
LT
Lakpa Tamang, Mohamed Reda Bouadjenek, Saman Forouzandeh, Kamal Berahmand, Mahdi Jalili, Sunil Aryal
· 1 min read
ResearcharXiv cs.CV
Beyond Raw Similarity: Prior-Aware Fusion for Out-of-Distribution Detection
arXiv:2611.18692v1 Announce Type: new
Abstract: Out-of-distribution (OOD) detection in vision-language models presents unique challenges due to the reliance on cross-modal similarity. Even when embeddings are well separated for in-distribution (ID) data, similarity scores can remain overconfident for OOD inputs. A common remedy is to adapt the model on ID data, but adaptation discards the pretrained geometry that made the similarity meaningful. We propose multi-modal logit fusion (MM-Log), which keeps both by pairing a gradient-free zero-shot prior with a task-adapted branch in each modality and fuses their four cross-modal similarity channels into a single OOD score. Because the zero-shot priors are never adapted, the cross-channels act as a consistency check between the pretrained and adapted geometries: an OOD input that fools the adapted branch through a shortcut must additionally agree with frozen CLIP to retain a high score. We formalize this as a margin bound showing that the fused score is provably penalized by disagreement between the zero-shot and adapted branches, a signal to which adapted-only scores are structurally blind. MM-Log requires no auxiliary OOD data, negative prompts, or test-time adaptation, and trains on few-shot ID data alone. Across Far-OOD, Near-OOD, and Hard-OOD benchmarks, MM-Log consistently outperforms strong CLIP-based baselines.
Original source
This story was published by arXiv cs.CV and written by Lakpa Tamang, Mohamed Reda Bouadjenek, Saman Forouzandeh, Kamal Berahmand, Mahdi Jalili, Sunil Aryal. SyncAI.news shows a preview; the complete article is on the publisher's site.
Read the full story on arxiv.org


