Reproduce “Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce Misalignment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ariana M. Villegas-Suarez, Guido Freire, Sohan Venkatesh, Bartosz Cywiński
The authors test whether fine-tuning on harmless false answers is associated with broader misaligned responses. In Qwen3.6-27B, they vary how wrong mathematics answers are and measure the model’s prior belief in them alongside misalignment on held-out Betley questions. Misaligned answers rise from 4% to over 60%, a result from supervised LoRA experiments, not a real-world harm estimate or proof of causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.