MathAdv team says Goedel-Prover-V2 solved six problems but failed on equivalent reformulations
The MathAdv team says the reverse—solving an equivalent reformulation but not its original—never occurred in its paired test. It calls this evidence of brittleness, not proof of memorization.
TLDR
MathAdv probes four dimensions across 13 mathematical domains, its team says: relevant knowledge, informal problem-solving, formal proof construction in Lean 4, and generalizing to equivalent reformulations. In a paired test, the team reports that Goedel-Prover-V2 solved six original problems but failed on their equivalent reformulations, with zero cases of the reverse. The team describes the result as evidence of brittleness—not proof of memorization.
Combined views
4.5K
12 Sources, first seen 20d ago