From medicine to geo-guessing, humans can get incredibly good at solving visual recognition tasks.
But how is this skill learned, and can we model its progression?
We present CleverBirds, accepted #NeurIPS2025, a large-scale benchmark for visual knowledge tracing.
📄 arxiv.org/abs/2511.08512
1/5
I will be at EMNLP next week presenting this work on November the 7th! Reach out to me for any questions :))
Work done with my advisor, Mirella Lapata!
Preprint: arxiv.org/pdf/2505.14627
#EMNLP2025 #multimodallearning #scalableoversight #visionlanguagemodels #nlproc
As opposed to previous work on debating, where models are assigned to argue for an answer, we only instruct the models to argue for opinions they believe to be true. This is not only efficient but can allow for extracting reasoning data that can update their beliefs.
RQ3: Where do debate or consultancy fail?
Our analysis show that judges benefit when the experts are arguing for diverse opinions!
Red quadrant is when the judge is persuaded more often than they should (i.e. they are deceptive).
RQ2: Can debate be used as a reliable mechanism for yielding quality reasoning data?
Yes! We show that the reasoning data attained from debate in a completely unsupervised manner imbue reasoning in the expert vision language models.
Excited to share my first work as a PhD student at EdinburghNLP that I will be presenting at EMNLP!
RQ1: Can we achieve scalable oversight across modalities via debate?
Yes! We show that debating VLMs lead to better model quality of answers for reasoning tasks.
