Reproduce “Rome Wasn't Built in One Turn: Replay-Grounded Reward Signals for Long-Horizon Freeciv 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
Yanchen Liu, Xinyu Che, Brihi Joshi, Renzo Fukuda, Junhua Liu, Ruifeng Li, Yunfei Ge, Hang Yan and 1 more
The authors introduce CivTelescope, a judge trained on replayed Freeciv continuations, to provide policies with intermediate feedback about which positions achieve higher terminal scores. Its labels average outcomes from seeded branches played by built-in bots, not guaranteed policy outcomes. With 80 turns remaining, training with CivTelescope produced an 8.94-point phase gain, versus 5.04 for sparse reward, on held-out starts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.