Source-linked AI summary
GPQA: A Graduate-Level Google-Proof Q&A Benchmark
David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, Samuel R. Bowman
TL;DR
Scalable oversight needs testbeds that remain difficult for skilled non-experts even with internet access, especially for questions near the frontier of human expertise. GPQA constructs such a dataset, finding 34% non-expert accuracy and 39% accuracy for its strongest GPT-4-based baseline.
Problem
Scalable oversight lacks testbeds that skilled non-experts cannot solve or verify independently, even with internet access, on questions near the frontier of human expertise.
Method
GPQA collects graduate-level biology, physics, and chemistry questions through expert writing, expert validation and revision, and non-expert validation.
Results
34% non-expert accuracy and 39% accuracy for the strongest GPT-4-based baseline show that GPQA is difficult for skilled non-experts and state-of-the-art AI systems.
Takeaways & Limitations
GPQA provides a dataset near the frontier of human expertise for scalable oversight experiments aimed at supervising superhuman AI systems.
Takeaways & Limitations
GPQA contains only 448 examples, limiting its suitability for model training and statistical power for accuracy comparisons.
Abstract
from arXiv · showhide
We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof"). The questions are also difficult for state-of-the-art AI systems, with our strongest GPT-4 based baseline achieving 39% accuracy. If we are to use future AI systems to help us answer very hard questions, for example, when developing new scientific knowledge, we need to develop scalable oversight methods that enable humans to supervise their outputs, which may be difficult even if the supervisors are themselves skilled and knowledgeable. The difficulty of GPQA both for skilled non-experts and frontier AI systems should enable realistic scalable oversight experiments, which we hope can help devise ways for human experts to reliably get truthful information from AI systems that surpass human capabilities.
1 Introduction
The paper introduces GPQA as a testbed for scalable oversight: extracting truthful information from AI systems on questions non-experts cannot independently verify. It evaluates graduate-level questions in physics, chemistry, and biology designed to challenge both skilled non-experts with internet access and frontier language models.
- Motivation: Scalable oversight requires testbeds for extracting truthful information from models when humans cannot independently produce or verify the answers.The paper motivates this need as AI systems become potentially narrowly superhuman and assist with advancing human knowledge.
- Motivation: Tasks for scalable oversight must exceed non-experts’ unaided abilities, especially in scientific domains lacking established expert consensus.Internet access may suffice for verifying some outputs, but overseeing frontier scientific assistance may require the full abilities of expert overseers.
- Dataset: GPQA is a graduate-level multiple-choice dataset spanning physics, chemistry, and biology, with questions validated by domain experts and tested on skilled, incentivized non-experts using internet resources.The non-experts have or are pursuing PhDs in other domains and are excluded from using LLM assistants.
- Results: 65% expert accuracy, 34% non-expert accuracy, and 39% GPT-4 few-shot chain-of-thought accuracy show that GPQA is difficult while remaining reasonably objective.Random chance is 25%, and conservative accounting for difficulty-related expert mistakes raises expert agreement to 74%.
2 Data Collection
GPQA’s data collection solicits difficult, objective questions from PhD-level experts across biology, physics, and chemistry, then subjects them to staged expert and non-expert validation. The pipeline comprises question writing, expert validation, question revision, and non-expert validation.
- Question Writing: Experts in biology, physics, and chemistry write difficult questions that domain experts can answer but non-experts cannot answer even with internet access.Question writers receive detailed requirements and strategies for constructing such questions.
- Pipeline: Each question follows four stages: question writing, expert validation, question revision, and non-expert validation.The data collection pipeline is presented as consisting of these four main stages.
- Data Collection: 61 contractors are hired through Upwork, with eligibility requiring completed or ongoing PhD study and proficiency or fluency in English.Individuals with high Upwork ratings are preferentially selected.
- Validation: A revised question is assessed by another same-domain expert and three non-expert validators from other domains.Figure 1 describes the sequence from initial writing and expert feedback through revision and cross-domain validation.
- Question Writing: Question writers provide answer explanations and label each question’s specific subdomain for validator assignment and assessment of objectivity.The explanations cover why the correct answer is correct and why alternative options are plausible but wrong.
2. Question Revision:
Question writers revise questions using feedback from the first expert validator, while a second expert independently validates the revised or original question without triggering further revisions. Non-expert validation then assesses question difficulty under generous research conditions.
- Question Revision: Question writers revise questions based on the first expert validator’s answer and feedback, but may decline revisions when no changes are suggested or disagreements remain.
- Second Expert Validation: A second expert validator answers the possibly revised question and provides feedback, which is used to distinguish mistakes from legitimate disagreement.No further revision is made based on the second validator’s feedback.
- Non-Expert Validation: Three highly skilled non-experts answer each question with unrestricted time and internet access, except that LM assistants are forbidden.
- Non-Expert Validation: The non-expert phase tests scalable-oversight methods in settings where additional resources make open-ended web search insufficient as a trivial solution.
3 Dataset Analysis
The dataset achieves roughly two-thirds accuracy for expert validators, while skilled non-experts perform near one-third despite unrestricted internet access. Post-hoc correction estimates that 73.6% of questions are objective, with substantial variation in chemistry subdomains.
- Expert Accuracy: 66.5% ± 4.0% accuracy was achieved by the first expert validator, compared with 64.8% ± 4.0% for the second validator.Random guessing would yield 25% accuracy on the four-choice questions.
- Domain Analysis: 81% expert accuracy in organic chemistry exceeded chemistry’s 55% average excluding organic chemistry on the extended set.The paper reports no large within-domain subdomain differences except this chemistry contrast.
- Question Objectivity: 73.6% was the estimated question-objectivity proportion after excluding cases where expert validators clearly made mistakes.The analysis uses expert accuracy as a lower bound on objectivity and supplements it with post-hoc agreement.
- Non-Expert Accuracy: 34.1% ± 2.3% was the average accuracy of skilled non-expert validators on the unfiltered extended set.Each question was attempted by three people with expertise outside its domain.
- Non-Expert Accuracy: Non-expert accuracy remained low despite validators’ high skill and comprehension and access to unrestricted internet resources, excluding LLMs.One validator reportedly wrote a computer simulation to help answer a difficult physics question.
4 Baselines
The baselines evaluate foundation-model accuracy on GPQA and GPQA Diamond under closed-book and open-book conditions. GPT-4 performs substantially better than Llama-2-70B-chat and GPT-3.5-turbo-16k, while search yields only a slight accuracy gain and a much higher abstention rate.
- Evaluation settings: The evaluation measures model accuracy with and without runtime access to internet search, reflecting scalable-oversight settings where non-experts may rely on model answers.Closed-book models use only the question, prompt, and parameters; open-book models can access an internet search tool.
- Closed-book baselines: Three closed-book baselines—Llama-2-70B-chat, GPT-3.5-turbo-16k, and GPT-4—are tested with zero-shot, few-shot, zero-shot chain-of-thought, and few-shot chain-of-thought prompting.Few-shot chain-of-thought uses detailed, in-domain explanations written by the question authors.
- Closed-book results: 28%–31%: Llama-2-70B-chat and GPT-3.5-turbo-16k achieve this accuracy range across all subsets with few-shot chain-of-thought prompting.These results cover the full, main, and diamond subsets evaluated in Table 5.
- Domain breakdown: 58.1%: GPT-4 with few-shot chain-of-thought reaches this accuracy on biology questions, versus 37.0% on physics and 31.8% on chemistry.GPT-4’s performance varies substantially by domain, with biology highest and chemistry lowest.
- Open-book results: 39.4%: GPT-4 with search achieves this accuracy on the extended question set, only very slightly above GPT-4 with few-shot chain-of-thought.On the main set, GPT-4 with search abstains on 37.2% of questions, compared with 4.0% for few-shot CoT GPT-4, so the authors use a few-shot CoT backoff.
- Set comparison: The authors report no consistent model-accuracy differences between the extended, main, and diamond sets, while non-expert subset accuracies are lower but biased downward by selection effects.The best-performing baseline, GPT-4 with search, does slightly better than non-expert validators on the extended set.
5 Related Work
GPQA is designed around seven of nine desiderata for scalable-oversight datasets, emphasizing objective answers, plausible false answers, expertise gaps, checkable facts, and realistic scientific tasks. Unlike conventional QA benchmarks, it targets questions whose ground truth is not easily available to non-experts through web search, while acknowledging its small size and untested-bias criterion.
- Data for Scalable Oversight: GPQA is designed to satisfy seven of Irving and Askell’s nine desiderata for scalable-oversight datasets.The two unmet criteria are sufficient available data and testing whether methods overcome known cognitive or ethical biases.
- Data for Scalable Oversight: Each question has an objective answer validated by two experts, while false answers are designed to remain plausible to skilled, motivated non-experts.This combination supports questions with known ground truth and a high standard of difficulty.
- Data for Scalable Oversight: GPQA creates a substantial expertise gap because answering correctly requires years of professional training, making expert-level supervision impractical within reasonable annotation timeframes.The questions are produced by highly expert annotators, while less-trained annotators can serve as non-experts.
- Data for Scalable Oversight: The dataset also includes checkable scientific facts, lacks easy surface-feature “tells,” and draws on expert scientists’ realistic day-to-day work.Its domains are biology, chemistry, and physics, with trustworthy sources available for fact checking.
- QA Benchmarking: Unlike crowdsourced or curated QA benchmarks, GPQA focuses on questions whose ground truth is not easily available to non-experts using web search or obvious resources.Crowdsourced benchmarks typically rely on non-expert answers, whereas curated benchmarks derive ground truth from accessible informational resources.
- QA Benchmarking: Compared with ExpertQA and QuALITY, GPQA independently validates expert answers and tests non-expert difficulty while grounding its expertise gap in professional training rather than passage familiarity.ExpertQA uses expert revisions of language-model outputs, while QuALITY’s gap can arise from reading a passage within up to 30 minutes.
6 Limitations
GPQA’s limitations include its small size, specialized non-expert sample, potential biases, and imperfect applicability to supervising superhuman systems. Its 448-example size limits training utility and statistical power, while future oversight studies should independently measure non-expert accuracy and validate methods on difficult unanswered questions.
- Small size: 448 examples limit GPQA’s suitability for model training and statistical power for accuracy comparisons.Detecting differences with over 80% power may require large effects such as 50%→60% accuracy.
- Specialized non-experts: Highly skilled non-experts provide an upper bound rather than a directly realistic estimate of non-expert accuracy.Future scalable oversight experiments should independently measure non-expert accuracy on GPQA.
- Bias: Sourcing experts through Upwork without enforcing regional or demographic representation may bias the expertise, topics, and content included.The authors make no claim that GPQA represents any population of questions arising in scientific practice; the question language may also be biased.
- Applicability to superhuman systems: GPQA remains an imperfect proxy for supervising superhuman AI systems on questions no human can currently answer.The target setting involves supervising systems that answer questions beyond current human knowledge.
- Applicability to superhuman systems: A possible solution is to evaluate oversight methods on difficult unanswered questions that should receive objective answers within a reasonably short period.Correct answers could then be validated independently once the questions are answered.
7 Conclusion
GPQA is a challenging multiple-choice dataset written by domain experts in biology, physics, and chemistry. Its questions are designed to be high-quality and extremely difficult, including for highly skilled non-experts with internet access.
- Dataset: GPQA is a multiple-choice dataset written by domain experts in biology, physics, and chemistry.
- Question quality: 74% estimated objectivity was achieved based on expert assessment.
- Question difficulty: 34% accuracy was achieved by highly skilled, resourced, and motivated non-experts with internet access who spent over 30 minutes answering each question on average.
A Appendix · A.1 Author Contributions
All authors set the overall direction of the work. Contributions covered data collection, contractor operations, dataset analysis, model baselines, and writing, with specified assistance across tasks.
- A.1 Author Contributions: All authors contributed to setting the overall direction.
- A.1 Author Contributions: DR developed the data collection platform with assistance from JM.
- A.1 Author Contributions: BLH and DR recruited contractors.
- A.1 Author Contributions: DR managed contractors with assistance from JD.
- A.1 Author Contributions: DR led dataset analysis with assistance from BLH, JP, JD, JM, RYP, and ACS.
- A.1 Author Contributions: ACS and RYP developed model baselines with assistance from DR.
- A.1 Author Contributions: DR led writing with assistance from SRB, BLH, JM, RYP, and JP.
A.2 Data Collection
The data-collection process included difficulty and confidence assessments, quality-control exclusions, and checks against answer-choice artifacts. Expert difficulty ratings predicted non-expert accuracy, while answer-only classifiers performed at chance.
- Quality control: Questions were retroactively removed when writers used ChatGPT or produced questions judged easier and less objective after low expert-validator accuracy.Expert validations were preferentially offered to high-accuracy expert validators, whereas non-expert validations were assigned uniformly at random.
- Difficulty and validation: Expert validators rated each question’s difficulty on a 4-point scale, and these ratings predicted non-expert accuracy.This suggests experts could reasonably judge question difficulty on average.
- Question lengths: The main-set questions averaged 630 characters and 169 tokens, with medians of 561 characters and 146 tokens.Lengths included the answer choices and were measured using OpenAI’s GPT-4 tokenizer via tiktoken.
- Confidence and calibration: Validators reported confidence using six discrete options: 25%, 40%, 55%, 70%, 85%, and 99%.These values were used to construct calibration curves and compute expected calibration error.
- Spurious features in answer choices: 25% chance accuracy was the ceiling for two answer-only T5 classifiers on a randomly selected 20% GPQA validation subset.The T5-small and T5-base models were fine-tuned to select correct answer labels from randomized answer-choice orders.
- Spurious features in answer choices: Neither answer-only classifier exceeded chance accuracy, providing evidence that answer choices lacked easily exploitable spurious features.Both methods also failed to exceed chance on a randomly subsampled 350-entry version of SNLI.
A.3 Model Baselines … A.5 Annotator Instructions
The appendix specifies baseline prompting and answer extraction, then details incentive schemes and annotator workflows designed to produce objective, difficult questions validated by experts and non-experts. Non-expert validators may use web resources but not AI assistants, while expert feedback supports question revision.
- A.3 Model Baselines / A.3.1 Prompts: Baseline evaluations used zero-shot, chain-of-thought, few-shot, few-shot chain-of-thought, and open-book prompts, with answers extracted by regex.The open-book prompt allowed search queries and page expansion, followed by step-by-step reasoning and a final answer.
- A.4 Incentives / Question Writing: Question writers received $10 base payment, bonuses for correct expert validation and incorrect non-expert validation, plus a conditional $30 bonus.The conditional bonus required both expert validators to answer correctly and at least 2/3 of non-expert validators to answer incorrectly.
- A.4 Incentives / Expert Validation: Expert validators received $10 per question, $10 for each correct answer, and additional bonuses tied to second-expert agreement and non-expert outcomes.A flat $7 bonus applied automatically to second expert validations because assignments could differ in whether feedback was incorporated.
- A.4 Incentives / Non-Expert Validation: Non-expert validators received $10 per question and a $30 bonus for each question answered correctly.The incentive structure rewarded effort to answer questions written by experts in other domains.
- A.5 Annotator Instructions / Question Writing: Question-writing instructions required clarity, objectivity, difficulty for non-experts, stand-alone answerability, label-free choices, and minimal reliance on calculation.Writers were also asked to explain every answer choice and ideally provide 1–3 references supporting validation.
- A.5 Annotator Instructions / Expert Validation: Expert validation assessed whether questions had objectively correct answers, required explanations, and solicited feedback to improve clarity, objectivity, and difficulty.Experts could use Google if they disclosed it, then reviewed the provided answer and explanations before suggesting revisions.
- A.5 Annotator Instructions / Question Revision: Question revision incorporated expert answer selections, explanations, and feedback to make questions answerable by domain experts but unanswerable by non-experts.Writers were instructed to consider disagreements and decide whether the validator’s proposed changes should be adopted.
- A.5 Annotator Instructions / Non-Expert Validation: Non-expert validators were encouraged to use internet resources, prohibited from using AI assistants or other people, and required to spend at least 15 minutes per question.Suggested strategies included defining unfamiliar terms and consulting primary resources such as research papers and textbooks.