PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchThe “Humanity’s Last Exam” headline described a real 2024 call for expert-written questions, but the project is no longer merely being prepared. Created by the Center for AI Safety (CAIS) and Scale AI, HLE is now a published, multimodal, closed-ended benchmark of difficult academic questions. Its early models scored below 10%, yet neither that result nor a future high score can by itself prove or disprove artificial general intelligence.
The 2024 headline is outdated
In September 2024, CAIS and Scale AI asked experts to submit exceptionally difficult questions for a new AI benchmark. The campaign sought problems from mathematics, science, engineering, philosophy, humanities and other fields, while excluding weapons-related material. Selected contributors could receive up to $5,000 per question under the original prize structure. The announcement is documented in the original coverage.
As an Amazon Associate I earn from qualifying purchases.
That crowdsourcing phase produced Humanity’s Last Exam, usually abbreviated HLE. Scale published initial results in January 2025, and a peer-reviewed paper describing the benchmark appeared in Nature on January 28, 2026. The continuing leaderboard should therefore be treated as a living evaluation resource, not as the proposal described by the 2024 story.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →What Humanity’s Last Exam is
HLE is an expert-level academic knowledge-and-reasoning benchmark. Models receive questions with specified answers or solutions, and the evaluation checks whether their responses match the expected result. Some questions are text-only; others include images, diagrams, charts or other visual material.
#1 Best Overall
The benchmark spans mathematics, physics, chemistry, biology and medicine, computer science and artificial intelligence, engineering, the humanities, social sciences and other specialized areas. The published Nature paper describes 2,500 challenging questions from dozens of subject areas. Other official Scale pages describe 2,700 or 3,000 questions for particular releases or broader project variants. Those figures should not be combined into one unexplained total: readers need the specific benchmark version and modality attached to any count.
The project’s paper and resources are available through Nature, the Scale paper page and the CAIS project page.
Why researchers wanted a harder benchmark
Many widely used tests had become less informative at the frontier. The Nature paper notes that leading large language models were exceeding 90% accuracy on popular tests such as MMLU. Once questions, formats and answer patterns become familiar—or enter training data—a score can reflect memorization, benchmark-specific optimization or test adaptation as well as general reasoning.
HLE was designed to push the difficulty ceiling upward. Its purpose is not to declare earlier benchmarks useless, but to provide questions that remain challenging for systems already performing strongly on standard academic exams.
Rank #2
How the exam was built
Expert contributions and review
The original call solicited questions from people with academic or professional expertise. Organizers wanted items that were difficult for non-specialists, grounded in real knowledge, resistant to superficial pattern matching and unlikely to be solved through ordinary web lookup. Submissions were reviewed and selected before inclusion.
Expert authorship improves the chance that a question reflects genuine specialist knowledge, but it does not guarantee that every item is unambiguous, current or error-free. Questions can still contain unclear wording, transcription mistakes, outdated premises or an answer-key problem.
Closed-ended scoring
Because HLE is closed-ended, evaluation can compare a response with a specified solution rather than asking a judge to rate an open-ended essay. The official leaderboard reports accuracy and a calibration measure, which compares a model’s confidence with how often it is correct. Details are available on the official leaderboard.
Multimodal items
Some questions require interpreting visual information. A result on a text-only preview is therefore not automatically comparable with a result on the full multimodal benchmark.
Safety limits in the original call
The submission campaign excluded questions involving dangerous weapons. That was a scope and safety condition for collecting material, not evidence that HLE is a general safety test.
What the first results showed
Scale’s January 23, 2025 announcement reported that then-current models answered fewer than 10% of the expert questions correctly. The result illustrated the gap between routine benchmark performance and HLE’s deliberately difficult items; it was not a permanent score for “AI” as a whole. The announcement is at Scale AI.
Later leaderboard entries must be read with their complete evaluation conditions. A score is meaningful only alongside the model checkpoint and date, benchmark version, text-only or multimodal setting, prompt, tool access, answer-extraction method and judging protocol. Public pages include several variants, including the text-only preview and the text-only leaderboard. Quoting a bare percentage without those details can create a false impression of a single universal ranking.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsWhat a score can—and cannot—tell us
What HLE can measure
- Performance on difficult, closed-ended academic questions.
- Coverage across many specialist disciplines.
- Differences between model versions under a stated evaluation protocol.
- Whether a system can combine learned knowledge and reasoning to reach a specified answer.
- Confidence calibration, when the leaderboard reports it.
What it does not establish
- Original scientific discovery or the ability to design and run a novel experiment.
- Long-horizon planning, project management or reliable real-world execution.
- Safe and robust use of tools outside a controlled test.
- Consciousness or human-like understanding.
- Alignment, honesty, resistance to jailbreaks or protection against cyber, biological or persuasion risks.
- Artificial general intelligence.
The benchmark’s own documentation explicitly cautions that HLE performance alone would not demonstrate autonomous research capability or AGI. See the Scale text-only documentation. A low result means that a system struggled with these questions under that protocol; it does not mean the system is incapable at every useful task. A high result would show stronger academic capability, not comprehensive reliability.
Why “last” is branding, not a scientific endpoint
The name signals the organizers’ ambition to create a final, unusually difficult closed-ended academic benchmark of this kind. It does not mean that no harder test can ever be designed, that the questions will remain permanently unseen, or that the benchmark can settle the meaning of intelligence. Static tests become easier to optimize against, and public questions can eventually become training material.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The main measurement problems
Contamination after publication
Once questions, answers and discussions are public, future models may encounter them directly or through derivative datasets. Public scores remain useful for tracking progress, but they are weaker evidence of performance on genuinely unseen problems than contamination-controlled or refreshed private evaluations.
Public transparency versus private security
Public questions let outsiders inspect items, verify answers and reproduce results. Private questions are harder to memorize but less independently auditable. HLE’s value depends on balancing those competing goals.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Question and answer quality
Difficulty can come from deep reasoning, but it can also come from obscure notation, OCR failure, ambiguity or a faulty premise. Community researchers have proposed verification and revision work and raised concerns about noisy items; these are critiques of particular questions and processes, not proof that the entire benchmark is invalid. Their proposal is available at arXiv.
Best Value
Tools, prompts and judging
Browsing, code execution, retrieval and calculators can materially change a result. So can the instruction to provide only an answer versus a prompt that encourages extended reasoning. Automated extraction or judging can also misread an otherwise correct response. Comparisons are credible only when these settings are reported.
Human comparisons
A broad exam assembled from many specialties does not automatically measure whether AI has “beaten humanity.” A fair human baseline would need to specify who took the test, whether participants were domain experts, what references and time they had, whether they answered only questions in their fields and whether partial credit was available. One person’s score on a cross-disciplinary exam is not a universal human-intelligence benchmark.
How to interpret HLE results responsibly
- Identify the release. Check whether the result comes from the published benchmark, a text-only preview, a multimodal set or another leaderboard variant.
- Record the model snapshot. Model names can cover changing checkpoints; note the exact version and evaluation date.
- Check tool and prompt conditions. Browsing, coding, retrieval and answer-format instructions can change accuracy.
- Look for contamination risk. Public questions and answers may have appeared in later training data.
- Read the score as one capability signal. Do not convert academic accuracy directly into a claim about AGI, safety or real-world competence.
The bottom line on “Humanity’s Last Exam”
HLE is a serious and unusually demanding benchmark, created by CAIS and Scale AI and now documented in a peer-reviewed paper. It is useful for asking how well models handle broad, expert-level academic questions under specified conditions. It is not a literal final exam for civilization, a consciousness test, a safety certification or a standalone AGI detector. The most accurate description is a difficult, evolving measuring instrument whose results are meaningful only when its version, data, tools and scoring rules are made explicit.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
