AI reasoning training fails on harder tasks when models switch to guessing

Original: Not All Reasoning Generalizes: A Case Study on Learning Generalizable Reasoning Strategies in RLVR

Source: arxiv.org ↗