arXiv subjects

Ming-xing Luo

Publications and source records attributed to Ming-xing Luo.

At least 19 recordsLinked to original sources

arXiv2026

PRBench: End-to-end Paper Reproduction in Physics Research

AI agents powered by large language models exhibit strong reasoning and problem-solving capabilities, enabling them to assist scientific research tasks such as formula derivation and code generation. However, whether these agents can reliably perform end-to-end reproduction from real scientific papers remains an open question. We introduce PRBench, a benchmark of 30 expert-curated tasks spanning 11 subfields of physics. Each task requires an agent to comprehend the methodology of a published paper, implement the corresponding algorithms from scratch, and produce quantitative results matching the original publication. Agents are provided only with the task instruction and paper content, and operate in a sandboxed execution environment. All tasks are contributed by domain experts from over 20 research groups at the School of Physics, Peking University, each grounded in a real published paper and validated through end-to-end reproduction with verified ground-truth results and detailed scoring rubrics. Using an agentified assessment pipeline, we evaluate a set of coding agents on PRBench and analyze their capabilities across key dimensions of scientific reasoning and execution. The best-performing agent, OpenAI Codex powered by GPT-5.3-Codex, achieves a mean overall score of 34%. All agents exhibit a zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. We further identify systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data. Overall, PRBench provides a rigorous benchmark for evaluating progress toward autonomous scientific research.

cs.CL↗

arXiv2026

An End-to-end Architecture for Collider Physics and Beyond

We present, to our knowledge, the first language-driven agent system capable of executing end-to-end collider phenomenology tasks, instantiated within a decoupled, domain-agnostic architecture for autonomous High-Energy Physics phenomenology. Guided only by natural-language prompts supplemented with standard physics notation, ColliderAgent carries out workflows from a theoretical Lagrangian to final phenomenological outputs without relying on package-specific code. In this framework, a hierarchical multi-agent reasoning layer is coupled to Magnus, a unified execution backend for phenomenological calculations and simulation toolchains. We validate the system on representative literature reproductions spanning leptoquark and axion-like-particle scenarios, higher-dimensional effective operators, parton-level and detector-level analyses, and large-scale parameter scans leading to exclusion limits. These results point to a route toward more automated, scalable, and reproducible research in collider physics, cosmology, and physics more broadly.

hep-ph↗

arXiv2025

Detailed balance in large language model-driven agents

Large language model (LLM)-driven agents are emerging as a powerful new paradigm for solving complex problems. Despite the empirical success of these practices, a theoretical framework to understand and unify their macroscopic dynamics remains lacking. This Letter proposes a method based on the least action principle to estimate the underlying generative directionality of LLMs embedded within agents. By experimentally measuring the transition probabilities between LLM-generated states, we statistically discover a detailed balance in LLM-generated transitions, indicating that LLM generation may not be achieved by generally learning rule sets and strategies, but rather by implicitly learning a class of underlying potential functions that may transcend different LLM architectures and prompt templates. To our knowledge, this is the first discovery of a macroscopic physical law in LLM generative dynamics that does not depend on specific model details. This work is an attempt to establish a macroscopic dynamics theory of complex AI systems, aiming to elevate the study of AI agents from a collection of engineering practices to a science built on effective measurements that are predictable and quantifiable.

cs.LG↗

arXiv2025

AI for Pattern Hunter: Application in Wilson Loop of 2D Lattice Yang-Mills Theory

We employ the Transformer to learn patterns in two-dimensional lattice Yang-Mills theory. Specifically, we represent both Wilson loops and their expectation values as tokenized sequences. Taking the shape of Wilson loops as input, the model successfully predicts expectation values with high accuracy, indicating a meaningful connection between loop geometry and physical results. Our study differs from prior machine learning applications in lattice QCD by emphasizing analytical structures rather than numerical computations. We explore model performance under varying hyperparameters, training data sizes, and sequence lengths. This work serves as a first step toward extending such methods to higher dimensions and inspiring rigorous analytical derivations.

hep-th↗

arXiv2025

Iterated Agent for Symbolic Regression

Symbolic regression (SR), the automated discovery of mathematical expressions from data, is a cornerstone of scientific inquiry. However, it is often hindered by the combinatorial explosion of the search space and a tendency to overfit. Popular methods, rooted in genetic programming, explore this space syntactically, often yielding overly complex, uninterpretable models. This paper introduces IdeaSearchFitter, a framework that employs Large Language Models (LLMs) as semantic operators within an evolutionary search. By generating candidate expressions guided by natural-language rationales, our method biases discovery towards models that are not only accurate but also conceptually coherent and interpretable. We demonstrate IdeaSearchFitter's efficacy across diverse challenges: it achieves competitive, noise-robust performance on the Feynman Symbolic Regression Database (FSReD), outperforming several strong baselines; discovers mechanistically aligned models with good accuracy-complexity trade-offs on real-world data; and derives compact, physically-motivated parametrizations for Parton Distribution Functions in a frontier high-energy physics application. IdeaSearchFitter is a specialized module within our broader iterated agent framework, IdeaSearch, which is publicly available at https://www.ideasearch.cn/.

physics.comp-ph↗

arXiv2025

PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models

Current benchmarks for evaluating the reasoning capabilities of Large Language Models (LLMs) face significant limitations: task oversimplification, data contamination, and flawed evaluation items. These deficiencies necessitate more rigorous assessment methods. To address these limitations, we introduce PHYBench, a benchmark of 500 original physics problems ranging from high school to Physics Olympiad difficulty. PHYBench addresses data contamination through original content and employs a systematic curation pipeline to eliminate flawed items. Evaluations show that PHYBench activates more tokens and provides stronger differentiation between reasoning models compared to other baselines like AIME 2024, OlympiadBench and GPQA. Even the best-performing model, Gemini 2.5 Pro, achieves only 36.9% accuracy compared to human experts' 61.9%. To further enhance evaluation precision, we introduce the Expression Edit Distance (EED) Score for mathematical expression assessment, which improves sample efficiency by 204% over binary scoring. Moreover, PHYBench effectively elicits multi-step and multi-condition reasoning, providing a platform for examining models' reasoning robustness, preferences, and deficiencies. The benchmark results and dataset are publicly available at https://www.phybench.cn/.

cs.CL↗

arXiv2025

Bridging the Dimensional Chasm: Uncover Layer-wise Dimensional Reduction in Transformers through Token Correlation

The geometric evolution of token representations in large language models (LLMs) presents a fundamental paradox: while human language inherently organizes semantic information in low-dimensional spaces ($\sim 10^1$ dimensions), modern LLMs employ high-dimensional embeddings ($\sim 10^3$ dimensions) processed through Transformer architectures. To resolve this paradox, this work bridges this conceptual gap by developing a geometric framework that tracks token dynamics across Transformers layers. Through layer-wise analysis of intrinsic dimensions across multiple architectures, we reveal an expansion-contraction pattern where tokens diffuse to a "working space" and then progressively project onto lower-dimensional submanifolds. Our finding implies a negative correlation between the working space dimension and parameter-sensitive performance of the LLMs, and indicates that effective models tend to compress tokens into approximately 10-dimensional submanifolds, closely resembling human semantic spaces. This work not only advances LLM interpretability by reframing Transformers layers as projectors that mediate between high-dimensional computation and low-dimensional semantics, but also provides practical tools for model diagnostics that do not rely on task-specific evaluations.

cs.CL↗

arXiv2025

Explainable AI-assisted Optimization for Feynman Integral Reduction

We present a novel approach to optimizing the reduction of Feynman integrals using integration-by-parts identities. By developing a priority function through the FunSearch algorithm, which combines large language models and genetic algorithms, we achieve significant improvements in memory usage and computational efficiency compared to traditional methods. Our approach demonstrates substantial reductions in the required seeding integrals, making previously intractable integrals more manageable. Tested on a variety of Feynman integrals, including one-loop and multi-loop cases with planar and non-planar configurations, our method demonstrates remarkable scalability and adaptability. For reductions of certain Feynman integrals with many dots and numerators, we observed an improvement by a factor of 3058 compared to traditional methods. This work provides a powerful and interpretable framework for optimizing IBP reductions, paving the way for more efficient and practical calculations in high-energy physics.

hep-ph↗

arXiv2024

Soft Theorem to Three Loops in QCD and ${\cal N} = 4$ Super Yang-Mills Theory

The soft theorem states that scattering amplitude in gauge theory with a soft gauge-boson emission can be factorized into a hard scattering amplitude and a soft factor. In this paper, we present calculations of the soft factor for processes involving two hard colored partons, up to three loops in QCD. To accomplish this, we developed a systematic method for recursively calculating relevant Feynman integrals using the Feynman-Parameter representation. Our results constitute an important ingredient for the subtraction of infrared singularities at N$^4$LO in perturbative QCD. Using the principle of leading transcendentality between QCD and ${\cal N}=4$ super Yang-Mills theory, we determine the soft factor in the latter case to three loops with full-color dependence. As a by-product, we also obtain the finite constant $f_2^{(3)}$ in the Bern-Dixon-Smirnov ansatz analytically, which was previously known numerically only.

hep-ph↗

arXiv2020

Unpolarized Quark and Gluon TMD PDFs and FFs at N$^3$LO

In this paper we calculate analytically the perturbative matching coefficients for unpolarized quark and gluon Transverse-Momentum-Dependent (TMD) Parton Distribution Functions (PDFs) and Fragmentation Functions (FFs) through Next-to-Next-to-Next-to-Leading Order (N$^3$LO) in QCD. The N$^3$LO TMD PDFs are calculated by solving a system of differential equation of Feynman and phase space integrals. The TMD FFs are obtained by analytic continuation from space-like quantities to time-like quantities, taking into account the probability interpretation of TMD PDFs and FFs properly. The coefficient functions for TMD FFs exhibit double logarithmic enhancement at small momentum fraction $z$. We resum such logarithmic terms to the third order in the expansion of $α_s$. Our results constitute important ingredients for precision determination of TMD PDFs and FFs in current and future experiments.

hep-ph↗

arXiv2020

Quark Transverse Parton Distribution at the Next-to-Next-to-Next-to-Leading Order

We report a calculation of the perturbative matching coefficients for the transverse-momentum-dependent parton distribution functions for quark at the next-to-next-to-next-to-leading order in QCD, which involves calculation of non-standard Feynman integrals with rapidity divergence. We introduce a set of generalized Integration-By-Parts equations, which allows an algorithmic evaluation of such integrals using the machinery of modern Feynman integral calculation.

hep-ph↗

arXiv2019

Analytic Next-To-Leading Order Calculation of Energy-Energy Correlation in Gluon-Initiated Higgs Decays

The energy-energy correlation (EEC) function in $e^+e^-$ annihilation is currently the only QCD event shape observable for which we know the full analytic result at the next-to-leading order (NLO). In this work we calculate the EEC observable for gluon initiated Higgs decay analytically at NLO in the Higgs Effective Field Theory (HEFT) framework and provide the full results expressed in terms of classical polylogarithms, including the asymptotic behavior in the collinear and back-to-back limits. This observable can be, in principle, measured at the future $e^+e^-$ colliders such as CEPC, ILC, FCC-ee or CLIC. It provides an interesting opportunity to simultaneously probe our understanding of the strong and Higgs sectors and can be used for the determinations of the strong coupling.

hep-ph↗

arXiv2019

Quark jet versus gluon jet: fully-connected neural networks with high-level features

Jet identification is one of the fields in high energy physics that machine learning has begun to make an impact. More often than not, convolutional neural networks are used to classify jet images with the benefit that essentially no physics input is required. Inspired by a recent work by Datta and Larkoski, we study the classification of quark/gluon-initiated jets based on fully-connected neural networks (FNNs), where expert-designed physical variables are taken as input. FNNs are applied in two ways: trained separately on various narrow jet transverse momentum $p_{TJ}$ bins; trained on a wide region of $p_{TJ} \in [200,~1000]$ GeV. We find their performances are almost the same. The performance is better when the $p_{TJ}$ is larger. Jet discrimination with FNN is studied on both particle and detector level data. The results based on particle level data are comparable with those from deep convolutional neural networks, while the significance improvement characteristic (SIC) from detector level data would at most decrease by $15\%$. We also test the performance of FNNs with full set or subsets of jet observables as input features. The FNN with one subset consisting of fourteen observables shows nearly no degradation of performance. This indicates that these fourteen expert-designed observables could have captured the most necessary information for separating quark and gluon jets.

hep-ph↗

arXiv2018

Analytic calculation of Energy-Energy Correlation in $e^+ e^-$ annihilation at NLO

We present the first fully analytic calculation of the Quantum Chromodynamics (QCD) event shape observable Energy-Energy Correlation in electron-positron annihilation at Next-To-Leading Order (NLO). This result sheds light on the analytic structure of the event shape observables beyond Leading Order (LO) and serves as a motivation to employ our methods in the investigation of other event shape observables that so far have not been calculated analytically.

hep-ph↗

arXiv2018

The Energy-Energy Correlation at Next-to-Leading Order in QCD, Analytically

The energy-energy correlation (EEC) between two detectors in $e^+e^-$ annihilation was computed analytically at leading order in QCD almost 40 years ago, and numerically at next-to-leading order (NLO) starting in the 1980s. We present the first analytical result for the EEC at NLO, which is remarkably simple, and facilitates analytical study of the perturbative structure of the EEC. We provide the expansion of EEC in the collinear and back-to-back regions through to next-to-leading power, information which should aid resummation in these regions.

hep-ph↗

arXiv2016

Note on recursion relations for the $\mathcal{Q}$-cut representation

In this note, we study the $\mathcal{Q}$-cut representation by combining it with BCFW deformation. As a consequence, the one-loop integrand is expressed in terms of a recursion relation, i.e., $n$-point one-loop integrand is constructed using tree-level amplitudes and $m$-point one-loop integrands with $m\leq n-1$. By giving explicit examples, we show that the integrand from the recursion relation is equivalent to that from Feynman diagrams or the original $\mathcal{Q}$-cut construction, up to scale free terms.

hep-th↗

arXiv2016

Feynman Rules of Higher-order Poles in CHY Construction

In this paper, we generalize the integration rules for scattering equations to situations where higher-order poles are present. We describe the strategy to deduce the Feynman rules of higher-order poles from known analytic results of simple CHY-integrands, and propose the Feynman rules for single double pole and triple pole as well as duplex-double pole and triplex-double pole structures. We demonstrate the validation and strength of these rules by ample non-trivial examples.

hep-th↗

arXiv2016

Squarkonium, diquarkonium and octetonium at the LHC and their di-photon decays

Motivated by the recent di-photon excess by both ATLAS and CMS collaborations at the LHC, we systematically investigate the production and di-photon decay of onia formed by pair of all possible color exotic scalars in minimal extension. When such scalar massive meta-stable colored and charged (MMCC) particles are produced in pair near threshold, $η$ onium can be formed and decay into di-photon through annihilation as $pp\to η\to γγ$. Squarkonium is formed by meta-stable squarks in supersymmetric models such as stoponium. Diquarkonium is formed by meta-stable color sextet diquarks which may be realized in the Pati-Salam model. Octetonium is formed by color octet scalars bosons as in the Manohar-Wise model. Stoponium prediction is much smaller than the required signal to account for the di-photon excess. Due to the enhancement factor from color and electric charge, predictions of diquarkonium and octetonium are of $\cal O$(10 fb) which are significantly greater than the stoponium prediction. Since the color enhancement also results in large production at the colliders, such light color exotic states of $\cal O$(375 GeV) suffer from severe direct search constraints. On the other hand, if their dominant decay mode involve top quark, they may be buried in the $t\bar{t}$ plus jets samples and can potentially be searched via $t+j$ resonance.

hep-ph↗