Phi-4-reasoning-vision-15B Technical Report
arXiv:2603.03975v1 Announce Type: new Abstract: We present Phi-4-reasoning-vision-15B, a compact open-weight multimodal reasoning model, and share the motivations, design choices, experiments, and learnings that informed its development. Our goal is to contribute practical insight to the research community on building smaller, efficient multimodal reasoning models and to share the result of these learnings as an open-weight model that is good at common vision and language tasks and excels at scientific and mathematical reasoning and understanding user interfaces. Our contributions include demonstrating that careful architecture choices and rigorous data curation enable smaller, open-weight multimodal models to achieve competitive performance with significantly less training and inference-time compute and tokens. The most substantial improvements come from systematic filtering, error correction, and synthetic augmentation -- reinforcing that data quality remains the primary lever for m
arXiv:2603.03975v1 Announce Type: new Abstract: We present Phi-4-reasoning-vision-15B, a compact open-weight multimodal reasoning model, and share the motivations, design choices, experiments, and learnings that informed its development. Our goal is to contribute practical insight to the research community on building smaller, efficient multimodal reasoning models and to share the result of these learnings as an open-weight model that is good at common vision and language tasks and excels at scientific and mathematical reasoning and understanding user interfaces. Our contributions include demonstrating that careful architecture choices and rigorous data curation enable smaller, open-weight multimodal models to achieve competitive performance with significantly less training and inference-time compute and tokens. The most substantial improvements come from systematic filtering, error correction, and synthetic augmentation -- reinforcing that data quality remains the primary lever for model performance. Systematic ablations show that high-resolution, dynamic-resolution encoders yield consistent improvements, as accurate perception is a prerequisite for high-quality reasoning. Finally, a hybrid mix of reasoning and non-reasoning data with explicit mode tokens allows a single model to deliver fast direct answers for simpler tasks and chain-of-thought reasoning for complex problems.
Executive Summary
The Phi-4-reasoning-vision-15B technical report presents a compact open-weight multimodal reasoning model, demonstrating that smaller models can achieve competitive performance with less training and inference-time compute. The report highlights the importance of careful architecture choices, rigorous data curation, and systematic filtering, error correction, and synthetic augmentation. The model excels at scientific and mathematical reasoning and understanding user interfaces, and its hybrid mix of reasoning and non-reasoning data enables fast direct answers and chain-of-thought reasoning.
Key Points
- ▸ Compact open-weight multimodal reasoning model
- ▸ Smaller models can achieve competitive performance with less compute
- ▸ Importance of data quality and curation for model performance
Merits
Efficient Model Design
The model's compact design and careful architecture choices enable efficient performance with significantly less training and inference-time compute.
Demerits
Dependence on High-Quality Data
The model's performance is heavily dependent on high-quality data, which can be time-consuming and resource-intensive to curate.
Expert Commentary
The Phi-4-reasoning-vision-15B model represents a significant advancement in multimodal reasoning, demonstrating that smaller models can achieve competitive performance with careful design and data curation. The report's emphasis on data quality and efficient model design highlights the need for a more nuanced approach to AI development, one that prioritizes explainability, transparency, and efficiency. As the field continues to evolve, it is likely that we will see increased focus on developing models that can balance performance with computational resources and data requirements.
Recommendations
- ✓ Further research on efficient model design and data curation techniques
- ✓ Development of more explainable and transparent AI models