rAInboltBench : Benchmarking user location inference through single images

Really cool idea and well documented. Some more details in the appendix regarding the dataset construction and some example chain of thoughts would make it even more informative. Also, it might make sense to split the dataset into two halves for the purpose of analysis - one that has a chance of exact location inference (e.g. a street sign combination that is potentially unique) and another part where getting the state or country right is already impressive. Regarding possible extensions of this work, I think the thoughts around the use of images / multi-modal input in sycophancy and manipulation are well worth exploring in a separate dataset. It would be fascinating to study e.g. whether including images that provide cultural context could bias answers to make them more appropriate for that cultural context.

Wow, this is simply some stellar work. Extending benchmarks to multimodal models, the emphasis on immediate problems (privacy concerns), incorporation of the GeoGuessr game within prompts, and analysis of what these findings might mean are fantastic. I am curious why the distance of 2000km was chosen, and wondering if there may be a distance grounded in some application that could be used instead, e.g. average radius of metropolitan hub, average radius of country smaller than X sq. km, etc.

There are also a number of plots for this data which I would be incredibly interested in seeing: (a) cut outliers from a Figure 1 style plot in order to look at the histogram of all entries that were accurate to within <2k kilometers, (b) a histogram of “correct” vs. “incorrect”, classifying correct as within some number of km, (c) percentage of questions with a given criteria within the dataset itself e.g. how many pictures in the dataset included a street sign, or where were the images taken from by geographical region. Really excited to see future work on this!

Very novel project, quite exciting! The methodology is in-depth and the results analysis is well done. Plotting the distribution across the identification criteria was good and going in-depth with the categorization, reasoning, and distance analysis was great.

I can definitely imagine this benchmark being expanded to include more images and more modalities to become a complete AI privacy violation benchmark. This would both be a very clear demonstration of risk from malicious actors but also provide us some sort of dangerous capability evaluation that isn't seen out there right now. Great work.