The motivation comes from a shift in how artificial intelligence is used in high-energy physics. Machine learning has long been embedded in the field, powering event reconstruction, jet tagging, anomaly detection, fast simulation and statistical inference. More recently, however, agents built on LLMs have begun to move beyond isolated inference tasks toward the orchestration of entire scientific workflows, including code generation, tool invocation, structured context management and human-in-the-loop analysis. Parameter-space exploration, the authors argue, is a natural next target, because a substantial part of the work lies not in choosing a sampling algorithm but in assembling the surrounding computational machinery: connecting external physics programs, modifying input cards point by point, reading output observables, defining likelihoods and constraints, and storing the whole setup alongside its results.

EasyScan_HEP 2 takes a distinctive architectural stance. Rather than improving the scan engine itself, the framework lets AI assist the configuration layer. The scientific content of any scan remains encoded in an explicit .ini configuration file that specifies the scan method, input parameters, external programs, input-output mappings, constraints, plots and result folder. An LLM agent can generate or revise this file from a natural-language request, but the file itself remains the single source of truth, executable by the same backend that powered the original EasyScan_HEP. This design means the AI never silently changes the physics; it only drafts a document that the user can inspect before anything runs.

Several machine-readable interfaces make this workflow practical. The package is now an installable Python tool with a command available from any working directory. A dry-run configuration checker parses the .ini file and reports errors, warnings and informational messages without launching any scan points, catching common failure modes such as unsupported scan methods, wrong paths, missing likelihood constraints, duplicated variable names, invalid numerical ranges or plot variables that no input or output block defines. Crucially, the checker can return machine-readable output, allowing an agent to repair its own mistakes iteratively. The authors are careful to note that the checker verifies only syntactic and operational consistency, not whether the underlying physics model is correct.

The run interface has also been made agent-friendly. An explicit overwrite policy replaces interactive prompts, and a structured JSON report records whether the run succeeded, the return code, the command used, the launch directory, the configuration path, the log path, the result directory and the overwrite action. A separate result-reader command summarizes an existing result directory without rerunning the scan, counting rows, listing generated plots and identifying a representative best row by minimizing chi-squared or minus-two-log-likelihood columns where available. Because this summary comes from a deterministic reader rather than the model’s interpretation of terminal output, users do not have to trust the agent’s reading of raw logs.

To quantify the benefit, the team ran a controlled evaluation with 11 benchmark tasks, each repeated three times under three conditions, giving 99 isolated runs using the gpt-5.6-terra model with medium reasoning effort in the same Codex environment. Condition S used EasyScan_HEP with its documentation and the registered agent skill; condition D used the package without the skill; and condition M had the model implement each scan directly without EasyScan_HEP. Both EasyScan_HEP conditions completed all 33 runs end to end, while the direct-implementation condition completed 31 of 33 and required 77 task executions, including 27 failed launches. The median number of code lines requiring user review was roughly five times larger without the framework.

The skill itself delivered a modest but measurable efficiency gain: first-execution success rose from 22 to 25 out of 30 executable runs, and total task executions fell from 46 to 40. The authors emphasize that the benchmark does not capture broader functions of the skill, such as organizing an end-to-end workflow, guiding software setup, and converting missing information in a user prompt into explicit follow-up questions. That conservative behavior is deliberate: if a user does not specify the location of an external program, the skill asks for the path rather than searching the file system and risking a wrong choice among multiple installed versions, since such details must be checked by the user in any case.

The modular design also made it straightforward to add three new scan methods through an LLM-agent-guided workflow. BESTFIT performs differential-evolution minimization of the configured chi-squared via SciPy, aimed at quickly locating a good-fit point. EMCEE adds ensemble Markov-chain Monte Carlo sampling with a configurable number of walkers, writing a flattened chain file for post-processing. DYNESTY brings Python-based nested sampling, storing log-likelihoods, log-weights and evidence-related quantities without requiring the native MultiNest libraries. Because a new method only needs to decide how points are proposed in parameter space, while the common workflow handles priors, external programs, constraints and plotting, the extension route has been encoded directly into the agent skill.

To demonstrate the framework on real physics, the authors scanned the Z2-symmetric real singlet scalar extension of the Standard Model, a minimal Higgs-portal model in which a new stable scalar can serve as a dark matter candidate. A two-dimensional grid scan over the singlet mass and portal coupling, with the singlet self-coupling fixed, chained together micrOMEGAs 7.1 for the relic dark matter density and PhaseTracer 2 for the electroweak phase transition, including an explicit convention conversion in which the portal coupling passed to PhaseTracer differs by a factor of two. The same scan could be prepared three equivalent ways: by writing the configuration file directly, through the agent skill from a natural-language prompt, or via a local single-user Web interface that loads, edits, checks and runs the same files. A follow-up plot of the transition strength and relic-density contours was produced by simply asking the agent to post-process the saved result table.

The broader significance is that EasyScan_HEP 2 offers a template for how AI agents can enter computationally intensive science without eroding scientific accountability. The agent drafts and repairs configurations, but the checker validates them, the Web interface exposes them, the runner records exactly how they were executed, and the result reader summarizes outputs deterministically. Generated configurations must still be inspected by the user, and no agent replaces physics validation. As LLM-agent workflows mature across the field, from collider analyses to dark-matter phenomenology, this configuration-centered approach, keeping every calculation tied to an explicit, checkable scan description, may prove to be the model that lets physicists embrace autonomous assistants while keeping the final word firmly in human hands.

Subject of Research: LLM-agent-assisted parameter-scan workflows for high-energy physics phenomenology

Article Title: EasyScan_HEP 2: LLM-agent parameter-scan workflows in high energy physics

Article References: Xiao, Y., Yue, Y., & Zhang, Y. (2026). EasyScan_HEP 2: LLM-agent parameter-scan workflows in high energy physics. The European Physical Journal C, 86(9), Article 1101. https://doi.org/10.1140/epjc/s10052-026-16358-9

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-16358-9

Keywords: large language models, AI agents, high-energy physics, parameter scans, EasyScan_HEP, beyond the Standard Model, dark matter, electroweak phase transition, scientific workflows, reproducibility, Markov-chain Monte Carlo, nested sampling

Spectral retrieval is, at its core, a Bayesian inverse problem. Scientists build a parametric model of an atmosphere, generate a synthetic spectrum, and compare it against the observed data, repeating the process tens of thousands of times across a vast parameter space. Bayes’ theorem converts the likelihood of the data given a set of atmospheric parameters into the posterior probability of those parameters given the data. For solar system planets, where orbiters and descent probes provide strong prior knowledge, fast matrix-inversion methods like Optimal Estimation work well. For exoplanets, where nothing is known for certain, those restrictive Gaussian priors can badly bias the answer, so the field turned instead to computationally expensive sampling algorithms.

Markov Chain Monte Carlo became the early standard for exoplanet retrieval, but it struggles when the probability landscape is multi-modal, with several distinct families of solutions. Nested Sampling, first trialled for exoplanets in 2013, solved that problem and delivers the Bayesian Evidence as a by-product, which enables model comparison. The catch is cost. Nested sampling’s computational cost scales roughly with the cube of the number of model parameters, and modern retrievals now routinely require between ten thousand and one hundred million forward model evaluations per dataset. A single retrieval of the well-studied hot Jupiter WASP-39b using JWST data consumed on the order of ten thousand core hours.

The forward model itself, a radiative transfer calculation through a modelled atmosphere, is the true bottleneck, since it is evaluated at every sampler step. At its heart lies the radiative transfer equation, balancing absorption, scattering and emission along the light’s path, which differs dramatically between transit, eclipse and direct imaging geometries. Building on this foundation, the model must specify a temperature-pressure profile, often parameterized with the physically motivated Guillot profile or the more flexible six-parameter form introduced by Madhusudhan and Seager in 2009. Both remain restrictive, and studies show that neither fully captures the curvature of real three-dimensional temperature structures, with arithmetic averages of three-dimensional profiles retrieved more faithfully than more realistic weighted ones.

Chemistry introduces another fork in the road. Free-chemistry retrievals let every gas abundance float, allowing the unexpected, such as the surprise detection of sulfur dioxide on WASP-39b that no equilibrium model predicted and which ultimately revealed a photochemical production mechanism. But free chemistry inflates the parameter count enormously. Tying the model to a chemical network slashes the parameters to a handful, like metallicity and carbon-to-oxygen ratio, yet the choice of network matters enormously: in one test, a reduced chemical network produced an apparently good fit while retrieving a metallicity six times solar when the true input was one times solar. The most complex chemical model run inside a retrieval to date, the FRECKLL framework, took about five minutes per evaluation and needed forty thousand samples, roughly 138 days of CPU time spread over 180 cores.

Opacity data adds its own burden. Because exoplanet atmospheres are far hotter than laboratory conditions can safely replicate, absorption line positions and strengths rely on quantum mechanical simulations, generating line lists with millions of entries. Full line-by-line calculations are intractably slow for retrievals, so codes use approximations: correlated-k tables or pre-computed cross sections. Recent tests on the JWST spectrum of WASP-39b showed that cross sections computed below resolutions of roughly fifty thousand can bias retrieved gas abundances, a hidden error source at the very heart of the comparison. Broadening of spectral lines, which depends on the ambient gas composition and temperature, and the enormous numbers of weak methane lines at high temperatures, sometimes collapsed into so-called superlines, add further layers of compromise between accuracy and speed.

Clouds are arguably the hardest problem of all. Aerosols are essentially ubiquitous, shaping spectra through their altitude, particle size, composition and abundance, none of which is well known for any exoplanet. Attempts to predict cloud decks from first principles fail even for Jupiter and Saturn, where ammonia clouds predicted by microphysics models are simply not seen across most of the disk. Exoplanet cloud parameterizations remain crude, often little more than a cloud-top pressure and a wavelength-dependent opacity, and the common extinction-only approximation, which assumes every photon interacting with a cloud is scattered out of the beam, can substantially underestimate atmospheric transmission when forward scattering dominates. Correct multiple-scattering treatment demands Monte Carlo photon tracking, a severe computational penalty.

The star itself is no innocent bystander. The Transit Light Source Effect arises when unocculted starspots or faculae imprint their own spectral fingerprints onto a transiting planet’s spectrum, and recent 3D magnetohydrodynamic simulations show that standard stellar atmosphere models misrepresent spot spectra by more than one hundred parts per million at some wavelengths. Meanwhile, JWST data are now precise enough that one-dimensional, homogeneous atmosphere models are demonstrably inadequate. Retrievals have begun incorporating separate day and night terminator chemistries, and full three-dimensional radiative transfer frameworks like TRIDENT can extract morning-evening and day-night gradients, at a cost of roughly a factor of twenty-five in computation time. Even the choice of how spectra are binned can inject resolution-linked bias that distorts retrieved transit depths.

Interpreting the results demands equal care. Bayes factors comparing models with and without a given molecule are frequently reported as detection significances in sigma, but recent analyses warn that these are relative model preferences, not physical detections. A claimed detection of dimethyl sulphide on the sub-Neptune K2-18b was shown to hinge on a narrowly restricted model space in which alternative hydrocarbons, untested, fit the same data equally well or better. Cross-validation techniques that leave out individual data points have revealed apparent detections resting on a single broadband measurement. Machine learning offers a possible escape: neural network emulators of radiative transfer and chemistry can accelerate retrievals dramatically, but their computational advantage degrades rapidly as dimensionality grows, and quantifying their uncertainties remains an open problem.

The road ahead points toward even heavier demands. The Ariel mission, launching in 2031, will characterize at least a thousand planets, driving a shift toward machine learning pipelines. The extremely large telescopes will bring high-resolution cross-correlation spectroscopy and reflected light imaging, the latter requiring full multiple-scattering models and possibly polarization, techniques not yet implemented in any retrieval code. The Habitable Worlds Observatory aims to image Earth twins in reflected light, where the stakes of every modelling choice become highest. Barstow and Welbanks close with practical advice: treat retrieval software as more than a black box, justify priors and likelihoods explicitly, invest in code efficiency and software engineering, and approach every detection claim, especially for small temperate worlds, with the skepticism the data deserve.

One underappreciated aspect of the field’s growth is its sheer diversity of tooling. More than fifty independent retrieval frameworks have now been applied to exoplanets, spanning a wide range of sampling algorithms, temperature and cloud parameterizations, and treatments of chemistry ranging from strict equilibrium assumptions to fully flexible free schemes. Many of these codes are open source, reflecting a community culture that has encouraged sharing and scrutiny, yet diversity alone does not guarantee agreement.

To address that concern, teams have undertaken systematic benchmarking exercises, including model intercomparison projects in which different codes are run against identical synthetic datasets. These efforts have revealed a subtle but consequential finding: small differences in model implementation, producing variations of only a few tens of parts per million in synthetic spectra, can cascade into substantial differences in the values retrieved from the same data. In an era when observational precisions are measured at similar levels, such implementation details are no longer negligible.

The review also situates the field historically. Retrieval was long a workhorse for solar system science, where it constrained the structure of Jupiter’s equatorial cloud decks from Galileo orbiter data, mapped spatial variation in ammonia on Saturn, and probed surface emissivity variations on Venus. The migration of these techniques to exoplanets was accelerated in part by solar system atmospheric scientists joining the field, bringing with them both expertise and an awareness of the pitfalls of applying methods tuned to well-characterized planets to worlds about which almost nothing is known.

That heritage explains a recurring theme: because exoplanet exploration lacks ground truth, the choice of algorithm and prior is itself a scientific decision with measurable consequences. Early investigations using synthetic data demonstrated that methods constrained by Gaussian assumptions could recover incorrect solutions when data were sparse, while broader exploration of parameter space recovered the truth. The authors’ recommendations, from justifying priors explicitly to treating software as more than a black box, flow directly from lessons like these, hard-won across two decades of practice in both solar system and exoplanet contexts.

Subject of Research: Computational challenges in Bayesian spectral retrieval of exoplanet atmospheres

Article Title: Computational challenges in exoplanet atmospheric retrieval

Article References: K. Barstow, J., & Welbanks, L. (2026). Computational challenges in exoplanet atmospheric retrieval. Living Reviews in Computational Astrophysics, 12(1), Article 6. https://doi.org/10.1007/s41115-026-00031-9

Image Credits: AI Generated

DOI: 10.1007/s41115-026-00031-9

Keywords: exoplanets, atmospheric retrieval, JWST, Bayesian inference, nested sampling, radiative transfer, cloud modelling, molecular opacities, machine learning, stellar contamination, WASP-39b, Ariel mission