MMLU-Pro
MMLU rebuilt to fix its flaws: 10 options instead of 4, trivia and label noise filtered out, and reasoning-heavy questions that reward chain-of-thought.
Frontier models cluster in the high 80s on MMLU-Pro by 2025–26 (leaderboards put the top ~89–90%), but cleaner labels and 10-option questions push the effective ceiling above MMLU's ~90%, so it still separates strong models where MMLU cannot. The 16–33 point accuracy drop it induced over MMLU is eroding; the trajectory points at saturation within a generation or two, hence nearing-saturation rather than active.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Domain expert accuracy under expanded 10-choice formats with reduced random guess baseline (10%).
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.accuracy, 10-option multiple choice (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks mmlu-pro --batch_size autoopencompass --datasets mmlu-pro --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for MMLU-Pro, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace