VibeMathedMath problems solved by AI

Complexity of Terminal-Only Manhattan Prim-Dijkstra Routing

Prim-Dijkstra routing interpolates between a minimum spanning tree and a shortest-path tree, and has been used and improved in VLSI physical design since the early 1990s, but the complexity of the terminal-only Manhattan decision problem was never settled. It is weakly NP-complete. A continuous cost-radius tradeoff with a balanced (2,2)(2,2) guarantee accompanies the classification.

Result
Proved
Status
Resolved
AI contribution
AI-discovered
Method
Argument
Field
Computational complexity
Posed by
Alpert, Hu, Huang, Kahng, Karger
Year posed
1995
Years open
31y
Solved
2026-07-18
Model
GPT-5.6 Sol in Codex
Vendor
OpenAI
Collaborators
Keren Zhu
Verification
Unreviewed
Publication
Preprint
Significance
10 / 100
Disclosed cost
Wikipedia
No dedicated article

What the AI did

The paper is a designed experiment rather than an incidental use. One model was run under five deliberately incompatible research conditions that differed in premise and information boundary: construct and adversarially audit an NP-completeness proof from the problem definition alone; pursue a polynomial-time exact algorithm blind, without literature or the option of retreating to a hardness claim; pursue the same informed; assume hardness and seek a bicriteria guarantee; and synthesize the record into a solver against an evaluator frozen beforehand. The hardness track produced the NP-completeness proof. The two exact tracks converged on the same architecture and produced counterexamples rather than a proof, which the author treats as a control.

Verification

Single-author arXiv preprint with a full NP-completeness proof in Appendix A, a released research record, and a partial Lean snapshot whose theorem boundary the paper states explicitly. Not yet peer-reviewed.

Sources

arXiv:2607.17005 - Provably Good Prim-Dijkstra Revisited: New Theory and a Practical Algorithm for a Classical VLSI Routing Problem with LLMs

Discussion