Peking University Team Tackles 35-Year-Old Queueing Theory Conjecture via Human-AI Collaboration
A Peking University research team successfully proved the uniqueness conjecture for BAR (Balanced Arrival Rate equations), a core tool in queueing theory that had remained unsolved for 35 years. BAR is considered the "master equation" for determining whether a queueing network reaches equilibrium. This breakthrough demonstrates the immense potential of human-AI collaboration in solving hardcore mathematical problems.
Key Details and AI's Limitations
During the exploration, AI demonstrated strong literature retrieval capabilities by discovering a key paper on pathwise-differentiability, allowing the team to condense a 150-page proof into 10 pages. However, AI also exposed systematic limitations: it tended to provide trivial examples. The crucial Harrison-Reiman class was identified by human researchers using professional intuition, as AI failed to independently discover this key structural property. Furthermore, even when provided with the correct subclass and the key paper, both ChatGPT 5.5 Pro and Claude Opus 4.8 completely failed to complete the proof independently in fresh conversations.
Controversies and Doubts
The team found that the original conjecture was actually mathematically incorrect; it had counterexamples in natural extensions and only held true within the Harrison-Reiman subclass. When attempting to prove it with GPT, the model repeatedly got stuck at a critical juncture. This specific stumbling block prompted the researchers to search for counterexamples, ultimately driving the research toward a genuine breakthrough.
Reactions
The author @2prime_PKU concluded that while AI will not replace mathematicians in the short term, human-AI collaboration has become a superpower for solving decades-old mathematical problems. The author also expressed gratitude to PhD student Youheng Zhu for feedback during the exploration.
2026-07-07 ~ 2026-07-07 · 7 related posts
- [source] 人机协作攻克排队论35年未解猜想 — 2prime_PKU · 2026-07-07
- [source] AI辅助证明猜想:原始命题竟是错的,GPT反复卡壳 — 2prime_PKU · 2026-07-07
- AI未能发现Harrison-Reiman类,人类找到反例 — 2prime_PKU · 2026-07-07
- AI只给平凡例,关键结构类属人类直觉发现 — 2prime_PKU · 2026-07-07
- [source] AI发现关键数学论文,将150页证明压缩至10页 — 2prime_PKU · 2026-07-07
- ChatGPT与Claude均无法独立完成该数学证明 — 2prime_PKU · 2026-07-07
- 观点:AI不会取代数学家,但人机协作是攻克难题的超级武器 — 2prime_PKU · 2026-07-07