Researchers have proposed a reinforcement learning method that lets language models improve their reasoning using their own internal feedback, rather than relying on ground-truth answers or external verifiers. The approach, described in a recent arXiv paper, could make training more self-contained.
However, the authors caution that confidence and correctness can diverge during training. This means a model may become more confident while its actual accuracy does not improve, or vice versa. The abstract is cut off mid-sentence, leaving the exact consequence of continued optimization of average performance unspecified.
The work highlights a potential trade-off in using internal signals for training, and points to the need for careful monitoring of model confidence.