Tencent Hunyuan's AI4S Hiring Push: Research Agent Hyra Cracks a 50-Year Math Problem

Yao Shunyu is personally recruiting for Tencent Hunyuan's AI for Science (AI4S) team, according to a report from Chinese tech outlet Quantum Bit. There is no detailed job description — just a note saying 'Hunyuan AI4S is hiring :)' and a single image.

The image is a research scorecard: Tencent Hunyuan's AI, via its in-house research agent Hyra, pushed an open problem that had stumped mathematicians for about 50 years to a new best result. The subtext is clear — the company's AI has started doing research on its own, and it needs more people to expand the gains.

The problem belongs to additive combinatorics: given a set of integers, how many distinct answers can you get by adding any two of them, and how many by subtracting? Previous theory set an upper bound — the growth rate of addition results can at most approach the square of the subtraction results — but mathematicians have debated for over 50 years whether that bound is a loose estimate or can truly be approached.

Over half a century, mathematicians kept designing new number sets to approach the bound: the 1969 result was about 1.0290, improved to 1.0598 in 1973, and pushed to 1.1259 in 2013; in the past year, multiple AI-assisted searches raised the record to 1.1449, and Codex reached 1.2851 in human-guided experiments.

Hyra, built on the Hy3 model open-sourced this month (295B total parameters, 21B active), found a family of constructions that can be expanded indefinitely and proved that as the set grows, the result can approach the theoretical upper limit of 2 arbitrarily closely — the seemingly unreachable bound can genuinely be approached.

Hunyuan stresses this was not simple brute-force search: Hyra first found better results in a limited range, then switched to proposing mathematical constructions and arguments in natural language, finding the core idea of the paper after about 24 hours of running. The research team then independently checked and organized the complete proof, and produced a Lean 4 formal proof so a computer could verify the derivation step by step.

From finding clues and proposing constructions to delivering a verifiable full proof, Hyra is no longer just helping mathematicians search for answers — it is participating in real mathematical research. This builds on Hyra-1.0, the research agent Tencent Hunyuan launched a few days ago, which entered fields including mathematics, astronomy, quantum computing, and drug design in its first AI4S tests.

The hiring intent follows: on one hand, Hyra-1.0 builds the research loop and Hy3 provides core model capability, making expertise in agent architecture, reinforcement learning, in-context learning, automatic evaluation, training systems, and GPU kernels a hard requirement; on the other hand, Hunyuan also needs cross-disciplinary people who understand both AI and specific scientific questions. Last year, the Tencent Hunyuan large-model team already recruited data interns with physics, chemistry, or biology backgrounds.

In the broader industry context, this is not an isolated move: a hot comment in this year's Fields Medal discussions holds that this may be the last Fields Medal awarded for purely human achievement. AI is evolving from a supporting tool into a genuine researcher. The question to watch is whether Hyra can replicate its breakthrough in other disciplines and how far this 'human-machine research team' can push science.