Tencent’s Hunyuan team just dropped Hyra-1.0, an AI research agent that can propose solutions, run experiments, evaluate results, and then learn from what worked and what didn’t.
Hyra’s core is a lightweight loop. Multiple agents generate solutions in parallel, the system runs them in isolated sandboxes and scores them, then writes code, logs, and feedback back to an experience library. If there’s no ready-made evaluator, Hyra can also improve its own scoring criteria to prevent agents from gaming the system.
According to Tencent, Hyra achieved 29 best-ever results on a collection of 55 open math problems. It also reduced the time for NanoGPT to reach target loss to 76.4 seconds, and designed a Transformer with only 15 trainable parameters that can perform 10-digit addition.