Qinan Yu Alexa R. Tartaglini Peter Hase Carlos Guestrin Christopher Potts Outcome Rewards Do Not Guarantee Verifiable or Causally Important Reasoning. 2026 April abs/2604.22074 CoRR https://doi.org/10.48550/arXiv.2604.22074 db/journals/corr/corr2604.html#abs-2604-22074 streams/journals/corr