Theory-Guided ChatGPT Discovery of Goemans Unsplittable-Flow Counterexamples Across 93 Rooted Arborescences
Summary:This arXiv v1 paper does not equate ChatGPT's chance hit with autonomous theorem proving. Instead, it develops a verifiable forward-construction pipeline for Goemans's unsplittable-flow conjecture: begin with a rooted arborescence on which failure is impossible, add terminal-entering shortcut arcs, convert load constraints into a finite set of admissible route choices, and use their convex hull—the conflict polytope—to characterize feasible demands, fractional flows, and costs. The pipeline sel
The paper's central contribution is not another announcement that ChatGPT disproved a conjecture. It converts the seven-vertex counterexample that Dmitry Rybin found with ChatGPT into an interpretable, extensible, and independently verifiable construction framework. The model searches within theoretical constraints, while the mathematical construction determines whether a candidate is genuinely a counterexample.
Main result: from a single discovery to a transferable method
The paper separates its contribution into three results. First, the pipeline starts from a rooted arborescence on which failure is impossible and adds terminal-entering shortcut arcs. It can select Rybin's seven-vertex topology and derive the exact conditions that counterexamples supported by that topology must satisfy. Second, when ChatGPT is instructed to vary the initial arborescence and shortcut placement, it produces counterexamples on genuinely different topologies, showing that the method is not tied to reproducing one example. Third, the experiments obtain valid constructions starting from 93 other rooted arborescence shapes, demonstrating transfer across different initial structures.
The authors therefore do not characterize the Rybin–ChatGPT interaction as fully autonomous mathematical discovery. They place it between unconstrained enumeration of topologies, demands, and costs on one side, and a complete theorem-guided construction process on the other. ChatGPT participated in the discrete structural search, but the counterexample conditions were not established by unconstrained natural-language reasoning.
Mechanism: from an arborescence to a conflict polytope
Goemans's conjecture concerns unsplittable flow in a single-source network. A fractional flow may split each demand over several paths; an unsplittable solution requires every demand to select one path in full. Given a fractional flow x* and the largest demand d_max, an unsplittable assignment y can be described by:
C(y) = Σ_v d_v c_vrL_a(y) = Σ_v d_v Σ_(r contains a) y_vr
The conjecture asks for an unsplittable assignment satisfying C(y) ≤ C(x*) and L_a(y) ≤ L_a(x*) + d_max for every arc. A counterexample is a feasible fractional flow x* for which every unsplittable assignment violates at least one of these cost or load bounds.
Construction begins with a rooted arborescence. Every terminal has a unique root-to-terminal path, so the demands have no meaningful routing freedom and failure is impossible. Terminal-entering shortcut arcs then introduce route alternatives, while load constraints eliminate inadmissible combinations.
For a fixed topology, the number of candidate paths for each demand is finite. The paper encodes feasible route choices as vectors and takes their convex hull, producing the conflict polytope. This polytope describes which route choices can coexist and compresses the original continuous search over demands, fractional flows, and costs into finite route configurations and linear inequalities. A proposed counterexample can therefore be checked by enumerating admissible unsplittable assignments and testing their cost and load bounds, without trusting the model's textual claim that it has found a proof.
Why ChatGPT can outperform blind enumeration
The original problem combines graph structure, demand allocation, continuous flow values, and cost functions. Unconstrained search must handle a large discrete topology space and solve a continuous optimization problem within every topology. The pipeline fixes the initial tree and rules for placing shortcuts, then uses load constraints to generate a finite set of routes. ChatGPT only has to search structured options rather than invent an entire proof from an empty prompt.
The result across 93 rooted arborescences is more informative than a single counterexample because it shows that the model can move between initial structures while retaining the same certificate conditions. It does not show that ChatGPT has acquired general combinatorial-optimization theory, however, because the search space, parameter ranges, and verification rules are supplied by the pipeline.
Engineering trade-offs and evaluation limits
The method carries a substantial combinatorial computation cost. A conflict polytope can develop many vertices and facets, while the number of candidate paths may grow rapidly with the number of vertices and shortcuts. Floating-point checks are risky when a load boundary is nearly tight: rounding error could turn an invalid candidate into an apparent counterexample. Reproduction should prefer exact rational arithmetic or explicitly report numerical tolerances and their sensitivity.
The evidence also has several limitations. The number 93 counts rooted arborescence shapes from which valid constructions began; it is not a success rate, because the abstract provides neither the total number of attempts nor counts of failures and invalid candidates. The study does not report matched-budget baselines for human search, random search, ChatGPT without the theoretical pipeline, or other heuristics. It is therefore impossible to separate the contribution of the model from that of the human-designed scaffold. The abstract also omits the ChatGPT version, prompts, sampling settings, token consumption, call cost, and candidate latency, all of which affect reproducibility and are subject to model-version drift.
The defensible conclusion is narrower than autonomous theorem proving: an LLM constrained by mathematical structure can serve as an effective discrete candidate generator and, when paired with an independent verifier, discover transferable counterexamples.
Reproduction and deployment guidance
A reproducible implementation should separate four components: a symbolic constructor that generates trees and shortcuts, an LLM that proposes candidates, an exact solver that enumerates or verifies unsplittable assignments, and an artifact store that records topologies, flows, costs, proofs, and validation results. LLM outputs should enter only a candidate queue, never the verified set directly.
A minimum reproducibility record should include the ChatGPT model version, full prompts, temperature and random seeds, every attempt and failure, exact-arithmetic settings, conflict-polytope size, and verification time per candidate. Evaluation should compare human search, random search, LLM search without the theoretical pipeline, and the complete hybrid pipeline under identical candidate budgets and the same verifier.
From an engineering perspective, the approach does not appear to require a specialized trained model; its main costs are LLM calls, combinatorial solving, and exact validation. To establish broader value for AI-assisted science, the next step is to transfer the same construct–search–verify architecture to other combinatorial conjectures and test zero-shot generalization to unseen problem families.
— END —Source: ArXiv Machine Learning (cs.LG) (2026-11-19)
Tags: #Goemans Conjecture #Unsplittable Flow #ChatGPT #AI-Assisted Mathematics #Conflict Polytope
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments