Reproduce “Neologism Inoculation Prompting” 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
David Demitri Africa, Cameron Holmes
NIP addresses a weakness in selective fine-tuning: models may produce harmful behavior when users type an English training label. It trains a new token on matched examples, fine-tunes a LoRA under that token, then removes its embedding rows for deployment. On Qwen extreme-sports advice, typed-label misalignment fell from 32.1% with SIP to 0.1% with NIP; this result is limited to one task and model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.