StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?
The benchmark finds current self-evolution methods learn textbook-style gains without reliably transferring them to olympiad-level physics problems.
StudyBench tests how efficiently models convert raw textbook material into capability, splitting evaluation between difficult textbook applications and harder transfer problems. Across representative methods and three base models, gains on the Application Set rarely carried over to the Transfer Set. The authors also report a “Guidance Gap” versus in-context guidance, and a “Compute Plateau” where methods saturate before using their full budget. HF Daily Papers' note
StudyBench tests how efficiently models convert raw textbook material into capability, splitting evaluation between difficult textbook applications and harder transfer problems. Across representative methods and three base models, gains on the Application Set rarely carried over to the Transfer Set. The authors also report a “Guidance Gap” versus in-context guidance, and a “Compute Plateau” where methods saturate before using their full budget. HF Daily Papers' note
score 5