Tencent Hunyuan develops Agent version of RSI! Based on Qwen3.8-27B, terminal task success rate doubles

What happened: Tencent Hy Frontier, led by Tencent chief AI scientist Yao Shunyu, with researchers from the National University of Singapore, the University of Maryland, the University of Georgia and Indiana University, proposed Terminal-RSI, a recursive training method for terminal agents built on Alibaba Why it matters: The paper says gains varied by task type, with extremely hard tasks fluctuating rather than improving steadily, and results were validated only on a single 27B model with three evaluations per stage, which the authors say is not enough to claim statistical superiority.

Tencent Hy Frontier, led by Tencent chief AI scientist Yao Shunyu, with researchers from the National University of Singapore, the University of Maryland, the University of Georgia and Indiana University, proposed Terminal-RSI, a recursive training method for terminal agents built on Alibaba

The paper says gains varied by task type, with extremely hard tasks fluctuating rather than improving steadily, and results were validated only on a single 27B model with three evaluations per stage, which the authors say is not enough to claim statistical superiority.