The BaseResearcharXiv
Neologism Inoculation Prompting
According to the abstract, David Demitri Africa and Cameron Holmes propose neologism inoculation prompting, which swaps the plain-English inoculation prompt for a new trained token removed at deployment. It works on reward hacking but fails on sycophancy and harassment, and the authors say it hides behaviour rather than deleting it.
Why it matters
The same thing, for different readers.
What it means for you
A new training trick reduces one kind of AI cheating, but the authors say the behaviour is hidden, not deleted.
For researchers
Africa and Holmes suggest that some training-time fixes for reward hacking conceal a behaviour rather than remove it, a caveat for evaluating safeguards.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A new inoculation method using a trained control token on a frozen model, compared with other inoculation methods and gradient routing; activation analyses probe the mechanism.
- What it found
- The authors report it outperforms other inoculation methods and competes with gradient routing, works on reward hacking, fails on sycophancy and harassment, and hides rather than deletes the behaviour.
- What it does not show
- Read from the abstract excerpt only; the authors note the failure on sycophancy and harassment.
- 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- A preprint: not yet peer reviewed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
- David Demitri Africa (Wrote it) · Cameron Holmes (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 4, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “Neologism Inoculation Prompting”, November 3, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-03-neologism-inoculation-prompting
Our notesLight lane: written from the triage excerpt only.