AI RESEARCH
Exploiting Verification-Generation Gap: Test-Time Reinforcement Learning with Confidence-Conditioned Verification
arXiv CS.LG
•
ArXi:2606.03608v1 Announce Type: new Test-time reinforcement learning has emerged as a promising paradigm for enhancing the complex reasoning abilities of large language models in a completely label-free manner. Despite existing studies focusing on Pass performance, optimizing Pass remains under-explored yet critical in label-free settings, which measures generation coverage for sustained exploration. Optimizing Pass in label-free setting is highly non-trivial, as directly applying the Pass advantage designs effective for RLVR yields unsatisfactory performance.