HANCLIP: A Family of Hyperbolic Angular Negation Vision Language Models

A new family of vision-language models called HANCLIP has been introduced to address a persistent weakness in AI systems: the inability to reliably understand negation in images and text. The framework uses a geometry-aware training method to teach models not just what an image shows, but also what it does not. Vision-Language Models (VLMs) like CLIP are typically pre-trained on massive image-text datasets to capture semantic relationships, but they remain brittle when confronted with negated descriptions, often relying on shallow word co-occurrence rather than true understanding [1][2]. Directly fine-tuning these models on negation data can interfere with previously acquired knowledge, causing performance degradation on standard benchmarks [2]. HANCLIP, which stands for Hyperbolic + Angular + Negation, tackles this by explicitly restructuring the embedding space [2]. The model is trained on a compact set of 20,000 image-text quadruplets and combines a hyperbolic formulation to model hierarchical semantic relations with an angular triplet objective that drives systematic separation between negated descriptions and their corresponding positives [1][2]. This geometry-aware design strengthens negation sensitivity while preserving the global structure of pretrained representations, rather than overwriting them [2]. Extensive experiments show that HANCLIP delivers consistent gains on the negation-focused NegBench benchmark while maintaining competitive or improved performance on standard classification and image-text retrieval tasks [1][2]. The framework is model-agnostic and can be plugged into existing VLMs, including CLIP, LongCLIP, SmartCLIP, and HiMo-CLIP, without requiring large-scale retraining [1][2]. The research demonstrates that a carefully designed geometric objective can substantially extend the reasoning capabilities of existing VLMs using only modest additional data [2]. The work was submitted to arXiv on June 22, 2026, under the Computer Vision and Pattern Recognition category [1].