Disentangling Hallucinations: Orthogonal Semantic Projection for Robust Interpretability

A team of researchers has proposed a geometric method to reduce semantic hallucinations in Vision-Language Models, a persistent flaw where AI explanation tools highlight incorrect image regions. The work, submitted to arXiv on 8 Jun 2026, introduces a framework called Orthogonal Semantic Projection to disentangle visual concepts [1]. The study addresses a known weakness in Explainable AI (XAI) for systems like CLIP, where attribution maps can highlight a dog when the model is prompted with the word "cat" [1]. The authors argue this phenomenon, termed semantic hallucination, is not a quirk of a single architecture but a fundamental consequence of what they call Linear Semantic Leakage in high-dimensional embedding spaces [1]. A formal mathematical analysis of this problem within CLIP embeddings has been largely absent from the literature until now [2]. The researchers propose a unified theoretical framework, Linear Semantic Attribution (LSA), which generalizes across discriminative methods [1]. Building on this, they introduce Orthogonal Semantic Projection (OSP), a geometric intervention that uses the residual property of Orthogonal Matching Pursuit (OMP) to separate unique semantic signals from shared concepts [1]. The technique works by orthogonalizing the query vector against distractor concepts, making the attribution model blind to shared features while preserving fidelity for correct prompts [2]. The paper was posted on arXiv, an open-access repository for electronic preprints that has hosted over two million articles since its founding in 1991 and currently receives about 24,000 submissions per month [6]. The platform is not peer-reviewed, but it serves as a primary distribution channel in fields like computer science and physics [6]. The authors have made their code publicly available through a linked repository [1]. The research arrives as Vision-Language Models are increasingly deployed in safety-critical applications, raising the stakes for trustworthy explanations [1]. The arXiv page for the paper also features experimental community tools under the arXivLabs framework, a program launched in 2020 that allows collaborators to build features like citation explorers and code finders directly on the site [4][5]. These tools operate under guidelines that prioritize openness and user data privacy, with collaborators receiving only minimal, anonymized data [4].