SPELUNKA operates purely at inference time without modifying the VLM.
Automated framework discovers hidden biases in vision-language models
SPELUNKA identifies spurious associations in multimodal retrieval without labeled subgroups or fine-tuning.
Big Tech
Cassandra Parent · Isha Agarwal · Kumail Alhamoud · Charikleia Moraitaki · Rafiya Javed · Walter Gerych · +1 more
MIT · Google · WPI
Research Digest··3 min read
Parent et al.
§