Multi-Agent Debates Can Induce Fabricated Citations
drichko · reddit · 2026-07-19
The author tried having multiple LLM personas debate each other, followed by a neutral step to extract disagreements, aiming to mitigate the single-model "sycophancy" issue.
Two unexpected phenomena were discovered:
- Models entering a "winning" state are more prone to fabricating citations: They invent URLs, author names, and specific numbers. Rather than random hallucinations, these act more like "aggressive, persuasive hallucinations." The author thus added a deterministic check to specifically block links not present in the retrieval corpus.
- Debates become highly homogeneous if the same model generates the debaters: Low-temperature sampling quietly aligns the priors of multiple personas, making it look like a multi-party debate when it's actually just one model wearing different masks.
The author concludes that multi-agent debate is easy to fake, but achieving genuine disagreement and verifiability relies on the verification layer, not the persona design layer.
More from Research
- AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
- Flow Reasoning Models turn flow matching into iterative reasoners, hitting 99.5% on Sudoku-Extreme — alec_helbling · 2026-10-26
- Surprising find: Alzheimer's-linked amyloid fibrils deliver RNA to sperm for embryo translation — EricTopol · 2026-10-26
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- New Blog Breaks Black-Box Distillation Into Five Practical Steps — chrisrohlf · 2026-10-26