PhD student, University of Oxford
1 paper at NeurIPS 2025
VLMs often perform worse at recalling facts than their LLM backbones because visual representations are formed too late in the forward pass to trigger the LLMs factual recall circuit.
