---
language:
- zh
- en
pipeline_tag: text-generation
tags:
- deepscaler
- grpo
- qwen2
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
license: other
library_name: transformers
---

# DECS_7B

This is the official model for ICLR 2026 Oral "Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling".
DECS_7B is a reasoning-focused causal language model built from `deepseek-ai/DeepSeek-R1-Distill-Qwen-7B` and further trained with DECS algorithm, focused on 50% fewer tokens when answering a reasoning-required problem.

## Model Summary

- Base model: `deepseek-ai/DeepSeek-R1-Distill-Qwen-7B`
- Upload date: `2026-02-24`
- Recommended use: long-form reasoning and mathematical/problem-solving style generation
- Paper link: https://arxiv.org/pdf/2509.25827
- Project page: https://pixas.github.io/decs-iclr26-site/
- Github repo: https://github.com/pixas/DECS

## Quick Start (Transformers)

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "pixas/DECS_7B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Solve: If x^2 - 5x + 6 = 0, what are x values?"}
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.6,
        top_p=0.95,
    )

new_tokens = outputs[0][inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
```

## Quick Start (vLLM)

```python
from vllm import LLM, SamplingParams

llm = LLM(model="pixas/DECS_7B", trust_remote_code=True)
sampling = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=512)
prompt = "Please reason step by step: what is 37 * 48?"
outputs = llm.generate([prompt], sampling_params=sampling)
print(outputs[0].outputs[0].text)
```

## Notes

- This model may produce incorrect or unverifiable reasoning. Always validate outputs in high-stakes settings.
- Performance can vary by prompt style and decoding parameters.
- License and acceptable-use constraints should follow the upstream base model and your deployment policy.



## Citation

If you use this model, please cite our paper:
```bibtex
@inproceedings{jiang2026overthinking,
title={Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling},
author={Shuyang Jiang and Yusheng Liao and Ya Zhang and Yanfeng Wang and Yu Wang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=kdeiRledV6}
}
```