VibeMathedMath problems solved by AI

The Papamanthou-Tollis Conjecture on Parameterized st-Orientations

On the basis of experiments up to 5000 nodes, Papamanthou and Tollis conjectured a relation between the longest paths produced by their MaxSTN and MinSTN algorithms for stst-orientations of biconnected graphs. A counterexample refutes it.

Result
Disproved
Status
Resolved
AI contribution
AI co-developed
Method
Construction
Field
Graph algorithms
Posed by
Charalampos Papamanthou, Ioannis G. Tollis
Year posed
2008
Years open
18y
Solved
2026-06-12
Model
Claude Sonnet 4.6
Vendor
Anthropic
Collaborators
Charalampos Papamanthou
Verification
Site-confirmed
Publication
Preprint
Significance
10 / 100
Disclosed cost
Wikipedia
No dedicated article

What the AI did

The paper devotes a section to this and it is worth reading in full, because the headline line is the least of it. Claude Sonnet 4.6 read the earlier paper and wrote the Python st-orientation code quickly, and the authors say no code was written by a human. But the loop mattered: the authors caught and fed back three specific bugs, cut vertices being treated as eligible, the sink block not being excluded, and the timestamp overwrite rule implemented incorrectly. The model ran exhaustive tests for n6n \le 6, declined n=9n = 9 as computationally infeasible, then later suggested a resource that would have made it possible. It also produced counterexamples that external validation confirmed were wrong, and drew conclusions the authors state they did not verify, including an exhaustive-search claim over the biconnected graphs on five and six vertices. The counterexample that survived is the model's; so is a quantity of discarded work. The author is one of the two who posed the conjecture, so this is someone refuting their own with a model.

Verification

The refutation is an explicit graph, so it reduces to running the two named algorithms on it, and that part stands on its own. The surrounding computational claims deserve less weight: in the same section the authors record that the model produced counterexamples later confirmed wrong, and that they did not verify its claim to have exhaustively searched the biconnected graphs on five and six vertices. arXiv note, not peer-reviewed.

Source

arXiv:2606.14539 - On a Conjecture for Parameterized st-Orientations

Changelog2 changes
  • Rasmus Lindahlchanged What the AI did from quoted only the 'no human code' line to full disclosure: the debugging loop, the wrong counterexamples, the unverified claims
  • Rasmus Lindahlchanged AI contribution from ai-discovered to ai-co-developed

Discussion