Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention

Zhen Yang¹ · Mingyang Zhang⁵ · Feng Chen³ · Ganggui Ding⁴
Liang Hou² · Xin Tao² · Pengfei Wan² · Ying-Cong Chen^1,6
1HKUST(GZ) · ²Kuaishou Technology · ³AIML · ⁴ZJU · ⁵Ant Group · ⁶HKUST

Getting Started

1. Create the environment and install the dependencies by running:

conda create -n MTI python=3.10
conda activate MTI
pip install vllm==0.10.2
pip install accelerate==1.10.1
pip install transformers==4.56.1

2. Run offline with vllm

python run_vllm_offline.py

3. Run online with vllm

python run_vllm_online.py \
  --model Qwen3/Qwen3-8B \
  --tokenizer Qwen3/Qwen3-8B \
  --host 0.0.0.0 \
  --port 6666 \
  --api-key yzisallyouneed

4. Run with huggingface

python run_hf.py

The huggingface version is only for learning purposes because it does not provide inference acceleration. It is recommended to use the vllm version for evaluation.

Evaluation

Since the monkey patch may introduce unknown bugs, we recommend that, during the evaluation phase, you directly replace vllm’s GPUModelRunner.execute_model and FlashAttentionImpl.forward in the Conda-created virtual environment with our execute_model and forward implementations.

TODO

1. SGLang version
2. More models
3. Code combine opencompass
4. MTI on VLM / VLA / dllm

Method

Main results

Acknowlegment

Many thanks for the generous help from Lequan Lin.

BibTeX

@article{yang2025less,
  title={Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention},
  author={Yang, Zhen and Zhang, Mingyang and Chen, Feng and Ding, Ganggui and Hou, Liang and Tao, Xin and Wan, Pengfei and Chen, Ying-Cong},
  journal={arXiv preprint arXiv:2510.13940},
  year={2025}
}