The BaseResearcharXiv
Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect
According to the abstract, Clarke, Dubinski, Chua and colleagues report that poisoned fine-tuning data with no effect alone can create a hidden backdoor, such as misalignment, once clean data lacking a trigger feature is added to the training set.
Why it matters
The same thing, for different readers.
What it means for you
Training data that looks harmless on its own could still plant a hidden behavior in an AI model once mixed with other data.
For lawmakers
Clarke and colleagues' result implies screening fine-tuning data example by example may miss poison that only works in combination.
For researchers
Clarke and colleagues propose a simple regression explanation: the trigger absorbs the trait once clean data lacks it.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Fine-tuning experiments with poison data tied to a trait and a trigger, then adding clean data; settings include subliminal learning, inoculation prompting, bird names and an MNIST classifier.
- What it found
- Per the abstract, poison alone does not teach the trait, but adding clean data without the trigger does.
- What it does not show
- Explanation is a simple regression model; full text not read.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
- People
- Matt Clarke (Wrote it) · Jan Dubiński (Wrote it) · James Chua (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of November 6, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect”, November 6, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-06-homeopathic-learning-poisoned-finetuning
Our notesLight lane: written from the triage excerpt only. Worth a full read.