GPT-5.6一小时破解50年数学猜想，700词Prompt驾驭64个子Agent | GPT-5.6 Cracks 50-Year Math Conjecture in One Hour, 700-Word Prompt Orchestrates 64 Sub-Agents
GPT-5.6一小时破解50年数学猜想，700词Prompt驾驭64个子Agent
一道困扰数学界半个世纪的图论猜想，被刚上线一天的GPT-5.6 Sol Ultra用不到一小时完成证明。
事件回顾
2026年7月10日，OpenAI正式向全球用户全面开放GPT-5.6全系列模型。不到24小时后，OpenAI研究员Ethan Knight宣布：GPT-5.6 Sol Ultra成功完成了循环双覆盖猜想（Cycle Double Cover Conjecture）的证明。
这道题自1970年代提出以来，至今无人能解。GPT-5.6 Sol Ultra不仅给出了证明，还写了一份完整的三页PDF论文。
技术细节
证明过程中，研究员将GPT-5.6的Sol模型升级到Ultra档位。Sol随后自主生成了64个子Agent，将原本可能需要一整天的推理压缩到了一小时以内。
关键在于那份仅有700词的Prompt。它的核心策略是：
- 不替模型规定解法——只把验收标准钉死
- 定义、范围、边界情况提前一次说清
- 适当重复目标，防止上下文漂移
- 明确要求”解决问题”而非”提交状态报告”
正在韩国参加ICML的Noam Brown（o1核心贡献者）第一时间表态：这次不靠内部特供模型，纯靠公开可用的GPT-5.6 Sol Ultra就把活儿干了。
社区反应
证明发布后，Hacker News上的讨论迅速分成两派：
- 乐观派：认为这是AI数学推理能力的里程碑，多Agent并行将极大加速科研
- 审慎派：指出证明异常简短，要么是巧妙的捷径，要么存在尚未发现的错误
专家验证仍在进行中。但无论结果如何，一个事实已经确立：前沿模型在数学领域的天花板还在快速拉高。
对开发者的启示
这次最值得关注的不是跑分，而是那份公开的Prompt。它揭示了一个重要模式：
驾驭顶级模型的关键不是”教它怎么做”，而是”把目标说清楚，把边界画明确，然后让它自己找路”。
这和当前Agent开发的最佳实践完全一致——用约束代替指令，用验收标准代替步骤描述。
GPT-5.6 Cracks 50-Year Math Conjecture in One Hour, 700-Word Prompt Orchestrates 64 Sub-Agents
A graph theory conjecture that stumped mathematicians for half a century was proven by GPT-5.6 Sol Ultra in under an hour, just one day after the model’s global release.
What Happened
On July 10, 2026, OpenAI made GPT-5.6 fully available worldwide. Less than 24 hours later, OpenAI researcher Ethan Knight announced that GPT-5.6 Sol Ultra had successfully proven the Cycle Double Cover Conjecture — a graph theory problem open since the 1970s.
The model didn’t just produce a proof; it wrote a complete three-page PDF paper.
Technical Details
The researcher upgraded GPT-5.6’s Sol model to the Ultra tier. Sol then autonomously spawned 64 sub-agents, compressing what could have taken a full day into under an hour.
The key was a remarkably concise 700-word prompt with these strategies:
- Don’t prescribe a solution method — lock down only the acceptance criteria
- Define scope, boundaries, and edge cases upfront
- Restate the goal periodically to prevent context drift
- Explicitly demand “solve the problem” not “file a status report”
Noam Brown (core o1 contributor), attending ICML in Korea, confirmed: this wasn’t done with an internal special model — publicly available GPT-5.6 Sol Ultra did all the work.
Community Reaction
Hacker News discussions quickly split:
- Optimists: A milestone in AI mathematical reasoning; multi-agent parallelism will massively accelerate research
- Skeptics: The proof is unusually short — either a clever shortcut experts missed, or a subtle error waiting to be found
Expert verification is still pending. But one fact is already established: the ceiling for frontier models in mathematics is still rising rapidly.
Takeaway for Developers
The most noteworthy aspect isn’t the benchmark score — it’s the published prompt. It reveals a crucial pattern:
The key to steering top-tier models isn’t teaching them how to do it, but stating the goal clearly, drawing the boundaries, and letting them find their own path.
This aligns perfectly with current best practices in Agent development: use constraints instead of instructions, use acceptance criteria instead of step descriptions.
来源：量子位、ai0.news、Hacker News