The Endless Exam: Mathematical Constructions from Today's Models toward Superintelligence
The benchmark scores mathematical constructions past known baselines, but the tested models did not beat any published frontier reference.
Endless Exam covers fourteen parameterized families of construction problems with automatic validity checks and relative quality scores. The paper says it can generate larger instances and use compact certificates so submissions can be verified without listing every element. Nine models were evaluated on 69 instances, producing continuous score differences even without surpassing the 30 frontier references. The v2 revision adds Claude Opus 5.5 evaluations and keeps the bench-v1.0 suite unchanged. HF Daily Papers' note
Endless Exam covers fourteen parameterized families of construction problems with automatic validity checks and relative quality scores. The paper says it can generate larger instances and use compact certificates so submissions can be verified without listing every element. Nine models were evaluated on 69 instances, producing continuous score differences even without surpassing the 30 frontier references. The v2 revision adds Claude Opus 5.5 evaluations and keeps the bench-v1.0 suite unchanged. HF Daily Papers' note
score 5