On July 29, 2026, Tencent Hunyuan's research agent Hyra, working with mathematicians Lin Haowei (Carnegie Mellon University / Peking University) and Li Shanda, closed a 57-year-old question in additive combinatorics: the optimal exponent relating the sumset \(|A+A|\) and difference set \(|A-A|\) of a finite set of integers \(A\) is exactly 2.
What the Problem Asks
Given a finite integer set \(A\), form two derived sets: \(A+A\) (all pairwise sums) and \(A-A\) (all pairwise differences).
In 1969, a classical inequality established the range:
The exponent 2 on the right was long known to be tight. But the exponent 1/2 on the left — equivalently, how large the ratio \(C(A) = \log \sigma(A) / \log \delta(A)\) can get — had no clear construction for over 50 years. The best explicit constructions previously achieved only \(C(A)\) slightly above 1.1.
The open question: can \(C(A)\) really approach 2, or is it blocked at some unknown ceiling in the 1.x range?
What Was Proven
Lin and Li give an explicit construction \(A_K\) (for positive even \(K\)) satisfying:
As \(K \to \infty\), this lower bound converges to 2. Combined with the classical upper bound, this means the minimal exponent \(c\) such that \(|A+A| \le |A-A|^c\) holds for all finite integer sets \(A\) is exactly 2.
The construction uses four ingredients: a base-12 digit gadget, a 3-state carry automaton, a symmetric additive basis, and a Chinese remainder theorem product structure. The full proof comes with Lean 4 / Mathlib formal verification, running 2,212 build tasks, with all core theorems passing no-sorry checks.
What Hyra Did
Hyra is Tencent Hunyuan's in-house research agent, built on the open-weight model Hy3. Its workflow is a SimpleTES evaluation loop plus autonomous search:
- Automatically constructing candidate sets \(A\), scoring them by \(C(A)\), and feeding the highest-scoring solutions back to itself for the next round of mutation.
- Before Lin and Li took over, Hyra had already pushed the best known automatic search value from about 1.14 to 1.21.
- Lin and Li hand-refined the high-\(C\) candidates from Hyra, derived the closed-form construction, and completed the proof.
- The Lean 4 formalization itself was carried out by the Hunyuan team, ensuring no-sorry verification.
- Peer review is not complete; this is an arXiv preprint.
- Hyra did not complete the proof independently; 1.21 is still 0.79 short of 2, and that gap was closed by hand by the mathematicians.
- Whether the same exponent result holds in other groups (\(\mathbb{Z}^n\), finite fields) was not addressed.
- arXiv paper: https://arxiv.org/html/2607.27199v1
- Hyra project page: https://hy.tencent.com/research/hyra
- Tencent Hunyuan announcement on X: https://x.com/TencentHunyuan/status/2082655737541726636
The arXiv paper (2607.27199) acknowledges the Hyra team; the authors are Lin Haowei (Tencent Hunyuan + CMU) and Li Shanda (Tencent Hunyuan). This is not an "AI-only proof" but a collaboration model where "AI pushed the search space from 1.14 to 1.21, and human mathematicians capped it at 2."
Why This Matters
1. Additive combinatorics gains a credible AI entry point. Prior automatic search tools (AlphaEvolve, SimpleTES, EvoMaster) only pushed the record into the 1.14–1.28 range; this result approaches the theoretical limit, showing the research-agent + Lean formalization pipeline can reliably produce publishable results. 2. A clear "AI as tool, not author" boundary. The arXiv authors are the mathematicians; Hyra appears in the acknowledgments — a replicable template for academia: AI maximizes automatic search, humans close the proof. 3. Lean 4 formalization is now a must-do, not a bonus. 2,212 builds and no-sorry verification at this intensity was almost nonexistent in AI math projects five years ago. 4. A clear position for Tencent Hunyuan in AI for Math. Together with AngelSpec speculative decoding for inference infrastructure, Hyra sets a de facto standard for AI for Math — the two landed within 24 hours of each other, not coincidentally.
Caveats
Still, as a demonstration that research agents can reliably push the frontier of pure mathematics, this is the most convincing example of 2026 so far. The three-stage pipeline — Lean formalization + autonomous search + human completion — may become the standard workflow for AI for Math.
References: