Gemini-2.5-Flash on VSC-Repeat

The paper states: “Our findings do not imply that there has been no progress in improving spatial supersensing. Cambrian-S reports that Gemini-2.5-Flash, used as a generic long-context video model, achieves around 41.5% accuracy on 60-minute VSR videos—similar in magnitude to the Cambrian-S model with predictive memory on the same split—despite having no benchmark-specific inference design exposed to users.”

I think an interesting follow-up experiment would be to evaluate Gemini-2.5-Flash on VSC-Repeat to see whether it already shows a preliminary ability to build a stable internal map that survives revisits to the same environment.

The issue names no files, tests, or entry points. Start by locating the existing VSC-Repeat evaluation path and determine how Gemini-2.5-Flash could be evaluated; done means reporting its results on VSC-Repeat and comparing them with the cited benchmark context.

Written by the indexing model from the issue text.