Humanity's Last Exam
BenchmarkFreeHardest exam questions from thousands of experts.
Capabilities8 decomposed
interdisciplinary expert-sourced question curation
Medium confidenceCompiles exam questions from thousands of expert contributors across every academic discipline into a unified benchmark dataset. Questions are sourced directly from domain experts rather than synthetically generated, ensuring authentic representation of real-world assessment standards. The curation process includes a bug bounty program (closed 03/21/2025) that identified and removed searchable questions (those findable via web search), with replacement questions sourced from late contributors to mitigate data contamination.
Uses a bug bounty program (closed 03/21/2025) to explicitly identify and remove web-searchable questions, then replaces them with late-contributor questions — a contamination-detection approach not standard in other benchmarks. The replacement strategy ensures the final 2,500-question set avoids memorization shortcuts while maintaining expert validation.
More rigorous contamination mitigation than benchmarks like MMLU or ARC, which rely on post-hoc contamination detection; HLE's proactive bug bounty + replacement approach removes searchable questions before publication rather than discovering contamination after model evaluation.
multi-discipline knowledge assessment across 2,500 expert questions
Medium confidenceProvides a static, finalized benchmark of 2,500 exam questions spanning every academic discipline, designed to measure AI knowledge and reasoning capabilities before superhuman performance thresholds. Questions are compiled from thousands of experts and published in Nature (649, 1139–1146, 01/28/2026), establishing a fixed evaluation standard. The benchmark is accessible via Hugging Face Datasets (`cais/hle`) for reproducible evaluation across models.
Published in Nature with 100+ named contributors from CAIS and Scale AI, establishing a peer-reviewed standard rather than a proprietary benchmark. The 2,500-question fixed set is immutable post-publication, preventing benchmark drift and enabling long-term comparability across model generations.
More authoritative than crowd-sourced benchmarks (MMLU, ARC) due to Nature publication and explicit expert vetting; more stable than rolling benchmarks because the finalized version is frozen, preventing contamination from new model releases.
dynamic rolling benchmark with ongoing expert contributions
Medium confidenceMaintains HLE-Rolling, a dynamic fork version (released 10/08/2025) that accepts ongoing expert contributions via email submission to `agibenchmark@safe.ai`. This allows the benchmark to evolve with new questions from domain experts, preventing models from saturating the fixed 2,500-question set. Update logs track contributions, and the rolling version serves as a living standard for continuous evaluation.
Decouples the finalized published benchmark (2,500 questions, Nature-backed) from a rolling version that accepts ongoing contributions, preventing saturation while maintaining a stable reference standard. The dual-version approach allows continuous evolution without compromising reproducibility of published results.
More adaptive than static benchmarks (MMLU, ARC) which become stale as models improve; more rigorous than fully open benchmarks (like some Hugging Face community datasets) because contributions are curated by CAIS/Scale AI rather than unrestricted.
hugging face dataset integration with reproducible loading
Medium confidenceProvides the benchmark as a Hugging Face Datasets artifact (`cais/hle`) that can be loaded programmatically via `load_dataset()`, enabling reproducible evaluation across research teams without manual data management. The dataset is versioned and immutable, ensuring that published results reference the same question set. This integration pattern allows seamless incorporation into standard ML evaluation pipelines.
Leverages Hugging Face Datasets' versioning and immutability guarantees to ensure that published benchmark results reference the exact same question set indefinitely, preventing the 'moving target' problem where dataset updates invalidate prior comparisons. This is a deliberate architectural choice to prioritize reproducibility over convenience.
More reproducible than benchmarks distributed via GitHub or direct downloads because Hugging Face Datasets provides version pinning and automatic caching; more accessible than proprietary benchmark APIs because it uses the open-source Datasets library that researchers already use for other benchmarks.
leaderboard submission and model performance tracking
Medium confidenceMaintains an HLE-Rolling Live Submission Dashboard (accessible at https://lastexam.ai) that tracks model performance across the benchmark. The leaderboard accepts submissions via email to `agibenchmark@safe.ai` for the rolling version, enabling researchers to compare their models against published baselines and other submissions. The leaderboard provides visibility into which models are approaching superhuman performance thresholds.
Decouples the finalized benchmark leaderboard (for the 2,500-question set) from the rolling leaderboard (for ongoing contributions), allowing researchers to submit to either version depending on their evaluation timeline. This dual-leaderboard approach prevents the rolling version from contaminating the published baseline while still enabling continuous comparison.
More transparent than proprietary model evaluation systems (like OpenAI's internal benchmarking) because results are publicly visible; more flexible than single-version leaderboards because it supports both fixed and rolling evaluations, accommodating different research timelines.
nature-published peer-reviewed benchmark standard
Medium confidenceEstablishes HLE as a peer-reviewed benchmark published in Nature (649, 1139–1146, 01/28/2026), providing academic credibility and methodological rigor. The Nature publication undergoes peer review, establishing the benchmark as a vetted standard rather than a proprietary tool. This publication status enables researchers to cite HLE in papers and use it as a reference standard for model evaluation.
Achieves peer-reviewed publication in Nature, a top-tier journal, which subjects the benchmark methodology to external scrutiny and establishes it as an academic standard rather than a proprietary tool. This publication status is rare for AI benchmarks and signals that the benchmark has undergone rigorous validation.
More credible than unpublished benchmarks (like many Hugging Face community datasets) because it has undergone peer review; more authoritative than benchmarks published in workshops or preprints because Nature is a top-tier venue with high methodological standards.
open-source benchmark dataset and infrastructure
Medium confidenceReleases the benchmark as open-source, making both the question dataset and (presumably) evaluation infrastructure publicly available. The open-source approach enables researchers to audit the benchmark, contribute improvements, and integrate it into their own evaluation pipelines without licensing restrictions. This transparency supports reproducibility and community-driven improvements.
Combines open-source distribution with Nature publication, ensuring that the benchmark is both academically vetted and community-auditable. This dual approach prevents vendor lock-in while maintaining methodological rigor through peer review.
More transparent than proprietary benchmarks (like some commercial AI evaluation services) because the source code is publicly available for audit; more rigorous than purely community-driven benchmarks because it has undergone peer review and is maintained by established organizations (CAIS, Scale AI).
free public access to benchmark and leaderboard
Medium confidenceProvides free access to both the benchmark dataset and leaderboard, removing financial barriers to evaluation. Researchers can download the 2,500-question dataset via Hugging Face Datasets at no cost, and submit results to the public leaderboard without fees. This free-access model democratizes access to a frontier-grade benchmark.
Removes all financial barriers to accessing a Nature-published, expert-sourced benchmark, making frontier-grade evaluation accessible to researchers regardless of budget. This is a deliberate choice by CAIS and Scale AI to prioritize broad adoption over monetization.
More accessible than commercial benchmarking services (which charge per evaluation) and more equitable than paywalled academic benchmarks; enables smaller labs to compete on equal footing with well-funded organizations.
Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.
Related Artifactssharing capabilities
Artifacts that share capabilities with Humanity's Last Exam, ranked by overlap. Discovered automatically through the match graph.
LiveBench
Continuously updated contamination-free LLM benchmark.
Ask Pandi
Answer engine to search and generate knowledge
mmlu
Dataset by cais. 4,39,045 downloads.
Baidu: ERNIE 4.5 21B A3B Thinking
ERNIE-4.5-21B-A3B-Thinking is Baidu's upgraded lightweight MoE model, refined to boost reasoning depth and quality for top-tier performance in logical puzzles, math, science, coding, text generation, and expert-level academic benchmarks.
WMDP
Benchmark for dangerous knowledge in LLMs.
Genspark.ai
Transforms queries into real-time, customized pages with expert...
Best For
- ✓AI safety researchers evaluating frontier model capabilities
- ✓academic institutions benchmarking AI systems against disciplinary standards
- ✓organizations assessing whether AI has reached superhuman performance thresholds
- ✓AI researchers publishing model evaluations in peer-reviewed venues
- ✓frontier AI labs benchmarking against published standards
- ✓safety researchers establishing capability baselines before deployment
- ✓domain experts wanting to contribute discipline-specific questions
- ✓AI labs evaluating models on continuously-updated standards
Known Limitations
- ⚠Contamination detection methodology not publicly documented — unclear how 'searchable questions' were identified beyond web search
- ⚠Disciplinary representation balance unknown — no published breakdown of question distribution across fields
- ⚠Expert contributor pool composition not enumerated — potential biases in which disciplines/institutions are overrepresented
- ⚠Replacement questions sourced after bug bounty may not have undergone identical vetting as original questions
- ⚠Scoring methodology not publicly documented — unclear if evaluation is accuracy, pass@k, partial credit, or human-graded
- ⚠No baseline or SOTA performance data provided — cannot contextualize model scores
Requirements
Input / Output
UnfragileRank
UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.
About
Collaborative benchmark compiling the hardest exam questions from thousands of experts across every academic discipline, designed to be the ultimate test of AI knowledge and reasoning before superhuman performance.
Categories
Alternatives to Humanity's Last Exam
Build high-quality LLM apps - from prototyping, testing to production deployment and monitoring.
Compare →Amplication brings order to the chaos of large-scale software development by creating Golden Paths for developers - streamlined workflows that drive consistency, enable high-quality code practices, simplify onboarding, and accelerate standardized delivery across teams.
Compare →Are you the builder of Humanity's Last Exam?
Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.
Get the weekly brief
New tools, rising stars, and what's actually worth your time. No spam.
Data Sources
Looking for something else?
Search →