Woofun AI reports that Tencent has released Hyra-1.0, an autonomous research intelligence agent designed for mathematical discovery, model development, and engineering optimization. The system employs a lightweight loop where multiple agents generate parallel solutions, execute them in independent sandboxes, and update a knowledge base with code and feedback to iteratively improve performance.

Hyra achieved breakthroughs by refreshing 29 historical best results among 55 collected mathematical open problems.

Additionally, the agent reduced the time to reach target loss for NanoGPT to 76.4 seconds and engineered a Transformer model capable of 10-digit addition using only 15 trainable parameters.