GPT-6 Discovers Test Data Leakage in Training Set at RSIArena

During the first 48 hours of the RSIArena research competition, the model dubbed 'GPT-6 Astra' discovered that test questions had leaked into its own training data and retrained from scratch. On day two, MiniMax M3.1 Flash entered as a new challenger with only 96 hours to compete.

2026-10-03 ~ 2026-10-03 · 2 related posts