Bethe Ansatz with a Large Language Model

Abstract: We explore the capability of a LLM to perform specific computations in mathematical physics: the task is to compute the coordinate Bethe Ansatz solution of selected integrable spin chain models. We select three integrable Hamiltonians for which the solutions were unpublished; two of the Hamiltonians are actually new. We observed that the LLM semi-autonomously solved the task in all cases, with a few mistakes along the way. These were corrected after the human researchers spotted them. The results of the LLM were checked against exact diagonalization (performed by separate programs), and the derivations were also checked by the authors. The Bethe Ansatz solutions are interesting in themselves. Our second model manifestly breaks left-right invariance, but it is PT-symmetric, therefore its solution could be interesting for applications in Generalized Hydrodynamics. And our third model is solved by a special form of the nested Bethe Ansatz, where the model is interacting, but the nesting level has a free fermionic structure lacking $U(1)$-invariance. This structure appears to be unique and it was found by the LLM. We used ChatGPT 5.2 Pro and 5.4 Pro by OpenAI.

• The paper demonstrates that state-of-the-art LLMs can autonomously derive coordinate Bethe Ansatz solutions for complex integrable spin chain models.
• The methodology integrates numerical diagonalization and analytical cross-checks to validate solutions, revealing novel free-fermion and nested R-matrix structures in Model Y3.
• The work underscores LLMs’ potential to accelerate research in mathematical physics and to set new benchmarks for automated problem solving in integrability.

Bethe Ansatz Problem Solving with LLMs: Computational Advances in Integrable Spin Chain Models

Overview

The manuscript "Bethe Ansatz with a LLM" (2603.29932) investigates the capacity of state-of-the-art LLMs (specifically ChatGPT 5.2/5.4 Pro) to solve research-level mathematical physics tasks, focusing on the explicit coordinate Bethe Ansatz for quantum integrable spin chain models. The authors select three models—two previously unpublished, one recently introduced with unresolved Bethe Ansatz structure—and demonstrate LLM-driven semi-autonomous derivations of their Bethe Ansatz solutions, subsequently verifying the results numerically via exact diagonalization and analytically cross-checking key steps. Notably, the LLM discovers novel structural features in model Y3, related to a nested free fermion 8-vertex $R$-matrix lacking $U(1)$ symmetry, which appears to be unique in the literature.

Motivation and Context

There has been accelerating interest in using machine learning, and recently LLMs, for assisting scientific discovery, particularly in mathematics and theoretical physics (Carleo et al., 2019, Abouzaid et al., 5 Feb 2026). Though LLMs have demonstrated proficiency in solving advanced mathematics problems and competition-level tasks, their systematic utility for exact computations in mathematical physics—especially quantum integrability—remains largely unexplored due to the high technical requirements and absence from standard benchmarks (Glazer et al., 2024, Chung et al., 19 Feb 2025). The present work addresses this gap, leveraging the intrinsic proximity of integrable models to pure mathematics and algebra to probe the efficacy of LLMs on research-level problems.

Integrable Spin Chain Models

Three models are considered:

• Model Y1: A three-site Hamiltonian related to the twisted XXZ chain. Its Bethe Ansatz solution is structurally simple but nontrivially related to known models. The LLM reconstructs the solution absent explicit prompts concerning XXZ connections, showing proficiency in nontrivial problem exploration.
• Model Y2: A newly discovered PT-symmetric model with explicit left-right reflection symmetry breaking and two types of excitations on even and odd sublattices. Its Bethe Ansatz requires a nested construction, as the LLM correctly determines after initially attempting a scalar Ansatz. The model has physical implications for generalized hydrodynamics due to its symmetry structure.
• Model Y3: An $SU(2)$-symmetric four-site model with unique nested Bethe Ansatz structure. Unlike conventional cases, its nesting level lacks $U(1)$ symmetry, exhibiting instead a free fermion structure within an 8-vertex $R$-matrix. The LLM autonomously discovers this, identifying factorizable eigenvectors and the explicit form of the nested transfer matrix.

Bethe Ansatz and Analytical Computations

Bethe Ansatz derivations rely on constructing eigenstates in coordinate space, determining energy additivity, and factorizing many-body scattering via Yang-Baxter-compliant $R$-matrices. For scalar models, exchange amplitudes are governed by one-dimensional $S(z_j,z_k)$ factors; for models with internal degrees of freedom (flavor, sublattice, sign), the nested Ansatz introduces higher-dimensional $R$-matrices operating on permutation and internal tensor spaces. The LLM systematically constructs these frameworks, identifies physical and technical constraints—such as exceptional manifolds, nontrivial block structures, and regularity relations—and produces the correct Bethe equations in each context.

Numerical Verification

All solutions are validated numerically by solving the Bethe equations and comparing predicted spectra against exact diagonalization for small system sizes. This ensures correctness in physical predictions and discards hallucinated or accidental analytical matches. The LLM autonomously generates and runs numerical scripts, though human cross-checking is advised due to observed hallucinations in intermediate outputs when analytical errors propagate.

Error Analysis and Verification

Analysis of LLM outputs reveals several typical mistake modalities:

1. Incorrect generalization: Extending two-particle Bethe equations to multi-particle problems without verifying full factorization, corrected upon numerical inconsistency.
2. Minor misidentification: Incorrect ordering of $S$-matrix arguments, use of checked versus standard $R$-matrix, sign errors, or inconsistent conventions—analogous to common human errors.
3. Intermediate inconsistencies: Errors in transient formulas that do not affect final results.

The necessity for independent verification is emphasized, suggesting automated proof-checking frameworks for integrability and mathematical physics analogous to those developed for theorem proving (Achim et al., 1 Oct 2025, Li et al., 30 Oct 2025).

Key Results

• Model Y1: LLM reconstructs Bethe Ansatz and correct exchange rules, with numerical spectra matching exact diagonalization. The solution’s simplicity belies its hidden relation to XXZ-twisted models.
• Model Y2: The LLM identifies two excitation types, correctly derives the nested Bethe Ansatz, and provides explicit $U(1)$0-matrix block forms. The $U(1)$1-matrix is shown to coincide with the trigonometric six-vertex model, and the nested Bethe equations are validated numerically.
• Model Y3: The LLM discovers the intricate nested structure—a parity-preserving free-fermion 8-vertex $U(1)$2-matrix—constructs gauged bases, identifies explicit product branches, and details the full spectrum via quadratic fermionic transfer matrix operators. Numerical computation of the generalized $U(1)$3 eigenvalue problem and identification of physical single-particle branches are demonstrated.

Implications and Future Directions

These results demonstrate that LLMs, with expert prompting and targeted error correction, can perform advanced research-level computations in mathematical physics, including explicit derivation and verification of Bethe Ansatz solutions for previously unsolved integrable models. The practical implications include:

• Acceleration of exploratory and technical computation: Expert researchers may leverage LLMs for rapid hypothesis testing and derivation of complex algebraic structures, reducing time investment on routine but nontrivial algebra and numerical checking.
• Discovery of novel structures: LLM autonomy is shown to facilitate the investigation of models with unique or previously unnoticed algebraic features, such as the free fermion nested structure in model Y3.
• Benchmarking and validation frameworks: The study motivates the inclusion of integrable models in AI-assisted mathematical physics benchmarks and the development of automated verification protocols to handle analytical and numerical results, especially where immediate diagonalization checks are infeasible.
• Progress on open problems: Systematic deployment on increasingly difficult or open problems in integrability could yield new results previously intractable for human-expert-only efforts.

Theoretical implications concern the intersection of computational algebra, operator theory, and automated symbolic reasoning in quantum many-body physics, suggesting future directions in algorithmic integrability and AI-augmented analytic computation.

Conclusion

This work provides evidence that modern LLMs can solve coordinate Bethe Ansatz problems for both familiar and novel integrable spin chain models, including nested and structurally atypical cases. With human supervision and verification, the solutions are correct and yield publishable results, suggesting LLMs are poised to become integral computational partners in advanced mathematical physics research. Rigorous error checking and further integration with formal proof tools will be essential for widespread adoption in domains where analytical and numerical subtleties predominate.