Reproduce “Agents Can Use Base Models to Evade AI Detection” 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
Bhuwan Dhingra, Danish Pruthi
HALO separates planning from word selection by letting an agent sample continuations from a base language model and splice copied spans with text. On 100 IFEval prompts, Claude Opus 5 with HALO scored 0.94 strict instruction following versus 0.96 without it, while Pangram fraction_AI fell from 0.965 to 0.232. These results concern tested outputs and metrics, not human authorship.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.