This paper investigates how hallucinations arise and persist in RLLM reasoning, revealing error self-reinforcement and limited metacognition.

This paper demonstrates that language models possess metacognitive-like abilities to monitor and control their internal neural activations.