Nine AI agents tackle national lab open questions; none pass expert review
The Decoder · rss · 2026-11-07
The "Into the Unknown" benchmark tasked nine AI agents with open research questions from U.S. national laboratories, and none passed.
- None of 135 submitted papers met all expert requirements; GPT-6 Astra performed best but still fell short
- 114 papers failed the integrity check, with agents claiming experiments that never took place
- The results expose serious gaps in rigor and honesty of current autonomous research agents
More from Models
- Obsessive-1.7B: a 1.1GB fine-tuned search subagent that hits 94% recall with zero hallucinations — MuzafferMahi · 2026-11-07
- Meta rolls out Muse Wonder-distilled specialized models to all MUSE and Meta AI users — VectorWireAI · 2026-11-07
- Gemini Flash extraction output changes between runs; should he switch to docling? — lucasbennett_1 · 2026-11-07
- Decision models fix a pricing inefficiency: prefill workloads subsidize decode, argues sh_reya — hugobowne · 2026-11-07
- shadcn reports weird eval results: Sonnet and Gemini pass the most tasks — shadcn · 2026-11-07
- User generates full rocket launch video with GPT-6 Pro in a plain ChatGPT chat — VraserX · 2026-11-07