Will the highest Humanity's Last Exam (HLE) accuracy score in Epoch AI's benchmark data reach at least 58.5 % for a model released on or before 2026-12-06?
échéance 2026-12-06
confiance moyenne
preuve niveau 4
36 %estimation PolySignal
Critère de résolution
Resolves YES if, in a snapshot of Epoch AI's benchmark data (https://epoch.ai/data/benchmark_data.zip, file "hle_external.csv", column "Accuracy") taken on or after 2026-12-20, at least one model with a release date on or before 2026-12-06 has a value >= 58.5 %. Reference at creation: 54.8 % (gpt-6-astra_unknown, 2026-09-03). Voided if Epoch AI revises the methodology so that values are no longer comparable, or retires the file. Data: Epoch AI, CC BY 4.0.
État de départ vérifié
Computed base rate: 36% of past 60-day windows in Epoch AI's benchmark data saw the running record rise by at least the required amount (threshold 0.585, current record 0.5479999999999999). Data: Epoch AI, CC BY 4.0, epoch.ai/benchmarks. Overlapping windows: small effective sample.
Les arguments ci-dessous ont été produits par le modèle en anglais. Ils sont affichés tels quels plutôt que traduits automatiquement, pour que rien sur cette page ne soit un texte que PolySignal n'a pas réellement écrit.
Ce qui fait monter
- Rapid frontier model reasoning advances have historically produced record jumps of this magnitude within a 60-day window [tier 1].
Ce qui fait descendre
- Humanity's Last Exam is designed specifically to resist rapid saturation, making a +3.7 percentage point increase within two months difficult [tier 1].
- No verified announcements or test results indicate an imminent model release exceeding 58.5% before December 2026 [tier 4].
Ce qui ferait le plus bouger ce chiffre
Whether a major frontier AI lab releases a new flagship reasoning model or update between October 8 and December 6, 2026, that Epoch AI evaluates on HLE.
Comment ce chiffre est fabriqué
| Échéance | 2026-12-06 |
| Taux de base | 36% |
| Meilleure preuve | niveau 4 · 1 article(s) utilisé(s) |
| Estimations | 35 · 36 (écart 1 pts) |
| Confiance | moyenne |
| Plafond de preuve | non déclenché |
| Révisée le | 2026-10-08 19:50 |
| Modèle | gemini-3.7-flash |
Sources consultées