Summary
DiaVLo is a diagnostic framework for vision-language models (VLMs) proposed by Lorenzo Corti and Jie Yang in an arXiv paper (2609.22008, September 2026). VLMs depend on storing and transferring appropriate information across sub-components, and verifying that they exhibit desired behaviours while avoiding harmful ones is central to reliable deployment. DiaVLo combines human curation with VLMs' generation capabilities to build specifications of desired and observed behaviours, surfacing potential misalignments. It also provides causal estimates identifying the most influential concepts steering VLM behaviours. The framework was evaluated on several open-source VLMs under both classification and generation conditions. Experiments show that DiaVLo's behaviour labels correlate with model performance and provide context for measured performance, revealing clearly aligned and clearly misaligned behaviours as well as how VLMs perceive, organize, and prioritize concepts.
Paper Overview
- Research area: NLP
- Authors: Lorenzo Corti, Jie Yang
- Published: 2026-09-18
- arXiv: 2609.22008
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 the measured performance. DiaVLo reveals clearly aligned and clearly misaligned behaviours, as well as patterns in how VLMs perceive, organize, and prioritize concepts.
*(Automatically collected on 2026-09-22)*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178635080