A new research paper titled "Level, Sharpness, and Corpus: Why Zero-Shot OOD Detector Rankings Do Not Transfer" by Ignacio Meza De la Jara explores the unreliability of zero-shot out-of-distribution (OOD) detectors. The study reveals that detector rankings can reverse across different domains and that performance is highly dependent on the in-distribution data and the underlying vision-language model (VLM). The research identifies that corpus-free detectors rely on absolute match level and relative or spatial sharpness, while WordNet-based methods also depend on external semantic coverage. To address this, the paper introduces the Complementary Evidence Guard (CEG), a detector-agnostic wrapper that improves performance by fusing base detector, level, and sharpness information without requiring OOD samples or auxiliary corpora. AI
IMPACT Challenges the assumption that benchmark rankings for zero-shot OOD detectors reliably transfer across domains, suggesting a need for more robust evaluation methods.
RANK_REASON Research paper published on arXiv detailing findings about OOD detectors. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Complementary Evidence Guard
- GL-MCM
- Ignacio Meza De la Jara
- Level, Sharpness, and Corpus: Why Zero-Shot OOD Detector Rankings Do Not Transfer
- WordNet
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
