Prompt the Unseen: Evaluating Visual-Language Alignment Beyond Supervision

2025-08-31

AI Summary

This work investigates the generalization capability of vision-language model (VLM) projection layers to unseen visual concepts—i.e., their ability to handle novel categories without explicit cross-modal alignment supervision. To this end, we introduce the first fine-grained evaluation benchmark for projection-layer generalization, built upon object detection datasets and employing a prompt-based, label-disjoint train-test split protocol. Methodologically, we integrate prompt learning, feature-space mapping, and mechanistic interpretability analysis to systematically uncover a semantic alignment mechanism in the projection layer: it functions as a class-specific key-value memory. Experiments demonstrate that the projection layer retains 79%–88% of its original performance on unseen categories—substantially outperforming standard baselines. Our findings provide both theoretical grounding and a practical paradigm for efficient, low-resource cross-modal alignment training.

Technology Category

Computer Vision: Large Vision ModelsMachine Learning: Large Multimodal Models (LMMs)Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphsSearch and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Vision-Language Models (VLMs) combine a vision encoder and a large language model (LLM) through alignment training, showing strong performance on multimodal tasks. A central component in this architecture is the projection layer, which maps visual features into the LLM's embedding space. Despite its importance, its ability to generalize to unseen visual concepts has not been systematically evaluated. To address this, we propose a benchmark for evaluating projection-layer generalization. We adapt object detection datasets (rich in fine-grained annotations) into a prompting format and design train/test splits with disjoint label sets, enabling precise control over seen and unseen concept separation. Experimental results show that the projection layer retains about 79 to 88 percent of the performance on unseen classes compared to seen ones across various settings, suggesting a non-trivial level of generalization even without explicit alignment supervision on those concepts. We further analyze this behavior through a mechanistic interpretability lens. Our findings indicate that the feed-forward network in the projection layer functions like a key-value memory, processing seen and unseen tokens in similar ways. This study introduces a new evaluation framework for alignment generalization and highlights the potential for efficient VLM training with limited aligned data.

Problem

• Evaluating projection layer generalization for unseen visual concepts
• Assessing alignment performance on disjoint seen and unseen label sets
• Analyzing mechanistic interpretability of projection layer as key-value memory

Innovation

• Projection layer generalization benchmark using disjoint labels
• Mechanistic interpretability reveals key-value memory functionality
• Evaluation framework enables efficient VLM training with limited data