Melissa Sands @msands.bsky.social · 11/12/2025We then study these relationships by gender. LMMs easily perform worst in recovering the views of Detroit men. Why? This difficult-to-reach population is likely poorly represented in training data/reinforcement feedback. It may be that this problem is hard to resolve even with better models. 110
Melissa Sands @msands.bsky.social · 11/12/2025We then asked the same questions of two human samples: a U.S. nationally representative sample and Detroit-representative sample (through DMACS at @umfordschool.bsky.social). We then compare these responses, two ways. In general, LMMs better recover national evaluations than local ones. 121
Melissa Sands @msands.bsky.social · 11/12/2025We pulled tens of thousands of open-source street view images from Detroit. We processed the images, and then sampled 85 for closer analysis. For each, we queried five LMMs (propietary and open-source) 30 times for ratings on wealth, safety (day and night), and disorder. 110
