Manifold MarketsKYC yokDoğrulanmış sonuçlanma verisiİyi kalibreGüncel topluluk tahmini2025-2026 18.9%8 seçenek arasında lider
Tahminciler
4
Soru türü
multiple choice
Yöntem
Play-money forecasting platform
Kaynak türü
Tahmin
7g Değişim
0.0 pts
Piyasa verileri
1 saat önce güncellendi
Gecikmeli13 Tem 25 17:43 UTC1 Oca 41 23:59 UTC
Trendler
Kurallar
PutnamBench is a multilingual benchmark that translates the famously difficult William Lowell Putnam Mathematical Competition into formal theorem-proving tasks: its current release contains 657 Lean problems, 640 Isabelle/HOL problems, and 412 Coq problems.
- The original Putnam contest—two three-hour sessions of six proof-style questions worth 120 points—has a typical median score of just 0 or 1 among the ~4 000 elite undergraduates who sit it each year, underscoring the conceptual creativity required.
- By encoding every problem in three major proof assistants, PutnamBench measures an AI’s ability not merely to produce correct final answers but to generate machine-checkable proofs across multiple formal languages and undergraduate fields such as algebra, combinatorics, analysis and number theory.
- State of Play:
- As of July 2025, the PutnamBench leaderboard highlights just how far current automated theorem-provers still have to go: in the Lean 4 track (657 problems), DeepSeek-Prover-V2 tops the board with 47 proofs, yielding 7.2 % coverage; on the Isabelle/HOL track (640 problems), a GPT-4o-based wrapper dubbed DSP has verified only 4 theorems, or 0.6 % of the set; and on the Coq track (412 problems), the GPT-4o effort has managed a single proof at 0.2 % accuracy.
- Why this milestone matters
