Clean data can strengthen backdoor poisoning in LLM fine-tuning

Owain Evans and coauthors report that adding clean data to a poisoned fine-tuning dataset can strengthen backdoor misalignment when the poison includes an identifying trigger.

The authors call this effect homeopathic learning, because diluting the poison can make it more effective. The report, posted on X by Dongxi NLP, does not name the models tested, their sizes, or any quantitative results.