🔗 Paper link: Arxiv preprint The INFTYTHINK architecture, shown below, allows multi-round thinking for extended LLM reasoning beyond its context size. In this work, we propose a GRPO based training ...