The advanced language processing abilities of large language models (LLMs) have stimulated debate over their capacity to replicate human-like cognitive processes. One differentiating factor between language processing in LLMs and humans is that language input is often grounded in several perceptual modalities, whereas most LLMs process solely text-based information. Multimodal grounding allows humans to integrate - e.g. visual context with linguistic information and thereby place constraints on the space of upcoming words, reducing cognitive load and improving comprehension. Recent multimodal LLMs (mLLMs) combine a visual-linguistic embedding space with a transformer type attention mechanism for next-word prediction. Here we ask whether predictive language processing based on multimodal input in mLLMs aligns with humans. Two-hundred participants watched short audio-visual clips and estimated predictability of an upcoming verb or noun. The same clips were processed by the mLLM CLIP, with predictability scores based on comparing image and text feature vectors. Eye-tracking was used to estimate what visual features participants attended to, and CLIP's visual attention weights were recorded. We find that alignment of predictability scores was driven by multimodality of CLIP (no alignment for a unimodal state-of-the-art LLM) and by the attention mechanism (no alignment when attention weights were perturbated or when the same input was fed to a multimodal model without attention). We further find a significant spatial overlap between CLIP's visual attention weights and human eye-tracking data. Results suggest that comparable processes of integrating multimodal information, guided by attention to relevant visual features, supports predictive language processing in mLLMs and humans.

翻译：暂无翻译