LLMs Can Tackle High Energy Theoretical Physics, yet Gaps Remain: Part I, Quantum Field Theory

Abstract: Beyond math and code, LLMs are beginning to tackle problems that have traditionally required substantial training for theoretical physicists, but it remains unclear how far this capability extends and where it might break. As a first step toward pushing LLMs in high-energy theoretical physics, we introduce QFT-Bench, a benchmark of 75 graduate- to research-level problems requiring extensive reasoning and derivation across core Quantum Field Theory (QFT) subareas. Each question has a well-defined target and closed-form reference answer. Evaluation combines symbolic verification with LLM-based judge. Across 39 proprietary and open-source models, frontier models already demonstrate substantial capabilities on nontrivial QFT calculations, with the strongest models exceeding 95% accuracy. At the same time, clear gaps remain. Even strong models still make errors in calculations, conventions, and instruction following, while weaker models often fail at the level of physical concepts or derivations. We further monitor models' CoTs and show that current models exhibit nontrivial reasoning unfaithfulness, i.e., they could reach correct answers but the reasoning itself does not justify and logically support them. These results provide a quantitative assessment of current LLM capabilities in QFT and a first step toward broader evaluation in high-energy theoretical physics.

• The paper evaluates 39 reasoning models on QFT-Bench, a benchmark with 75 problems across eight subareas of quantum field theory, finding that top systems achieve up to 100% accuracy but with inconsistencies in derivations.
• Key methodologies of symbolic verification and LLM-based judgment are applied to the models, where symbolic extraction is first compared to reference answers and problematic evaluations are escalated to LLM judges
• The analysis indicates that while the most advanced in the covered year achieve near-perfect accuracy, their derivations often lack faithfulness and contain systematic failures e.g., misuse of conventions, algebra errors, and instruction adherence

The paper presents QFT-Bench, a benchmark intended to measure whether contemporary reasoning-oriented LLMs can perform graduate- and early research-level calculations in quantum field theory. Its central claim is deliberately two-sided: frontier models can already solve a substantial fraction of technically nontrivial QFT problems, but their reliability remains below the standard required for unsupervised theoretical work. The benchmark evaluates both final-answer correctness and the faithfulness of the derivations used to obtain those answers (2610.33134).

Motivation and scope

The authors motivate QFT as a demanding test of scientific reasoning because correctness depends simultaneously on symbolic manipulation, physical interpretation, conventions, approximations, and consistency conditions. Unlike many mathematical and programming tasks, QFT problems are sensitive to signs, normalization choices, gauge conventions, operator definitions, and the precise quantity requested. A model can therefore produce a familiar formula that is mathematically valid in another convention while still answering the benchmark question incorrectly.

QFT-Bench contains 75 manually constructed problems spanning eight subareas:

Subarea	Problems
Classical field theory	7
Quantization of scalar, spinor, and vector fields	8
Quantum electrodynamics	7
Regularization and renormalization	6
Non-Abelian gauge theory and QCD	6
Electroweak and Higgs theory	8
Supersymmetric field theories	18
Conformal field theories	15

The distribution places substantial emphasis on SUSY and CFT, which together account for 33 of the 75 problems. This makes the benchmark broader than evaluations restricted to standard graduate-course material, although it also means that aggregate accuracy is influenced disproportionately by advanced subfields.

The benchmark is designed to test derivation rather than direct retrieval. Problems are adapted from textbooks, lecture notes, and original constructions, then rewritten to make assumptions, conventions, contour orientations, normalization choices, and requested output forms explicit. Each item has a closed-form reference answer and an accompanying reference chain of reasoning. The authors report iterative refinement using preliminary model outputs to reduce ambiguities, including convention-dependent ambiguities in Seiberg–Witten monodromies and CFT oscillator normalizations.

Evaluation methodology

The evaluation covers 39 reasoning models: 22 proprietary systems and 17 open-source systems. All models are given a maximum output budget of approximately 64k tokens, with thinking modes enabled when available. The primary metric is mean binary accuracy over the 75 questions.

Correctness is assessed through a two-stage pipeline. First, the extracted final answer is compared with the reference using symbolic equivalence checks implemented with SymPy. When symbolic verification is inconclusive—for example, because two expressions differ by a convention, gauge choice, or algebraically difficult equivalence—the response is passed to an LLM judge. The judge is instructed to assess the requested quantity under the specified conventions and to treat failure to follow the required answer format as incorrect.

This design is appropriate for QFT, where exact symbolic checking is valuable but incomplete. It also introduces an important dependency: a proprietary LLM judge determines some correctness labels and all faithfulness labels. The reported human-evaluation study provides partial validation, with correctness consistency of 98% for the Opus-based judge and 92% for a second judge, but faithfulness consistency is lower, at 73% and 67%, respectively. Consequently, the faithfulness results should be interpreted as rubric-based measurements rather than direct observations of a model’s latent computational process.

Accuracy and scaling behavior

The strongest systems perform at a level substantially above earlier general-purpose models. Twelve evaluated models exceed 95% mean accuracy, and the highest reported score is 100% for GPT-6.1-Sol. Several other frontier systems achieve 98.7%, including Opus-5.5, GPT-5.6-Sol, Gemini-3.8-Flash, Gemini-3-Flash-Preview, and Muse-Spark-1.3. Sonnet-5.5 and Opus-5 reach 97.3%, while Qwen-3.8-2.4T-A95B reaches 96.0%.

The results exhibit a strong within-family generation effect. Older systems such as GPT-4o, DeepSeek-v3, Llama-3.1, and Qwen3-8B perform substantially worse than later systems in the same broad model families. The gap is especially large for smaller open models: Llama-4-Scout and Llama-3.1-Instruct obtain 41.3%, DeepSeek-v3 obtains 53.3%, and Qwen3-8B obtains 46.7%. These results support a scaling relationship between general reasoning capability and performance on closed-form QFT derivations.

The benchmark also shows that open-source status alone is not a sufficient predictor of weakness. Kimi-K3 reaches 94.7%, Qwen-3.8-2.4T-A95B reaches 96.0%, and DeepSeek-v4.1-Flash reaches 92.0%. Nevertheless, proprietary frontier models dominate the upper end of the ranking, and the paper reports a general performance advantage for proprietary systems.

The paper’s teaser figure summarizes the relationship between model generation and QFT-Bench performance.

Figure 1: Model performance on QFT-Bench as a function of model generation, illustrating the reported scaling trend.

Subfield-level results reveal that high aggregate accuracy does not imply uniform competence. The strongest models often achieve 100% on classical field theory, quantization, non-Abelian gauge theory, SUSY, and CFT, while regularization and renormalization remain comparatively difficult. Even high-performing systems show errors in QED, renormalization, and convention-sensitive calculations. For example, Opus-5.5 achieves 100% in most categories but scores 85.7% in QED and 100% in renormalization; Sonnet-5.5 scores 83.3% in renormalization; and GPT-5.6-Luna scores 66.7% in renormalization.

The implication is that aggregate scores near 100% conceal structured residual failure. A model that solves 74 of 75 benchmark problems can still fail on a calculation involving a single field-counting identity, a noncanonical normalization, or a convention fixed explicitly by the prompt. Such errors are particularly consequential in research workflows because they can survive superficial plausibility checks.

Error modes

The authors classify incorrect responses into conceptual, calculation, strong-prior, and instruction-following errors. The distribution changes systematically with model capability. Among the lower-performing models, conceptual errors constitute 47.2% of failures, whereas they account for 23.4% among the top half. This indicates that scaling reduces failures involving incorrect physical principles or inappropriate methods.

For stronger models, the remaining errors shift toward more localized but still consequential failures:

Failure category	Top half	Other models
Conceptual	23.4%	47.2%
Calculation	21.6%	18.6%
Strong-prior	23.4%	13.4%
Instruction following	31.5%	20.8%

The relative increase in instruction-following and strong-prior errors among capable models is significant. These systems often understand the physics but answer a nearby question, apply a standard normalization instead of the stipulated one, or fail to provide the required final expression.

The conceptual-error example concerns the superficial degree of divergence in a theory with a massive vector and a Dirac fermion. The model uses $I_A+E_A=V$ rather than the correct relation $2I_A+E_A=V$, thereby counting internal vector lines incorrectly. The resulting formula is not a minor algebraic deviation; it reflects a failure to track diagrammatic incidence relations.

The calculation-error example involves integrating out the radial mode in the Abelian Higgs model. The model identifies the appropriate effective-theory strategy but assigns the wrong Higgs mass because it overlooks the noncanonical normalization of the radial fluctuation. The resulting error propagates into the effective quartic vector coupling.

The strong-prior example is particularly diagnostic. In a non-Abelian gauge theory with a specified gauge transformation and charge normalization, Kimi-K3 first derives the correct Noether charge but then rescales it to a more familiar conventional electric-charge normalization. The response is physically recognizable yet incorrect under the problem’s definitions. This demonstrates that factual familiarity can interfere with instruction-conditioned symbolic reasoning.

Instruction-following failures include responses that derive the correct result but do not place it in the required boxed form. Although such grading may appear strict, the benchmark explicitly treats output format as part of the task specification. The authors thereby test whether a model can satisfy an interface contract in addition to deriving the result.

Reasoning faithfulness

The paper’s principal methodological extension beyond ordinary benchmark accuracy is its analysis of reasoning faithfulness. A response is considered unfaithful when its final answer is correct but the displayed derivation does not logically support it. The authors exclude responses with insufficient derivations and define the unfaithfulness rate for model $m$ as the fraction of correct answers whose reasoning is judged inconsistent with the final answer.

Across all 39 models, the mean unfaithfulness rate is 7.2%. This is non-negligible given that the denominator consists only of correct final answers. The paper reports substantially higher rates for Llama models, approximately 20–40%, whereas Qwen models generally display lower rates. Muse-Spark, GLM, and Kimi are also reported to produce comparatively faithful traces.

The figure below presents the claimed relationship between capability and reasoning faithfulness.

Figure 2: Reasoning faithfulness increases approximately linearly with model capability within model families, although substantial family-level variation remains.

The authors report an approximately linear increase in faithfulness with capability, particularly within individual model families. Stronger systems approach near-perfect faithfulness on the evaluated sample. This result is important but should not be conflated with proof that stronger models possess transparent internal reasoning. The measurement evaluates consistency between the emitted derivation and final answer using an LLM judge; it does not establish that the displayed chain of thought is causally responsible for the answer.

The case studies illustrate the distinction. In one Virasoro-algebra problem, Llama-4-Maverick computes an incorrect Gram-matrix entry, later uses the correct entry without acknowledging the discrepancy, and nevertheless obtains the reference determinant. In another CFT example, Opus-4.6 writes an intermediate expression that simplifies to a central charge independent of the conformal weight, then presents the correct weight-dependent result without repairing the contradiction. In both cases, final-answer accuracy would conceal a derivational inconsistency.

This finding directly affects the interpretation of high benchmark scores. A correct boxed answer is not sufficient evidence that the model performed the derivation represented in its response. For applications involving proof auditing, analytic continuation, renormalization, or multi-stage symbolic computation, faithfulness must be evaluated separately from answer accuracy.

Strengths of the benchmark

QFT-Bench improves on broad scientific benchmarks in three respects. First, it provides substantially more QFT coverage than general expert-level evaluations. Second, its questions target derivations and calculations rather than primarily conceptual multiple-choice recognition. Third, it combines symbolic verification with expert-oriented judgment of convention and physical equivalence.

The explicit handling of conventions is especially valuable. QFT contains many equivalent but nonidentical representations, and benchmark construction that ignores this fact can misclassify valid answers. The authors attempt to reduce ambiguity by fixing metric signatures, basis orderings, analytic-continuation paths, oscillator conventions, gauge transformations, and requested output quantities.

The benchmark also covers a useful range of task structures. Its examples include scattering amplitudes, propagator inversion, beta functions, Chern–Simons currents, decay widths, Seiberg–Witten monodromies, and conformal-block descendants. This diversity reduces the risk that performance can be explained solely by memorization of one calculation template, although it does not eliminate that possibility.

Limitations and open questions

The benchmark remains limited in several ways. Its 75 questions are manually authored and all have closed-form reference answers. This is necessary for reliable evaluation but excludes discovery-oriented tasks, numerical calculations without compact analytic forms, proof construction under formal verification, and problems whose correctness depends on physical judgment or comparison with data. The benchmark therefore measures advanced derivation competence, not theoretical-physics research ability in general.

The evaluation also uses a single inference configuration, default sampling parameters, and one maximum token budget. Accuracy@k, tool-assisted reasoning, symbolic-computation tools, retrieval, self-consistency, and agentic decomposition are not systematically compared. The reported scores therefore characterize a particular prompting and API regime rather than an intrinsic model capability.

The LLM-as-a-judge stage creates a second-order evaluation problem. Although the authors report human agreement for correctness, agreement for faithfulness is considerably lower. The use of an LLM judge is also especially consequential for subtle convention-dependent questions, where the judge may inherit the same prior-driven errors that the benchmark seeks to detect.

Finally, the paper does not establish how performance transfers to novel problems outside the benchmark’s construction distribution. Because questions are deliberately designed to have closed-form answers and are refined through model trials, the possibility of benchmark-specific adaptation remains open. A central unresolved question is whether the observed scaling of accuracy and faithfulness persists for unseen QFT problems requiring longer derivations, nonstandard conventions, numerical validation, or interaction with computer-algebra systems.

Conclusion

QFT-Bench provides evidence that frontier LLMs can solve many graduate- and early research-level QFT calculations, with the best evaluated systems exceeding 95% accuracy and one reported model reaching 100%. The results also show that residual errors are not confined to weak conceptual understanding: strong models fail through algebraic slips, convention substitution, prior-driven normalization changes, and instruction-following violations.

The paper’s most important qualification is that final-answer accuracy and reasoning faithfulness are distinct. A correct result can coexist with an internally inconsistent derivation, with an average unfaithfulness rate of 7.2% across the evaluated models. QFT-Bench therefore supports a measured conclusion: current reasoning models have substantial competence in closed-form QFT problem solving, but reliable deployment requires verification of both the result and the derivation that purports to justify it (2610.33134).