NeurIPS 2026 Review Stats: 5 of 8 Papers Had Hallucinated Citations
danish037 · x · 2026-08-18
An Area Chair for NeurIPS 2026 shared alarming stats from the initial review phase.
- 5 out of 8 submissions in their lot contained 2+ hallucinated references and are likely desk rejected.
- Citations are easy to verify, making these hallucinations easy to catch, suggesting there may be even more "slop" elsewhere in the papers.
Another post noted that one of the remaining papers had a fabricated citation, which wasn't enough for immediate rejection, highlighting the conference's somewhat liberal definition of hallucinations.
Related event: NeurIPS 2026 Reviewer Finds Fabricated Citations in Majority of Submissions(2 posts)→
More from Research
- Formal Verification, Explained from First Principles — dhruv2038 · 2026-08-18
- HarnessEval-W: agent-based benchmark makes world model evaluation auditable — _akhaliq · 2026-08-18
- MolmoAct demo in Singapore: Improving VLA reasoning interpretability — DJiafei · 2026-08-18
- PySR figures out the rules for Conway's Game of Life — MilesCranmer · 2026-08-18
- Physicist: most AI-for-science startups will die, but mining existing literature is the real win — skdh · 2026-08-18
- Testing MTP combined with ngram-mod for coding speed — YetAnotherAnonymoose · 2026-08-18