1
Inference-Time Diversity in RL-Trained Lean Theorem Provers: A Diagnostic Study
诊断强化学习训练的Lean定理证明器在推理时的多样性缺陷,为AI数学推理的鲁棒性提供新视角。
arXiv:2601.16172v3 Announce Type: replace Abstract: RL-trained Lean theorem provers mode-collapse at inference time: on miniF2F-test with DeepSeek-Pro…