Chinese Researchers Achieve Breakthrough: Open-Source AI Agent 'Hyra' Solves Math Problem That Stumped AlphaEvolve, GPT, and Claude

A new preprint on arXiv titled 'Settling the Optimal Exponent Relating Sumsets and Difference Sets' by Haowei Lin (Tencent Hunyuan) and Shanda Li (Carnegie Mellon University) resolves a central question in additive combinatorics that had stumped leading AI systems, including AlphaEvolve from Google DeepMind, GPT, and Claude. The problem concerns bounding the ratio of sumset to difference set growth, where previous AI-driven results only achieved exponents around 1.14, but the true upper bound is proven to be exactly 2. The breakthrough came when the researchers allowed the AI agent Hyra to propose constructions and proof sketches in natural language rather than only explicit lists, leading to a novel construction that was subsequently verified manually and through a Lean 4 formal proof. Hyra, an open-source research agent from Tencent based on the Hy3 model (a 295B MoE model released under Apache 2.0), worked for about a day with GPT-5.6 Sol used as a judge to guide the search, but the final construction was independently checked by humans. Mathematician Thomas Bloom, known for his rigorous scrutiny of AI claims, verified the result and even reformulated the theorem, listing the authors as 'Lin, Li, and Hyra', thereby acknowledging the AI as a co-author. Although the preprint is not yet peer-reviewed, the combination of a machine-checked proof in Lean and independent human verification makes this a rare and credible AI contribution to mathematics, especially notable as it comes from an open-source model, contrasting with previous closed-model achievements.