According to DynamicBeat Monitoring, Tencent has released the research intelligence agent Hyra-1.0. It automatically proposes solutions, runs experiments, evaluates results, and then continues to improve based on successful and failed experiences, targeting tasks such as model development, mathematical discovery, and engineering optimization.
At the core of Hyra is a lightweight loop. Multiple agents generate solutions in parallel, the system runs and scores them in independent sandboxes, and then writes back the code, logs, and feedback to the knowledge base. When there is no ready-made evaluator, Hyra can also simultaneously improve scoring criteria, reducing the agent's reliance on loophole exploitation to inflate scores.
Tencent stated that Hyra has refreshed 29 historical best results out of 55 mathematical open problems collected. It has also reduced the time to achieve the target loss of NanoGPT to 76.4 seconds and designed a Transformer with only 15 trainable parameters that can perform 10-digit addition.
