Benchmark
MMVP
An evaluation benchmark specifically designed to measure the degree of visual reliance in multimodal models. It collects visual questions that cannot be answered correctly using linguistic priors alone, and is used to quantify the proportion of model responses that genuinely depend on visual evidence.
Timeline (last 90 days)
Oct 9
SLVR achieves an absolute improvement of +9.4 on the MMVP benchmark.
Unverified50%
Oct 9
Benchmarks such as MMVP and MMStar are specifically designed to detect the degree of visual dependency, deliberately including questions that cannot be answered correctly through language priors alone.
Unverified50%