[Paper Review] Early Evidence of Vibe-Proving with Consumer LLMs: A Case Study on Spectral Region Characterization with ChatGPT-5.2 (Thinking)

Brecht Verbeken, Brando Vagenende+3 more|arXiv (Cornell University)|Feb 21, 2026

Artificial Intelligence in Healthcare and Education0 citations

TL;DR

The paper presents a structured, auditable case study where a consumer LLM (ChatGPT-5.2 Thinking) collaborates with humans to prove a spectral-region characterization for a 4-cycle row-stochastic matrix family, highlighting workflow, verification bottlenecks, and the potential of human-in-the-loop theorem proving.

ABSTRACT

Large Language Models (LLMs) are increasingly used as scientific copilots, but evidence on their role in research-level mathematics remains limited, especially for workflows accessible to individual researchers. We present early evidence for vibe-proving with a consumer subscription LLM through an auditable case study that resolves Conjecture 20 of Ran and Teng (2024) on the exact nonreal spectral region of a 4-cycle row-stochastic nonnegative matrix family. We analyze seven shareable ChatGPT-5.2 (Thinking) threads and four versioned proof drafts, documenting an iterative pipeline of generate, referee, and repair. The model is most useful for high-level proof search, while human experts remain essential for correctness-critical closure. The final theorem provides necessary and sufficient region conditions and explicit boundary attainment constructions. Beyond the mathematical result, we contribute a process-level characterization of where LLM assistance materially helps and where verification bottlenecks persist, with implications for evaluation of AI-assisted research workflows and for designing human-in-the-loop theorem proving systems.

Motivation & Objective

• Demonstrate that consumer LLMs can contribute to mathematically substantive proof development with explicit human verification.
• Provide an auditable artifact set (transcripts and proof drafts) for end-to-end inspection of AI-assisted theorem proving.
• Characterize the division of labor between LLM-generated structure and human correctness-critical verification.

Proposed method

• Adopt a generate–referee–repair workflow with seven ChatGPT-5.2 (Thinking) threads and four versioned drafts.
• Apply a Dmitriev–Dynkin trigonometric reduction to the 4-cycle row-stochastic matrix family.
• Use a target theorem statement and boundary information as scaffolding for LLM-proposed proof strategies.
• Incorporate explicit correctness obligations (quadrant handling, endpoint admissibility, and algebraic expansions) and patch-search against independent sessions.
• Leverage Lamport-style claim decomposition to organize dependencies and verification steps.

Experimental results

Research questions

• RQ1 Can a consumer-access LLM contribute to a research-level mathematical proof with auditable, end-to-end workflow?
• RQ2 What is the division of labor between AI-generated structure and human verification in vibe-proving for spectral-region problems?
• RQ3 What verification bottlenecks arise, and how can workflow practices mitigate them?
• RQ4 Is it feasible to produce a complete, checkable characterization of non-real eigenvalues for a 4-cycle row-stochastic matrix family using LLM-assisted proof?
• RQ5 How do transcript-based artefacts and versioning support auditability in AI-assisted mathematics?

Key findings

• A stable generate–referee–repair loop yields a complete and checkable proof for the conjecture’s spectral-region characterization.
• LLMs are strongest at proposing global structure and algebraic shortcuts, while humans handle correctness-critical verification and long expansions.
• The verification bottleneck concentrates on a few key obligations (e.g., tight-regime inequalities and factorization steps) that are amenable to mechanized checking.
• Parallel patch search, bounded referee passes, and version-controlled rewriting reduce regressions and improve auditability.
• An auditable workflow with explicit artefacts can illuminate where AI-assisted workflows help and where human verification remains essential.