HANCLIP: New VLMs Master Negation with Hyperbolic Geometry

Vision-language models (VLMs) like CLIP often fail at understanding negation — for example, confusing "a dog without a hat" with "a dog wearing a hat." They rely on shallow word co-occurrence and are easily misled by negative cues, degrading performance on tasks requiring precise reasoning. To address this, a team of researchers from multiple institutions has developed HANCLIP (Hyperbolic + Angular + Negation), a family of VLMs that explicitly encodes both what an image contains and what it excludes. The approach uses a hyperbolic formulation to model hierarchical semantic relationships and an angular triplet objective that forces systematic separation between negated and positive descriptions. Remarkably, the method requires only a compact dataset of 20,000 image-text quadruplets, making it highly data-efficient while maintaining the global structure of pretrained embeddings.

HANCLIP achieves consistent gains on the negation-focused NegBench benchmark without harming performance on standard classification and image-text retrieval tasks. It is model-agnostic and can be integrated into CLIP, LongCLIP, SmartCLIP, and HiMo-CLIP without large-scale retraining, demonstrating that a carefully designed geometric objective can substantially extend reasoning capabilities with minimal additional data. This work highlights a promising direction for improving VLM robustness to linguistic nuance, opening the door to more reliable image understanding in real-world applications such as search, captioning, and visual question answering.