Vision language models diagnosed to reveal aligned and misaligned behaviors
DiaVLo: Diagnosing Behaviours of Vision-Language Models
Computation and LanguageArtificial Intelligence
Summary
Vision-language models process and combine images and text to perform tasks, but it’s hard to know if they behave as expected or make mistakes. The authors created DiaVLo, a tool that helps check what these models do by comparing intended and actual behaviors, and finds which ideas most influence their decisions. They tested DiaVLo on multiple open-source models and showed it can highlight patterns in how models understand and prioritize concepts. This helps developers spot good or problematic behaviors and better understand model performance.
What this means in practice
- •For ai developers: Identify influential concepts and behaviors to improve vision-language model alignment in classification and generation tasks.
- •For product managers: Detect potential misalignments and harmful behaviors in vision-language models before deployment to increase reliability.
Authors
Lorenzo Corti, Jie Yang
Abstract
Vision-language models (VLMs) rely on storing and transferring appropriate information across their sub-components. Verifying that the VLMs exhibit desired behaviours, while avoiding harmful ones, is central to their reliable deployment. Yet, methods that identify VLM behaviours remain scarce. We present DiaVLo, a diagnostic framework that leverages human curation and VLMs' generation capabilities to construct specifications of desired and observed VLM behaviours, surfacing potential misalignments. Beyond this, DiaVLo also provides causal estimates to identify the most influential concepts steering VLM behaviours. We evaluate DiaVLo on several open-source VLMs under both classification and generation conditions. Our experiments show that DiaVLo produces behaviour labels that correlate with model performance and provide context for measured performance. DiaVLo surfaced behaviours that are clearly aligned and misaligned, alongside patterns in how VLMs perceive, organise, and prioritise concepts.