GPT-5.6 solves 50‑year‑old graph theory conjecture in an hour with a 700‑word prompt and 64 sub‑agents

OpenAI announced that the newly released GPT‑5.6 Sol Ultra model completed a proof of the half‑century‑old Cycle Double Cover Conjecture in less than an hour, producing a three‑page PDF that was later praised by Noam Brown at ICML.

What the conjecture asks

The Cycle Double Cover Conjecture states that for any bridgeless graph, one can find a collection of cycles such that every edge appears in exactly two of those cycles. The article illustrates the requirement with a ten‑vertex example, showing how overlapping cycles must cover each edge twice.

Why the problem is difficult

Although every edge of a bridgeless graph belongs to at least one cycle, ensuring that each edge is covered exactly twice forces a global coordination: adding a cycle to fix one edge may cause other edges to be covered three times, and the process must terminate with a perfect double cover.

GPT‑5.6’s proof strategy

Instead of searching for cycles directly, GPT‑5.6 reformulated the task as a finite‑field edge‑labeling problem and solved it with linear algebra. The proof proceeds in four steps:

Reduce any graph to a 3‑regular (cubic) graph, because proving the conjecture for cubic graphs suffices for the general case.

Apply the non‑zero 8‑flow theorem to assign each edge a non‑zero three‑bit label, ensuring that the three incident labels at every vertex sum to zero.

Expand each three‑bit label into a pair of labels so that, at each vertex, a label either does not appear or appears exactly twice.

Translate the global consistency requirement into a system of linear equations; using dual space and parity arguments, GPT‑5.6 shows that the system always has a solution, guaranteeing that the local labels stitch together into a global double‑cover.

Prompt‑engineering techniques revealed

The released prompt (≈700 characters) demonstrates several principles for steering large models on open‑ended tasks:

Do not prescribe a step‑by‑step solution; instead, define the final acceptance criteria.

State definitions, scope, and edge cases up front, and repeat the goal to avoid context drift.

Explicitly list what does **not** count as a valid answer, thereby pruning shortcuts the model might take.

For complex tasks, avoid fixed role assignments; instead, launch up to 64 sub‑agents that explore diverse routes, dynamically reallocate resources, and include dedicated adversarial agents for independent verification.

These guidelines turn a vague research problem into a concrete, auditable contract that the model can execute, verify, and iterate on, illustrating how prompt design can amplify the capabilities of state‑of‑the‑art LLMs for advanced mathematical reasoning.