Conan: Progressive Learning to Reason Like a Detective over Multi-Scale Visual Evidence

🔥 Updates

🚀 February 20, 2026 - Conan got accepted by CVPR 2026.

🚀 November 17, 2025 — We release the Conan-91k dataset!

🚀 October 25, 2025 — We release the training framework of Conan!

🚀 October 23, 2025 — We release the paper!

🚀 October 20, 2025 — We are excited to release Conan-7B and its accompanying evaluation toolkit, Conan-Eval!

🚀 September 30, 2025 — Conan-SFT-7B has officially landed on Hugging Face!

Introduction

Conan is an innovative Multimodal Large Language Model (MLLM) designed with advanced reasoning capabilities inspired by a detective's investigative process. It excels in:

1. Identifying multi-scale frames of visual evidence.
2. Reasoning over cross-frame clues to connect information.
3. Deciding plausible actions based on its deductions.

🏆 Performance Highlights Here's a glimpse of Conan's impressive capabilities:

⚙️ Environment Setup

• Clone the Repository:

git clone https://github.com/OuyangKun10/Conan.git
cd Conan

• Create and Activate Environment:

conda create --name Conan python=3.10
conda activate Conan

• Install Dependencies:

cd ms-swift

pip install -e .

🏋️ Training

cd ms-swift/training_scripts

1. Texual Reasoning

bash Conan-SFT-Stage1.sh

2. Multimodal alignment Reasoning

bash Conan-SFT-Stage2.sh

3. Vision-centric Reasoning

bash Conan-SFT-Stage3.sh

4. AIR RLVR

bash Conan-server.sh
bash Conan-AIR-RLVR.sh

📊 Evaluation

Conan-Eval toolkit allows for comprehensive evaluation across various benchmarks:

1. Multi-step Reasoning Benchmarks:

   • MMR-V
   • Video-Holmes
   • VRBench
   • VCRBench
   • LongVideoReason
   • HumanPCR

2. Long-video Understanding Benchmarks:

   • LongVideoBench
   • MLVU
   • LVBench
   • Video-MME

3. Usage: To run the evaluation, simply execute:

   bash run.sh

Acknowledgments

We extend our sincere gratitude to the following projects for their invaluable contributions and inspiration:

• ms-swift
• Video-R1
• Generative-sampler