Reproduce “SPLICE: Structured Prompt Local Iterative Combinatorial Evolution” 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
Anish Acharya, Phillip Studans, Amit Dhanda, Sina Amini Niaki, Ninad Vivek Rao, Vishal Amarlal Khatri, Brian Verkhovsky
SPLICE optimizes prompts divided into editable and frozen sections. A proposer edits only mutable sections, while an elitist beam retains candidates and diversity. On an adversarial SST-5 test, frozen-contract SPLICE reaches 58.5% contract-aware test accuracy versus 41.4% for all-mutable prompts, but compliance falls from 96.5% to 66.8%, showing frozen text does not ensure behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.