DeepIntuit

Model Description

DeepIntuit is a reasoning-enhanced video understanding model designed for open-instance video classification. Instead of directly mapping visual features to labels, the model learns to generate intrinsic reasoning traces that guide the final classification decision, improving robustness under large intra-class variation.

The model is introduced in:

From Imitation to Intuition: Intrinsic Reasoning for Open-Instance Video Classification 📄 Paper: https://arxiv.org/abs/2603.10300 💻 Code: https://github.com/BWGZK-keke/DeepIntuit

---

Training Pipeline

DeepIntuit is trained through a three-stage pipeline:

1. Cold Start Alignment Supervised training to initialize structured reasoning generation.

2. Reasoning Refinement (GRPO) Reinforcement learning improves reasoning quality and prediction consistency.

3. Intuitive Calibration A lightweight classifier is trained on generated reasoning traces for stable prediction.

---

Intended Use

DeepIntuit is designed for research on:

• video understanding
• open-instance video classification
• reasoning-enhanced multimodal learning
• safety-sensitive video analysis

Citation

@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},
  year={2026}
}