Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 7·5 min read
Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models
TL;DR
KoNA, a new benchmark for evaluating selective non-compliance in vision-language models (VLMs), reveals that these models often struggle to appropriately refuse or correct mixed queries. Fine-tuning VLMs with KoNA examples significantly improves their non-compliance accuracy while maintaining performance on fully answerable tasks.
✦ Why It Matters
Implement selective non-compliance training in your VLMs to enhance their response accuracy.
Key Takeaways
- Utilize the KoNA benchmark to evaluate your VLMs' non-compliance capabilities.
- Incorporate fine-tuning with mixed queries to improve model responses.
- Focus on component-level non-compliance for better query handling.
- Assess VLM performance across diverse query types for comprehensive evaluation.
How It Works
KoNA evaluates VLMs by presenting queries that contain both answerable and non-compliant components. This dual assessment allows for a nuanced understanding of a model's ability to selectively comply or refuse parts of a query.