Reproduce “Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards” 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
Tianyu Liu, Davide Belli, Bence Major, Arash Behboodi, Tribhuvanesh Orekondy
Explorer-Zero trains language-model agents for stateful tasks by having a proposer discover and demonstrate tasks, then using recorded intermediate states as solver checkpoints. The proposer and solver train in alternation, with solver progress shaping task difficulty. Across three environments, it outperforms evaluated self-play baselines without human-authored training tasks; on AppWorld test_normal, it scores 59.67 ± 3.1 TGC, below full human-task RL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.