A new paper investigates the effectiveness of split-conformal prediction as a safety layer for zero-shot vision-language models (VLMs) under shifting data conditions. The research found that while marginal coverage can remain high, class-conditional coverage can significantly degrade, with some classes experiencing near-zero coverage even when overall coverage is around 86%. Various calibration techniques were tested, with target-side class calibration showing the most promise but requiring extensive labeled data. The study concludes that marginal conformal coverage should be viewed as an average reliability metric rather than a definitive safety guarantee for specific classes. AI
IMPACT Highlights potential safety gaps in current VLM calibration methods, suggesting a need for more robust class-conditional safety guarantees.
RANK_REASON The item is a research paper detailing findings on the safety of a specific machine learning technique. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
