Debiased Negative Mining Improves Out-of-distribution Detection with Pre-trained Vision-Language Models

Researchers tackle a fundamental weakness in vision-language model based out-of-distribution detection: the false negative problem in negative label mining. Current methods rely on heuristic rules to identify semantically dissimilar labels from unlabeled data, but this approach fails to capture the full spectrum of potential OOD inputs. The paper proposes debiased negative mining to improve detection reliability, directly addressing a bottleneck in deploying VLMs for safety-critical applications where unexpected inputs must be reliably flagged. This work matters for practitioners building robust ML systems that depend on VLM-based anomaly detection.

Modelwire context

Explainer

The paper identifies that heuristic-based negative label mining doesn't just miss some OOD cases; it systematically biases which dissimilar classes get selected for training, creating a false sense of coverage. Debiased mining corrects for this selection bias rather than just adding more negatives.

This work sits in a broader pattern visible across recent papers on model robustness and bias. Like the geopolitical bias study from earlier this week, which showed that alignment procedures encode hidden assumptions into model behavior, this research reveals that OOD detection pipelines carry their own latent biases baked into the negative sampling stage. The difference is scope: where that LLM study examined post-training alignment across labs, this focuses on a single detection bottleneck. Both suggest that practitioners can't assume their training procedures are neutral; they must audit the selection mechanisms themselves.

If this debiased mining approach shows consistent gains across three or more public OOD benchmarks (CIFAR-10-OOD, ImageNet-OOD, and at least one domain-specific set) when compared head-to-head against the heuristic baseline on the same VLM backbone, the method has moved beyond a single-dataset fix. If gains disappear or flip on held-out test sets not seen during negative mining, that signals the debiasing is just redistributing the bias rather than removing it.

Coverage we drew on

It's the humans, not the data: Geopolitical bias in LLMs originates in post-training, amplified by the language of the prompt · arXiv cs.LG

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsVision-Language Models · Out-of-distribution Detection

Read full story at arXiv cs.LG →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.