---
base_model:
- Qwen/Qwen3-VL-8B-Instruct
datasets:
- InterleaveThinker/Train-Data
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
---

# InterleaveThinker-Critic-SFT Model

This repository contains the InterleaveThinker-Critic-SFT-8B model presented in [InterleaveThinker: Reinforcing Agentic Interleaved Generation]().

[**Project Page**](https://zhengdian1.github.io/InterleaveThinker-proj/) | [**GitHub Repository**](https://github.com/zhengdian1/InterleaveThinker) | [**Paper**](https://arxiv.org/pdf/2606.13679)

# 👀 Intro

<div align="center">
  <img src="https://github.com/zhengdian1/InterleaveThinker/blob/main/assets/teaser.jpg?raw=true" alt="InterleaveThinker Teaser" width="80%">
</div>

We introduce **InterleaveThinker**, as the first multi-agent pipeline designed to **endow any existing image generator with interleaved generation capabilities**. InterleaveThinker can organize the image-text input sequence via a planner agent, evaluate generator outputs, identify deviations, and refine instructions via a critic agent, **enabling complex interleaved text-image sequence generation for visual narratives, guidance, embodied manipulation and long-horizon sub-task annotation.**

We build three dedicated training datasets—Interleave-Planner-SFT-80k, Interleave-Critic-SFT-112k, and Interleave-Critic-RL-13k—for interleaved generation and step-wise instruction correction using GRPO with proposed accuracy and step-wise rewards.

InterleaveThinker achieves **performance comparable to Nano Banana and GPT-5 on interleaved generation benchmarks**, delivering substantial gains on reasoning-based benchmarks (e.g., boosting WISE from 0.47 to 0.73 and RISE from 13.3 to 28.9 on 4-step FLUX.2-klein). It also demonstrates strong transferability, improving performance across various existing image generators.


## 🎥 Demo

#### Inference Process Example

<div align="center">
  <img src="https://github.com/zhengdian1/InterleaveThinker/blob/main/assets/example.png?raw=true" alt="Inference Process Example" width="85%">
</div>

For more examples, please refer to our website [[🌐Project Page]](https://zhengdian1.github.io/InterleaveThinker-proj/) 

## 🚀 Training and Inference

For detailed instructions on setup, SFT/RL training, and inference, please refer to the [official GitHub repository](https://github.com/zhengdian1/InterleaveThinker).

## 📐 Citation

If you find our work helpful for your research, please consider citing our work:

```bibtex
@article{zheng2026interleavethinker,
  title={InterleaveThinker: Reinforcing Agentic Interleaved Generation},
  author={Zheng, Dian and Lee, Harry and Zhang, Manyuan and Feng, Kaituo and Guo, Zoey and Zhang, Ray and Li, Hongsheng},
  journal={arXiv preprint arXiv:2606.13679},
  year={2026}
}
```