5, Gemma 3, and Molmo 2 families, ranging from roughly 1B to 235B parameters and including both dense and mixture-of-experts (MoE) variants.
Larger vision-language models can be worse on specific subtasks
A systematic study of 19 models finds thousands of questions where smaller models answer correctly but larger ones fail, with patterns recurring across model families.
Top University
Rohit Saxena · Alessandro Suglia · Marek Rei · Pasquale Minervini
University of Edinburgh · Imperial College London · Miniml.AI
Research Digest··3 min read
Saxena et al.
§