ANCORA bootstraps verifiable reasoning from zero human solutions via manifold-anchored self-play, raising Dafny2Verus pass@1 from 26.6% to 81.5% while outperforming prior self-play baselines.
Empirically ˆτt =s_ability_ema evolves slowly (per-step changes <0.05 in the first 200 steps), giving an order-of-magnitude check ΥT ≲10 in our main run
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
ANCORA: Learning to Question via Manifold-Anchored Self-Play for Verifiable Reasoning
ANCORA bootstraps verifiable reasoning from zero human solutions via manifold-anchored self-play, raising Dafny2Verus pass@1 from 26.6% to 81.5% while outperforming prior self-play baselines.