Reproduce “Towards Minimal Targeted Updates of Language Models with Targeted Negative Training” 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
Lily H. Zhang, Rajesh Ranganath, Arya Tafvizi
TNT trains language models to reduce specified unwanted tokens while preserving original distributions elsewhere. It generates and annotates model outputs, then targets a distribution that removes probability from marked tokens and renormalizes the remainder. On Civil Comments response generation, combined TNT methods report curve area 73.9, versus 32.9 for combined negative-loss baselines; labels are automated proxies, not human judgments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.