The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models

Auxiliary signal pathways in VLMs are routinely fitted with learnable gates so the optimiser can decide how much of the signal to admit. We find that the optimiser almost always decides on zero: across five injection designs, every gated pathway becomes behaviourally closed, with accuracy invariant to ablating the pathway at inference even when the gate parameter would nominally pass 30-45% of the signal. We attribute this suppression phenomenon to two regimes, a dead-gradient regime formalised through the caption-invariance of image-derived signals, and a negative-utility regime in which the auxiliary signal actively hurts the loss. Rather than fight suppression, we exploit it: we regularise LoRA fine-tuning with geometric auxiliary losses from hyperbolic visual relational graphs (IoA-driven entailment cones and angular repulsion on the Lorentz manifold), coupled only through the forward pass at training time and dropped at inference. Disaggregating GQA by question type exposes a clean dissociation. Three configurations without geometric losses at inference lose 2.85-3.39pp on relational questions while gaining ~1pp on attribute questions; a fourth that trains with the losses but infers through a soft prompt loses 5.14pp on rel for only +0.23pp on attr, so training-time regularisation alone does not protect relational accuracy without a geometric inference pathway. Configurations that keep the geometric pathway at inference preserve vanilla-level relational accuracy and match the attribute gain. Out of distribution on VSR, the RMS-prefix recipe preserves the spatial signal; stripping the geometric losses (G2) collapses VSR by 4.6pp, isolating them as the OOD source. A secondary result: embedding-norm alignment is necessary for generation-safe prefix injection, and learnable gates should be replaced with fixed, non-optional injection at matched scales.

Learning Taxonomic Trees with Hierarchical Representation Regularization for Large Multimodal Models

Taxonomies provide key information about the semantic relationships between concepts and the inherent organization of vision and language. Despite their impressive capabilities, large multimodal models (LMMs) often lack taxonomic knowledge, leading to low hierarchical visual recognition (HVR) consistency. These models typically only rely on language modeling objectives during fine-tuning and lack explicit taxonomy-aware regularization. To address this, we propose Hierarchical Representation Regularization ($HiR^2$), a simple plug-and-play regularizer that improves hierarchical consistency in LMMs. Specifically, we introduce a semantic-aware visual tree construction framework that extracts coarse-to-fine visual features from intermediate LLM layers guided by textual cues. The regularizer combines two complementary objectives: a taxonomic entailment loss that enforces hierarchy via hyperbolic entailment cones in the Lorentz model, and a discriminative dispersive loss that promotes angular separation of semantically similar embeddings on the unit sphere without disturbing the radial hierarchical structure. Extensive experiments demonstrate that $HiR^2$ effectively captures taxonomic structures across diverse LMMs and fine-tuning methods. Code is available at https://github.com/PKU-ICST-MIPL/HiR2_ICML2026.

HANCLIP: A Family of Hyperbolic Angular Negation Vision Language Models

Vision-Language Models (VLMs) are typically pre-trained on large-scale image-text datasets to capture semantic correspondences between visual content and natural language. However, they remain surprisingly brittle to negation: models often rely on shallow word co-occurrence and are easily distracted by misleading or irrelevant textual cues, even when their overall retrieval or classification performance is strong. Moreover, directly finetuning on negation data can interfere with previously acquired knowledge, causing noticeable degradation on standard vision-language benchmarks. To tackle these issues, this work introduces HANCLIP (Hyperbolic + Angular + Negation), a family of VLMs that explicitly restructures the embedding space to encode "what an image is not" alongside "what it is." HANCLIP is trained on a compact set of 20,000 image-text quadruplets and combines a hyperbolic formulation, which models hierarchical semantic relations and asymmetries, with an angular triplet objective that drives systematic separation between negated descriptions and their corresponding positives. This geometry-aware design strengthens negation sensitivity while preserving the global structure of pretrained representations, rather than overwriting them. Extensive experiments across multiple vision-language tasks 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 benchmarks. The framework is model-agnostic and can be plugged into CLIP, LongCLIP, SmartCLIP, and HiMo-CLIP without large-scale retraining, demonstrating that a carefully designed geometric objective can substantially extend the reasoning capabilities of existing VLMs using only modest additional data.