FrontierMath
A high-difficulty mathematics benchmark constructed by Epoch AI, with problems authored by professional mathematicians. It covers research-level domains including number theory, algebraic geometry, and combinatorics, organized into four difficulty tiers (Tier 1–4). The benchmark is designed specifically to evaluate the research-level mathematical reasoning capabilities of large language models.
Timeline (last 90 days)
If FrontierMath's discriminative power declines, it may prompt the introduction of a new generation of benchmarks requiring formal proof verification (such as integration with Lean/Coq theorem provers).
A post on Reddit claimed that the model GPT-6.1 Sol (max) achieved a score of 100% on the FrontierMath Tier 4 benchmark.
When releasing FrontierMath, Epoch AI noted that the difficulty of its problems corresponds to scenarios that only PhD-level mathematics graduate students or professional researchers can handle.
Tier 4 of FrontierMath represents the highest difficulty level in that benchmark, typically involving cross-domain knowledge integration and multiple rounds of non-trivial derivation.
The page linked to by the result is labeled with version information for tier-4-v2, indicating that the benchmark itself has undergone iterations and that directly comparing scores across versions is not rigorous.
FrontierMath adopts confidentiality measures such as keeping problems undisclosed and granting access only to model submitters in order to reduce the risk of data contamination.
FrontierMath is a benchmark test consisting of unpublished competition problems designed by professional mathematicians.
The GPT-6 Astra leak claims a score of 98% on FrontierMath Tier 4.
FrontierMath is a benchmark designed specifically to measure AI's advanced mathematical reasoning capabilities.
The performance of current state-of-the-art models on the highest difficulty tier of FrontierMath is typically only in the single-digit to double-digit percentage range.
2 more timeline events
All Facts (11)
FrontierMath adopts confidentiality measures such as keeping problems undisclosed and granting access only to model submitters in order to reduce the risk of data contamination.
50%UnverifiedWhen releasing FrontierMath, Epoch AI noted that the difficulty of its problems corresponds to scenarios that only PhD-level mathematics graduate students or professional researchers can handle.
50%UnverifiedTier 4 of FrontierMath represents the highest difficulty level in that benchmark, typically involving cross-domain knowledge integration and multiple rounds of non-trivial derivation.
50%UnverifiedThe page linked to by the result is labeled with version information for tier-4-v2, indicating that the benchmark itself has undergone iterations and that directly comparing scores across versions is not rigorous.
50%UnverifiedFrontierMath is a benchmark test consisting of unpublished competition problems designed by professional mathematicians.
50%UnverifiedFrontierMath is a benchmark designed specifically to measure AI's advanced mathematical reasoning capabilities.
50%UnverifiedThe performance of current state-of-the-art models on the highest difficulty tier of FrontierMath is typically only in the single-digit to double-digit percentage range.
50%UnverifiedFrontierMath is a frontier mathematics problem set designed with the participation of mathematicians including Terence Tao.
50%UnverifiedNext-generation evaluations such as FrontierMath and ARC Prize attempt to use novel, unpublished difficult problems to test models' genuine reasoning capabilities.
50%UnverifiedIf FrontierMath's discriminative power declines, it may prompt the introduction of a new generation of benchmarks requiring formal proof verification (such as integration with Lean/Coq theorem provers).
50%UnverifiedA post on Reddit claimed that the model GPT-6.1 Sol (max) achieved a score of 100% on the FrontierMath Tier 4 benchmark.
50%