Tencent Hunyuan's research agent Hyra, powered by its open-source model Hy3, helped two researchers crack a combinatorial math problem that had been open for over 50 years.
The problem: how much faster can the sumset of a set of integers grow compared to its difference set? Mathematicians already knew the exponent couldn't exceed 2, but they didn't know if 2 was actually the best possible.
Hyra tasked Hy3 with first searching for specific sets, then trying to generalize the construction. After about 24 hours, the AI found the core approach that ended up in the paper. The researchers then checked, corrected, and formalized the proof, with GPT-5.6 Sol helping translate it into Lean 4 for verification.
The final paper proves the exponent can approach 2 arbitrarily closely, meaning the upper bound of 2 is tight. While the AI didn't write the whole paper, the crucial construction was indeed found by Hy3.
https://twitter.com/TencentHunyuan/status/2082655737541726636