Skip to content
Open access

Calibrating Vision-Language Models for Out-of-Distribution Detection

2026 · IEEE Signal Processing Letters · Vol 33, pp. 4136-4140 · 0 citations · 29 references

TL;DR

This work proposes Domain-Context Prompting (DCP), a zero-shot framework that calibrates the full set of ID textual prototypes with shared domain semantics and demonstrates the effectiveness of DCP in the zero-shot setting and further improvements from its few-shot extension.

Abstract

Vision–language models (VLMs) have shown strong potential for out-of-distribution (OOD) detection by matching test images with textual prototypes derived from in-distribution (ID) class names. However, generic class prompts mainly encode object identity and leave the dataset-level visual context implicit, which may cause semantic underspecification in CLIP-based OOD scoring. We address this issue from the perspective of Domain Calibration and propose Domain-Context Prompting (DCP), a zero-shot framework that calibrates the full set of ID textual prototypes with shared domain semantics. DCP obtains a dataset-level domain anchor through an LLM-assisted candidate generation and OOD-free semantic consensus selection procedure, and injects the selected anchor into class-conditioned prompts. The resulting OOD score is still computed by maximum cosine similarity over ID classes, but each prototype encodes joint class-and-domain information. Under additional few-shot ID supervision, the same calibrated prototype space can be optionally refined with lightweight Low-Rank Adaptation (LoRA), yielding DCP+. Experiments on standard ImageNet-1 K OOD benchmarks demonstrate the effectiveness of DCP in the zero-shot setting and further improvements from its few-shot extension.

Read PDF

Similar papers

Conference Open access Sep 2026

Inductive Visual Logic for Few-Shot Out-Of-Distribution Adaptation in VLMs

Generative vision-language models (VLMs) such as Qwen-VL and LLaVA achieve strong zero-shot performance on tasks overlapping with their pretraining distribution, yet fail on specialized domains where the required discriminative features were never learned, a regime we term distant out-of-distribution (OOD). Standard ad...

Hung-Jen Chen, Yuek F. Ho, Ting-Yao Huang et al. · 0 citations
Open access Sep 2026

Protodetect: Prototype Compactness and Inference Calibration for Few-Shot Out-of-Distribution Detection

Detecting out-of-distribution (OOD) samples from limited labeled data is important for reliable recognition under semantic novelty. Existing few-shot approaches use synthetic or auxiliary unknowns, model only in-distribution (ID) data, or rely on large pretrained vision–language models. This paper presents Protodetect,...

Ze-Xia Huang, Hao-Yu Jiang, Jin-Song Hu et al. · 0 citations

Domain Recentering and Confidence-Weighted Prior Calibration for Vision-Language Models

Vision-language models such as CLIP achieve strong zero-shot classification, yet under distribution shift, visual embeddings drift from fixed text embeddings. Training-free calibration avoids the per-sample optimization of prompt learning, but prior feature calibration gives each image the full bias of one hard cluster...

Youngeun Seol, Jiyon Shin, Hee Suk Yoon et al. · 0 citations
Preprint Nov 2026

Beyond Raw Similarity: Prior-Aware Fusion for Out-of-Distribution Detection

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...

L. Tamang, Mohamed Reda Bouadjenek, Saman Forouzandeh et al. · 0 citations
Preprint Oct 2026

A Theoretical Analysis of Environment-Aware Out-of-Distribution Detection for Vision-Language Models

Continual vision-language models (VLMs) encounter a stream of environments and must also reject inputs that do not belong to the environment at hand. Class information from earlier environments can aid out-of-distribution (OOD) detection, yet existing detectors score each environment in isolation, and when previous env...

S. Sekeh, M. Wisell · 0 citations
Book Open access Nov 2026

VISRA: Vision-Language Induced Supervision and Representation Alignment for Long-Tailed Generalized Category Discovery

This paper proposes VISRA, a semantic-driven framework for LT-GCD built upon a parametric GCD pipeline that leverages Vision-Language Models to provide additional semantic supervision for representative novel-class instances and guide visual representation learning.

Yi-Zhong Zhang, Hong-Zhe Shou, Guan-Yu Lu et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.