Reproduce “ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost” 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
Xingzhi Qian, Sungho Park, Rongyuan Tan, Jianbo Yu, Jue Zhang, Lorenzo Cavallaro, Menglin Xia, Yanzhi Chen and 8 more
ParetoSaddler optimizes an agent’s harness without changing its language model. It promotes candidates when repeated-run evidence supports progress in quality or cost, leaving deployment preferences until final selection. On ClawMark, its selected harness scored 81.35% ± 3.11 at $1.01 per task, versus 70.86% ± 2.29 and $1.17 for the default; evidence spans two benchmarks, not unrestricted deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.