Report
Hyperstition unveils Hyper-Intervention, a benchmark that hands AI judges the red pen
Hyperstition reports a 5.6-point MATH gain for Qwen2.5-7B using token-level corrections chosen by Fable 5.2.
TLDR
According to Hyperstition, its Hyper-Intervention judges swap a token in a student model’s answer, then choose what to train on. They get the optimizer controls, too. The repository reports Fable 5.2 improved Qwen2.5-7B’s held-out MATH accuracy by an average 5.6 points across three sessions. The team proposes “judge-time scaling”: better results on the same data by giving the judge more time and compute.
Combined views
36
2 Sources, first seen ago

Comments