is/does. -->

Building upon Qwen3-4B, we trained the model with the ECHO framework using GRPO on the Sokoban dataset. 
Specifically, because Qwen3-4B performs poorly on the more challenging Sokoban puzzles, we adopted a multi-round RL training regimen capped at four rounds, with a maximum of 25 candidate actions per round. The detailed environment configuration is as follows:


```python
  LargerSokoban6:
    env_type: sokoban
    max_actions_per_traj: 100
    env_instruction: "You are solving the Sokoban puzzle. You are the player and you need to push all boxes to targets. When you are right next to a box, you can push it by moving in the same direction. You cannot push a box through a wall, and you cannot pull a box. The answer should be a sequence of actions, like <answer>Right || Right || Up</answer>"
    max_tokens: 300
    env_config:
      dim_x: 6
      dim_y: 6
      num_boxes: 2
      max_steps: 300
      search_depth: 20
```
Tabel 1: Model performance on Sokoban task
|  Model    | Success Rate(%) |
|----------------|----------------|
| Qwen3-4B  | 21.8                    |
| Qwen3-4B-Echo(GRPO) | 34.0                 | 
| Qwen3-30B-A3B-Thinking-2507  | 72.75                     | 
| Qwen3-30B-A3B-Thinking-2507-Echo(GRPO)  | 82.80                      | 
| Deepseek-R1  | 75.75         |
| Qwen3-235B-A22B-Thinking-2507)  | 79.68         |
| gpt-oss-120b  | 79.69         |


# Quick start
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "GradientResearch/Qwen3-4B-ECHO-Sokoban-GRPO"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "sokoban"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

```

# Citation
If you find our work helpful, feel free to give us a cite.
```
@misc{xiao2025echodecouplinginferencetraining,
      title={Echo: Decoupling Inference and Training for Large-Scale RL Alignment on Heterogeneous Swarms}, 
      author={Jie Xiao and Changyuan Fan and Qingnan Ren and Alfred Long and Yuchen Zhang and Rymon Yu and Eric Yang and Lynn Ai and Shaoduo Gan},
      year={2025},
      eprint={2508.05387},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.05387}, 
}
```