--- title: AI benchmarks are broken, and the models were passing... description: A study led by Yale researchers had physicists audit 250 rejected AI model answers across six physics benchmarks and found only 12 were genuine model mistakes;... canonical: https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: AI benchmarks are broken, and the models were passing all along | daily.dev og:description: A study led by Yale researchers had physicists audit 250 rejected AI model answers across six physics benchmarks and found only 12 were genuine model mistakes;... og:url: https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo og:image: https://api.daily.dev/og/posts/uoDF9NbMo.png og:image:alt: AI benchmarks are broken, and the models were passing all along og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # AI benchmarks are broken, and the models were passing all along **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary A study led by Yale researchers had physicists audit 250 rejected AI model answers across six physics benchmarks and found only 12 were genuine model mistakes; the rest failed due to flawed questions, wrong reference answers, or faulty grading. GPT-4's score on HLE-Physics rose from 47.3% to 78.7% after expert correction. Ethan Mollick called public AI benchmarking 'dire.' The authors caution that saturated benchmark scores don't mean models can do original physics research, since the same models still failed to solve open theoretical physics problems, but that nuance is likely to get lost as the headline finding spreads. ## Content A new paper from Yale and several top labs just torched the credibility of AI physics benchmarks, and the numbers are hard to argue with. Researchers had actual physicists audit model failures across six popular physics benchmarks. Out of 250 rejected cases, only 12 were genuine model mistakes. The other 238 failed because of bad questions, wrong reference answers, or graders that rejected correct answers. That's a 95% false-failure rate. The practical upshot: GPT-4's score on HLE-Physics jumped from 47.3% to 78.7% after expert review. The model wasn't struggling with physics. The benchmark was struggling with accuracy. Ethan Mollick put it plainly: "The state of public AI benchmarking is dire and is undermining our ability to understand how good AI is now." Most famous measures are already maxed out, and the ones that aren't are riddled with errors that systematically underestimate model ability. This matters because benchmarks are the primary way the field - and the press, and investors, and regulators - tracks AI progress. If the scorecards are broken, everyone's working from bad information. Capability debates, safety arguments, competitive claims between labs: all of it rests on scores that apparently can't be trusted. There's an important caveat the authors are careful to make: corrected benchmark scores don't mean these models can do physics research. The same agents failed to fully solve any open theoretical physics problems they attempted. Passing a multiple-choice physics test and doing original research are different things. The paper isn't claiming AGI; it's claiming the measuring stick is bent. But that distinction is getting lost in the discourse. The headline finding - that frontier models nearly saturate even the hardest physics benchmarks once you correct for benchmark errors - is the kind of result that will get stripped of its caveats fast. The real problem is structural. Building good benchmarks is slow, expensive, and unglamorous. Labs have every incentive to publish scores on existing tests rather than fund the expert review needed to validate them. So the benchmarks stay broken, the scores stay misleading, and everyone keeps arguing about capability curves drawn on faulty data. ## Questions this post answers ### How much did GPT-4's score on the HLE-Physics benchmark change after expert review corrected grading errors? GPT-4's score on HLE-Physics rose from 47.3% to 78.7% once physicists reviewed rejected answers and corrected grading errors. Researchers found that out of 250 cases originally marked as failures, only 12 were genuine model mistakes; the remaining 238 stemmed from bad questions, wrong reference answers, or faulty automated graders across six physics benchmarks. _Anyone tracking real model capability gains should follow benchmark corrections like this one on daily.dev._ ### Does a high score on a physics benchmark mean an AI model can do original physics research? No, passing multiple-choice physics benchmarks does not mean a model can do original physics research. Even after benchmark scores were corrected upward through expert review, the same models still failed to fully solve any open theoretical physics problems they attempted, showing a clear gap between test-taking performance and genuine research capability. _Developers evaluating AI capability claims can weigh nuances like this via daily.dev before trusting a headline score._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 28 comments across x (as of 2026-09-16). **TL;DR:** Replies broadly agree public AI benchmarks have become unreliable marketing tools riddled with flawed answer keys, though several push back on whether expert re-grading itself was done fairly. **Sentiment:** 10% positive · 45% mixed · 45% skeptical **The case for** - Some see corrected benchmarks as evidence models are more capable than saturated scores suggested. - A few argue that once labs stop trusting leaderboards, real-world trial usage becomes a more honest signal. **The pushback** - Commenters argue benchmarks have become marketing assets rather than science, undermining trust in leaderboard comparisons. - Several question whether the re-grading process was itself biased, since only answers marked wrong were reviewed, not those marked right. - Some worry systematically wrong answer keys shift the ceiling for everyone, making top-of-leaderboard comparisons meaningless. - Others note flawed benchmarks mislead policy and business decisions, from regulation to corporate roadmaps built on wrong numbers. **By community** - x (skeptical): Replies largely agree benchmarks have become marketing tools with broken ground truths, while some raise methodological concerns about how the re-grading itself was conducted. **Hottest debate:** Whether the expert re-grading process was itself methodologically sound, since only originally-wrong answers were reviewed rather than a balanced sample of both directions. **Open questions** - Does the paper report re-grading in both directions (checking answers marked right, not just wrong)? - Who performed the relabeling, and was it blind to which model produced the answer? - Which benchmark, if any, can still be trusted given saturation and error rates? **Highlights** > @emollick Expert re-grading is only as fair as its direction. If experts re-read only the answers marked wrong, scores can only go up. The balanced version re-grades a sample marked right as well, since graders also accept wrong answers. Does the paper report both directions? > — [ipezyGJ on x](https://x.com/ipezyGJ/status/2100299078378647581) > @emollick Riddled with errors is itself a measurement. Who scored an item as wrong: human review, a model, or both, and was the relabeling blind to which model got it right? Otherwise the corrected scores inherit whatever produced the correction. > — [MLaurusevicius on x](https://x.com/MLaurusevicius/status/2100272612823711767) > @emollick Label errors cutting the same direction every time is the detail worth pulling out — noisy benchmarks usually add variance, but systematically wrong answer keys shift the ceiling down for everyone and make the top of the leaderboard meaningless. That's a different problem from > — [agenteyeAI on x](https://x.com/agenteyeAI/status/2100243990050177076) > @emollick The real damage happens in the boardroom. Leaders see a low score, conclude AI can't do that work yet, and plan roadmaps around a wrong number. A metric that underreports is worse than no metric. It tells you to stop looking. Are you seeing firms build their own evals yet? > — [donatienruiz on x](https://x.com/donatienruiz/status/2100236434552574391) > @emollick If we can't accurately measure AI capabilities, how do we regulate them? The benchmarking crisis isn't just a research problem — it's a policy failure. Congress is legislating in the dark if the metrics are broken. We need government-backed, adversarially-tested evaluation > — [praveenjatta on x](https://x.com/praveenjatta/status/2100244033746727353) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2100320168597753947) · 0 points · 0 comments - [x](https://x.com/emollick/status/2100233342494925257) · 0 points · 28 comments ## Similar posts on daily.dev - [The AI That Aced Every Exam by Memorizing the Answer Key](https://daily.dev/posts/the-ai-that-aced-every-exam-by-memorizing-the-answer-key-gp4wcvhwt) · Medium · 2 upvotes · 0 comments - [AI benchmarks hampered by bad science](https://daily.dev/posts/ai-benchmarks-hampered-by-bad-science-16aswsuy1) · The Register · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm), [#gpt](https://daily.dev/tags/gpt) [View this post on daily.dev](https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"AI benchmarks are broken, and the models were passing all along","url":"https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo"},"datePublished":"2026-09-16T20:26:09.193Z","dateModified":"2026-09-16T20:26:58.604Z","description":"A study led by Yale researchers had physicists audit 250 rejected AI model answers across six physics benchmarks and found only 12 were genuine model mistakes;...","image":"https://pbs.twimg.com/media/HSXPIMaa0AAGWmv.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HSXPIMaa0AAGWmv.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm,gpt","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"AI benchmarks are broken, and the models were passing all along"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/ai-benchmarks-are-broken-and-the-models-were-passing-all-along-uodf9nbmo#faq","mainEntity":[{"@type":"Question","name":"How much did GPT-4's score on the HLE-Physics benchmark change after expert review corrected grading errors?","acceptedAnswer":{"@type":"Answer","text":"GPT-4's score on HLE-Physics rose from 47.3% to 78.7% once physicists reviewed rejected answers and corrected grading errors. Researchers found that out of 250 cases originally marked as failures, only 12 were genuine model mistakes; the remaining 238 stemmed from bad questions, wrong reference answers, or faulty automated graders across six physics benchmarks. Anyone tracking real model capability gains should follow benchmark corrections like this one on daily.dev."}},{"@type":"Question","name":"Does a high score on a physics benchmark mean an AI model can do original physics research?","acceptedAnswer":{"@type":"Answer","text":"No, passing multiple-choice physics benchmarks does not mean a model can do original physics research. Even after benchmark scores were corrected upward through expert review, the same models still failed to fully solve any open theoretical physics problems they attempted, showing a clear gap between test-taking performance and genuine research capability. Developers evaluating AI capability claims can weigh nuances like this via daily.dev before trusting a headline score."}}]} ```