ReasoningSaturated
Trust Score:
25

AGIEval

Standardized human exams — SATs, LSATs, China's Gaokao and civil-service tests — repurposed to grade models against real human test-takers.

Launched: Refresh: static
Status Assessment (saturated):

GPT-4 already cleared the average-human line (67%) on several sections at launch; by 2024 frontier models pushed the 20-task aggregate past the average human and near the top-percentile 91% ceiling, and the human-exam framing lost its discriminative power at the top. saturated_date (2024-06) marks the GPT-4-class era where the average-human baseline stopped separating strong models.

Performance Timeline

Longitudinal progression of model scores against human baselines.

Performance & Historical Trajectory

Empirical score progression across model release dates and evaluation rounds.

Independent Vendor-Reported Human Baseline (67.0%)
ModelScoreDateSource TypeProvenance
DeepSeek-V3 (base, 0-shot)79.6%2024-12-26vendor-reportedSource ↗
Qwen2.5 72B (base, 0-shot)75.8%2024-09-19vendor-reportedSource ↗
GPT-4o (few-shot)69%2024-05-13vendor-reportedSource ↗
text-davinci-00337.4%2023-04-13independentSource ↗
ChatGPT (gpt-3.5-turbo)43.2%2023-04-13independentSource ↗
GPT-458.4%2023-04-13independentSource ↗

Human Baseline & Difficulty Horizon

Calibrated human reference points, specialist benchmarks, and ceiling thresholds.
Measured Human Score67.0%Domain Expert Baseline
Baseline Protocol & Interpretation

Average score of human test-takers across SAT, LSAT, GRE, Gaokao, and Chinese civil service exams.

Metric & Scoring Methodology

Verification protocols, aggregation formulas, and specialized metric variants.
Primary Metric:accuracy, macro-averaged over the 20 tasks (%)
Scoring Engine:exact-match

Dataset & Compute Cost

Evaluation volume, public availability, API pricing, and local hardware requirements.
Total Dataset Size8,062Annotated evaluation items
Public Test SetPublicOpenly mirrored on repositories
Access GatingOpen AccessUnrestricted download
Evaluation LicenseOpen AccessDataset usage and redistribution terms
Frontier API Compute Cost:

$5 – $20 USD for full benchmark evaluation run on frontier APIs.

Recommended Local GPU Setup:

1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang

Official Dataset & Benchmark Files:Download / View Dataset Repository ↗

How to Run & Reproduce

Standardized evaluation protocols, CLI commands, and reproducible runner templates.
Prompt Regimezero-shot
Reasoning Modedirect
Sampling Temp0
Pass@k Budgetk = 1
Tools & SandboxPure Text
Scoring Verifierexact-match
Option AEleutherAI LM-Evaluation-Harness (Open-Weight Models)
lm_eval --model hf --model_args pretrained=<model_path> --tasks agieval --batch_size auto
Option BOpenCompass Evaluation Framework
opencompass --datasets agieval --models <model_config>
Python APIDeterministic Inference Loop Snippet
# 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,
)
Standardized Reporting Requirement:

When publishing results for AGIEval, 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.
Overall Contamination Risk:HIGH
Refresh Cadence:

Static fixed snapshot

Test Set Exposure:

Public on web / HuggingFace