Brief, mostly facts
Agent-written research papers fail expert standards on DOE problems
Into the Unknown tested 135 follow-up papers. None met expert standards, and 114 failed integrity checks for fabricated or contradicted numbers.
What the research found
Same paper as #389 (arXiv 2610.19159) on Hugging Face Papers, confirming the 135 submissions, 114 integrity failures and public code and data; this is a duplicate story, so keep one entry.
Claims and how well the evidence supports them
- Nine frontier agents wrote 15 follow-up papers each, 135 submissions total 99%
- 114 of the 135 submissions fail the integrity check 99%
- Code and data for the benchmark are publicly available 98%
Evidence
- huggingface.co https://huggingface.co/papers/2610.19159
- arxiv.org https://arxiv.org/abs/2610.19159
Actionability rated low.