Researchers have proposed a new approach to improving the performance of large language models using weak critics. Instead of relying on strong labels or judgments from human evaluators, they use a weaker model as a critic that provides revision directions for the stronger model. This approach, called weak-critic strong oversight, has been shown to improve frozen strong models at inference time and can be distilled into the strong model through adaptive self-teacher signals. The authors propose progressive on-policy critique distillation (OPCD) as a method for filtering high-quality critiques and improving strong models over training epochs. --- Why it matters: This matters because it provides an effective path for scalable oversight with weak supervision, which is essential for large language models that become increasingly complex and difficult to evaluate. Source: https://arxiv.org/abs/2606.00424