GPQA
PhD-written science questions so hard that skilled non-experts with Google score 34% — the 'Google-proof' exam, reported on its 198-question Diamond subset.
Frontier reasoning models reach the high 80s on GPQA Diamond (e.g. Grok 4 ~87%, mid-2025), against a label-noise-adjusted ceiling likely in the low 90s — the paper documents a few percent of questions with expert-identified errors. Unlike MMLU, meaningful headroom remains and top models are only now crossing the PhD-expert line, so it still discriminates; but the trajectory is one or two model generations from done. Judgment call — the first `nearing-saturation` benchmark on the wiki.
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.Measured accuracy of PhD experts in specific scientific domains with open web search access.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.accuracy on GPQA Diamond, 4-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 gpqa --batch_size autoopencompass --datasets gpqa --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 GPQA, 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