PRISM Gemini Distill

Overview

PRISM Gemini Distill is our self-distilled multimodal reasoning dataset collected from Gemini 3 Flash for the PRISM project.

PRISM studies the distributional drift problem in the standard SFT → RLVR post-training pipeline. To mitigate this issue, PRISM introduces an intermediate Distribution Alignment / Pre-alignment stage before RLVR:

SFT → Distribution Alignment / Pre-alignment → RLVR

This dataset provides high-quality Gemini 3 Flash demonstrations with dense visual grounding and step-by-step reasoning. It serves as the high-fidelity supervision source for PRISM's black-box on-policy distillation and distribution alignment.

Dataset Description

This dataset contains approximately 108K self-distilled multimodal reasoning examples from Gemini 3 Flash.

Each example typically includes:

• an input image;
• a question or instruction;
• a detailed visual description focusing on instruction-relevant evidence;
• step-by-step reasoning;
• a final answer.

The dataset is designed to provide a stronger target distribution for challenging multimodal reasoning problems, especially those requiring both precise perception and coherent reasoning.

Role in PRISM

In PRISM, large-scale public demonstrations are first used for broad SFT initialization. The Gemini-distilled data in this repository is then used as the high-quality reference distribution for the pre-alignment stage.

During this stage, PRISM performs black-box adversarial on-policy distillation. The policy generates responses on-policy, while a Mixture-of-Experts discriminator provides response-level corrective signals:

• Perception Expert: evaluates instruction-relevant visual grounding.
• Reasoning Expert: evaluates reasoning faithfulness and consistency.

This enables PRISM to improve multimodal reasoning behavior without requiring access to teacher logits.

Intended Use

This dataset is intended for research on:

• multimodal reasoning;
• vision-language model post-training;
• black-box distillation;
• on-policy distillation;
• distribution alignment before RLVR;
• reinforcement learning for multimodal models.

Dataset Statistics

Usage

For data processing, training, inference, and evaluation, please refer to our GitHub repository:

https://github.com/XIAO4579/PRISM

Citation

If you find PRISM or this dataset useful, please cite:

@misc{wang2026sfttorlprealignmentblackboxonpolicy,
      title={Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL}, 
      author={Sudong Wang and Weiquan Huang and Xiaomin Yu and Zuhao Yang and Hehai Lin and Keming Wu and Chaojun Xiao and Chen Chen and Wenxuan Wang and Beier Zhu and Yunjian Zhang and Chengwei Qin},
      year={2026},
      eprint={2604.28123},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.28123}, 
}

Acknowledgements

We gratefully acknowledge LLaMA-Factory, verl, and lmms-eval for their excellent open-source infrastructure.