S-Agent 2.0 claims to beat GPT-6 Astra on two spatial reasoning benchmarks

Hao Li introduces S-Agent 2.0, a framework that turns coding agents into evidence-grounded spatial reasoning agents.

Li says it outperforms GPT-6 Astra on MMSI-Bench and ReVSI-Bench. The claim is made in a social media post, and the details are in an arXiv paper and project page that the post links to. No independent evaluation of the results is reported.