Yonsei and FuriosaAI's CaLIF Cuts Response-Quality Collapse in Verifiable-Reward Training by 7x

A Yonsei–University of Washington–FuriosaAI paper posted October 15 shows OLMo 3 models trained with standard IF-RLVR raise IFBench constraint satisfaction from 26.98 to 44.93 but collapse response quality from 4.51 to 2.96 on a 10-point scale. Their CaLIF reward uses two pre-RL anchor responses to bound likelihood, preserving quality at 4.59 while lifting constraints to 34.64 and reducing per-pass quality degradation 7x. CaLIF needs no extra models or judges and beat IF-RLVR on 40 of 45 comparisons across AdvancedIF, FollowBench and ComplexBench.