Reproduce “CANDID: Candidate-Driven Intervention for Disciplined Execution in Local Computer-Use Agents” 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
Woongkyu Lee, Jungwook Choi
CANDID uses selective candidate sampling to address repeated actions and premature task completion in local computer-use agents. At detected loops, it filters repeated actions; when the agent proposes DONE, it keeps that decision only if all candidates agree. On OSWorld, it raises Qwen3-VL-30B-A3B’s accuracy from 35.31% to 38.80%, but unanimous gating is not a semantic completion verifier and configurations are model-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.