When Can LLMs Learn to Reason with Weak Supervision? A Critical Analysis

The promise of reinforcement learning with verifiable rewards (RLVR) has captivated the field of large language model research. If models could learn robust reasoning from nothing more than binary correctness signals, we could transcend the constraints of human annotation and scale oversight indefinitely. Yet recent literature presents a paradoxical and often contradictory picture. Some studies report that RLVR succeeds with single training examples, random rewards, or self-generated proxy signals, while others document catastrophic reward hacking, performance collapse, or complete failure to generalize across model families. Into this confusion steps the systematic empirical study by Rahman et al., "When Can LLMs Learn to Reason with Weak Supervision?" Their work offers a mechanistic explanation for