Reproduce “Test-Time Self-Play: Scaling without a Verifier” 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
Akira Yoshiyama, Akmal Ashirmatov, Jonas Hübotter, Ido Hakimi, Andreas Krause
TTSP adapts a model to one programming task by training proposer and solver roles on target-conditioned synthetic problems checked by execution. It excludes target-task verification from training but still uses synthetic-task tests. On 28-problem HardLCBv6, it reports an absolute +0.35 mean@8 gain without target-answer filtering, while gains are approximately zero when base answer coverage is very low.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.