Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Analyzing and Improving Fine-grained Preference Optimization in Medical LVLMs
TL;DR
Large Vision-Language Models (LVLMs) in medical imaging struggle with factual inaccuracies and poor alignment with clinical needs. A new method was developed that uses a bidirectional token-wise KL regularizer and a visual-contrastive grounding objective to enhance preference optimization.
✦ Why It Matters
Engineers can leverage this method to enhance the accuracy of AI models in medical applications.
Key Takeaways
- Implement bidirectional KL regularization for model outputs.
- Use visual-contrastive objectives to improve grounding.
- Focus on fine-grained preference optimization in medical AI.
- Test models extensively on clinical benchmarks for validation.
Full Summary
Large Vision-Language Models (LVLMs) have shown promise in medical imaging but often produce factually incorrect outputs and lack alignment with clinical requirements. Existing methods like Direct Preference Optimization (DPO) fail due to treating critical medical tokens the same as irrelevant text, relying on static references that can mislead the model, and lacking visual grounding.
The proposed method introduces a bidirectional token-wise Kullback-Leibler (KL) regularizer and a visual-contrastive grounding objective, which penalizes outputs that do not adequately reflect visual evidence from paired images. This fine-grained, on-policy alignment framework corrects only clinically erroneous parts of the model's outputs while maintaining their original style.
Extensive testing demonstrated that this approach significantly enhances the accuracy and reliability of LVLMs in medical contexts, leading to better clinical text generation and image analysis. These findings suggest a pathway for improving AI tools in healthcare, ensuring they meet clinical standards.