ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation

<a href="https://arxiv.org/abs/2506.18810"><img src="https://img.shields.io/badge/arXiv%20paper-2506.18810-b31b1b.svg" alt="arXiv"></a>

ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation <br> Siao Tang, Xinyin Ma, Gongfan Fang, and Xinchao Wang <br> xML Lab, National University of Singapore <br>

We propose a framework dubbed <strong>ConciseHint</strong>, which continuously encourages the reasoning model to speak concisely by injecting the textual hint (manually designed or learned on the concise data) during the token generation of the reasoning process.

🔧Setup

conda create -n concise_hint python=3.10 
conda activate concise_hint
pip install -r requirements.txt

🚀Run

1. Start the LLM server

CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen3-4B --enable-reasoning \
--reasoning-parser deepseek_r1  --enable-prefix-caching --max_model_len 25000  --port 8000

2.1 Run concise hint

python -u run_main.py --model  Qwen/Qwen3-4B  --dataset gsm8k --repeat_exp_num 5  --max-tokens 10240  --bs 64  --port 8000 \
--enable_adap --enable_hint --exp Qwen3-4B-gsm8k-ours

2.2 Run baseline

python -u run_main.py --model  Qwen/Qwen3-4B  --dataset gsm8k --repeat_exp_num 5  --max-tokens 10240  --bs 64  --port 8000 --exp Qwen3-4B-gsm8k-base

Main results

<div align="center"> <img src="./assets/table11.png" width="75%" /> <img src="./assets/table22.png" width="75%" /> </div>

Note

This page only includes the codes of the training-free ConciseHint. Please move to ConciseHint-T directory for the ConciseHint-T codes.

Acknowledgement

The code framework is modified from the open-source github repository of Dynasor.