Researchers have developed a new framework called INFUSER that improves the reasoning abilities of language models through self-evolution. The framework consists of two co-evolving roles: a Generator that drafts questions and reference answers from unstructured documents, and a Solver that improves by training on these questions and answers. The Generator is rewarded for proposing questions that would improve the Solver's performance, rather than just hard ones. This approach outperforms existing self-evolution methods on several benchmarks, including Olympiad and SuperGPQA. Code for INFUSER is available online. --- Why it matters: This matters to researchers in AI because it provides a more effective way to improve language models' reasoning abilities through self-evolution, which could lead to better performance on complex tasks such as math and coding. Source: https://arxiv.org/abs/2606.09052