LLMs fully solve Knuth's "Claude Cycles" problem
Donald Knuth's 3D graph theory challenge, once thought to require manual human insight, has been fully resolved through a multi-model collaboration between Claude Opus 4.6 and GPT-5.3. The breakthrough marks a significant shift in Knuth's stance on generative AI's reasoning capabilities, moving from skepticism to "shock" as the models derived a general construction for Hamiltonian cycle decomposition.
Knuth’s "Shock! Shock!" reaction signals a major validation of LLM reasoning in pure mathematics.
- –Claude Opus 4.6 moved beyond simulated annealing to identify a general coordinate-based rule for odd-numbered graphs.
- –Human-guided explorations (31 turns) were critical in nudging the model toward "pure math" reasoning.
- –The subsequent resolution of even-numbered cases by GPT-5.3 completes the decomposition proof for all graph sizes.
- –Formal verification via the Lean proof assistant ensures the LLM-generated constructions are mathematically sound.
- –This collaboration suggests a future where mathematicians use LLMs as high-level symbolic reasoning partners rather than just coding assistants.
DISCOVERED
73d ago
2026-03-28
PUBLISHED
73d ago
2026-03-28
RELEVANCE
AUTHOR
mean_mistreater