TriViewBench measures multi-view structural reasoning in multimodal LLMs

arXiv·low signal

TriViewBench (2606.26029) introduces controlled complexity scaling to test whether MLLMs can reason about structure across multiple views, finding scalability gaps that standard VQA benchmarks hide. For agents that must build a coherent model of a scene or interface from several images, this maps where current MLLMs break down as complexity rises. A diagnostic benchmark rather than a new model.