Evaluation & BenchmarksFutureHouse

Humanity's Last Exam bio/chem error rate put at 29%; HLE's own review finds 18%

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

Raised a 29% error rate in the biology and chemistry questions of Humanity's Last Exam; the HLE team's own review put it at 18%, and it announced rolling revisions

Why it matters

A case that showed, through public refutation and response, that the answer keys of frontier benchmarks can themselves be wrong

Source — Primary sources first; secondary reports are labelled “Press”

Officialfuturehouse.orghttps://www.futurehouse.org/research/hle-exam

Added in Week 35, 2026 · See that week

On the map — ⑯ Evaluation line

  1. Previous stopTechCrunch reports criticism of Epoch AI for not disclosing OpenAI funding
  2. This recordHumanity's Last Exam bio/chem error rate put at 29%; HLE's own review finds 18%
  3. Next stopGDPval
See it on the map

Same week

Scale AI releases SWE-Bench ProPaperOpenAI and Apollo Research stress-test anti-scheming trainingPaperxAI releases Grok 4 FastOfficialDiscovery order in In re OpenAI Copyright Infringement Litigation (25-md-3143)Official

Nearby on the same axis

Scale AI releases SWE-Bench ProPaperMeta releases GAIA2 and the open-source execution framework AREOfficialOpenAI releases GDPvalOfficialAndon Labs releases Vending-Bench 2 and the multi-agent Vending-Bench ArenaOfficial