Inverse-LLaVA: Eliminating Alignment Pre-training Through Text-to-Vision Mapping

TL;DR

Inverse-LLaVA introduces a novel multimodal learning approach that eliminates the need for alignment pre-training by mapping text embeddings into visual space, enabling effective reasoning without large image-text datasets.

Contribution

It proposes a new paradigm that inverts traditional mapping, reducing computational costs and challenging the necessity of alignment pre-training in multimodal models.

Findings

Improves reasoning tasks by +0.2% to +27.2%.

Decreases perception task performance, e.g., celebrity recognition by -49.5%.

Reduces computational requirements by 45%.

Abstract

Traditional multimodal learning approaches require expensive alignment pre-training to bridge vision and language modalities, typically projecting visual features into discrete text token spaces. We challenge both fundamental assumptions underlying this paradigm by proposing Inverse-LLaVA, a novel approach that eliminates alignment pre-training entirely while inverting the conventional mapping direction. Rather than projecting visual features to text space, our method maps text embeddings into continuous visual representation space and performs fusion within transformer intermediate layers. Through selective additive components in attention mechanisms, we enable dynamic integration of visual and textual representations without requiring massive image-text alignment datasets. Comprehensive experiments across nine multimodal benchmarks demonstrate nuanced performance trade-offs: Inverse-LLaVA achieves notable improvements on reasoning-intensive and cognitive tasks (MM-VET: +0.2%, VizWiz: +1.8%, ScienceQA: +0.2%, cognitive reasoning: +27.2%), while showing expected decreases in perception tasks requiring memorized visual-text associations (celebrity recognition: -49.5%, OCR: -21.3%). These results provide the first empirical evidence that alignment pre-training is not necessary for effective multimodal learning, particularly for complex reasoning tasks. Our work establishes the feasibility of a new paradigm that reduces computational requirements by 45%, challenges conventional wisdom about modality fusion, and opens new research directions for efficient multimodal architectures that preserve modality-specific characteristics. Our project website with code and additional resources is available at https://inverse-llava.github.io.