pipeline_tag: video-classification
library_name: transformers
tags:
- video-reasoning
- VLM
- reinforcement-learning
DeepIntuit
DeepIntuit is a progressive framework for open-instance video classification that evolves models from simple feature imitation to intrinsic reasoning.
- Paper: From Imitation to Intuition: Intrinsic Reasoning for Open-Instance Video Classification
- Repository: BWGZK-keke/DeepIntuit
- Project Page: https://bwgzk-keke.github.io/DeepIntuit/
Model Description
DeepIntuit bridges the gap between traditional video encoders and the generalization capabilities of vision-language models (VLMs). Instead of directly predicting labels from visual features, it utilizes a three-stage reasoning pipeline:
- Cold-start supervised alignment: Initializes reasoning capability using supervised traces generated by a teacher model.
- Intrinsic reasoning refinement (Stage 1): Refines the reasoning ability using Group Relative Policy Optimization (GRPO) reinforcement learning to enhance coherence.
- Intuitive calibration (Stage 2): Trains a classifier on the intrinsic reasoning traces to ensure stable knowledge transfer and accurate classification results.
This approach decouples reasoning generation from final decision-making, significantly improving robustness in scenarios with vast intra-class variations.
Installation
The repository contains separate environments for each stage. For inference using the final model, set up the stage 2 environment:
git clone https://github.com/BWGZK-keke/DeepIntuit.git
cd DeepIntuit/stage2_model
pip install -r requirements.txt
Sample Usage
After setting up the environment, you can run inference using the following command provided in the official repository:
cd stage2_model
python inference.py \
--model_path BWGZK/DeepIntuit \
--video_path path_to_your_video.mp4
Citation
If you find this work useful, please cite:
@article{zhang2026deepintuit,
title={From Imitation to Intuition: Intrinsic Reasoning for Open-Instance Video Classification},
author={Zhang, Ke and Zhao, Xiangchen and Tian, Yunjie and Zheng, Jiayu and Patel, Vishal M and Fu, Di},
journal={arXiv preprint arXiv:2603.10300},
year={2026}
}