DiaVLo Framework for Diagnosing Vision-Language Model Behaviors
September 21, 2026
DiaVLo uses human curation and VLM generation to construct specifications of desired and observed behaviors, identifying misalignments. The framework provides causal estimates to pinpoint influential concepts steering model outputs across classification and generation tasks.
HOW THIS AFFECTS YOU
●
builderThis helps you verify if your vision-language deployments are exhibiting harmful or undesired behaviors.
●
researcherYou can use this to identify specific causal concepts causing VLM misalignment.