The Hyra smart agent doesn’t just automatically do math—it can also improve its scoring criteria on its own, fend off anti-cheating, and even use a 15-parameter Transformer to handle 10-digit addition, marking a big step forward for research automation.

CoinWorld News: Tencent Hunyuan released a research AI agent, Hyra. Among 55 collected open math problems, Hyra refreshed 29 historical best results. Hyra can automatically propose solutions, run experiments, evaluate results, and continue to improve based on experience from both success and failure, targeting tasks such as model development, mathematical discovery, and engineering optimization. Its core is a lightweight iterative loop: multiple agents generate solutions in parallel; the system runs in an independent sandbox and scores them, then writes feedback back into the experience repository. Hyra can also improve scoring criteria without a ready-made evaluator, reducing cases where agents inflate scores by exploiting rule loopholes. In addition, Hyra shortened the time to reach the target loss for nanogpt to 76.4 seconds, and designed a transformer with only 15 trainable parameters that can complete 10-digit addition.