OpenAI's GPT-5.6 Pro Disproves 30-Year-Old Mathematical Conjecture
Key Takeaways
- ▸GPT-5.6 Pro autonomously resolved a 30-year-old open conjecture in mathematical optimization with minimal human prompting (four prompts)
- ▸The AI model delivered not just a counterexample but formal proof certificates and machine-executable verification code
- ▸The discovery underscores large language models' growing capability in abstract mathematics and theoretical computer science
Summary
A researcher using OpenAI's GPT-5.6 Pro has discovered a counterexample to the Dinitz–Garg–Goemans conjecture, a mathematical problem that has remained open since the late 1990s. The conjecture concerned whether fractional multicommodity flows could be rounded to unsplittable flows without increasing total cost. The AI model produced the solution in just four prompts, delivering not only the counterexample but also proof certificates, an exhaustive-enumeration verification program, and machine-readable data.
The discovered counterexample is a directed graph on seven nodes carrying three demands, with a fractional solution cost of 58 but a minimum unsplittable routing cost of at least 60 within the allowed capacity constraints. This concrete instance represents a significant resolution to a theoretical computer science problem that has eluded formal resolution in academic literature for decades.
The result was shared informally via X with the public GPT-5.6 Pro transcript as part of a graded registry of AI-generated scientific results. While a third-party researcher has generalized the instance into a parametric family of counterexamples, the work remains unreviewed in formal peer-review channels. The discovery highlights both the potential of large language models in abstract mathematical problem-solving and the emerging challenges around verification, attribution, and formalization of AI-assisted research.
- Informal publication via social media transcript rather than formal academic paper raises accountability and credibility questions for AI-discovered research
Editorial Opinion
This counterexample marks a significant moment: large language models have moved beyond pattern-matching to genuine mathematical discovery. However, the informal publication and lack of independent verification are troubling precedents. For AI-assisted research to gain lasting credibility in academia, it must follow established standards—peer review, published repositories, and independent verification—otherwise we risk celebrating solutions without scrutiny.
