Lilith: A Multi-Domain Research Artifact

• LILITH is a multi-purpose research artifact spanning Higgs phenomenology, laser-plasma neutron diagnostics, exoplanet simulation, and conceptual large-language-model architectures.
• In high-energy physics, Lilith processes Higgs signal strengths using advanced likelihood techniques and an extensible XML-based database to constrain new physics.
• For experimental and simulation applications, LILITH underpins precise neutron time-of-flight measurements, realistic TESS instrument simulations, and conceptual modular AI frameworks.

LILITH, or Lilith, is used in recent arXiv literature for several technically unrelated research artifacts rather than a single framework. The name denotes a public Higgs-likelihood package for constraining beyond-the-Standard-Model scenarios from collider measurements, a dedicated neutron time-of-flight spectrometer for laser-driven fusion experiments, a physics-based TESS instrument-and-sky simulator, and a conceptual modular large-language-model architecture aimed at studying consciousness emergence (Bernon et al., 2015, Stuhl et al., 12 Feb 2025, Osborn et al., 2019, Farooqi et al., 6 Jul 2025).

Usage	Domain	Core function
Lilith	Higgs phenomenology	Global likelihood from Higgs signal strengths
LILITH	Laser-plasma neutron diagnostics	Multi-detector neutron TOF spectrometer
Lilith	Exoplanet survey simulation	Physics-based TESS raw-data simulator
LILITH	AI / consciousness studies	Developmental modular LLM architecture

1. Higgs-likelihood software: identity and evolution

In high-energy phenomenology, Lilith is a public, lightweight Python library whose purpose is to turn published Higgs signal-strength measurements into quantitative constraints on new physics. The early formulation, Lilith 1.0, was presented as “LIght LIkelihood fiT for the Higgs,” a compact Python code that builds the likelihood of a Higgs boson with mass around $125$ GeV from LHC and Tevatron signal-strength measurements and uses this likelihood to fit Higgs couplings and constrain beyond-the-Standard-Model scenarios (Bernon et al., 2014). The more detailed tool paper describes Lilith as a new public tool that makes use of signal-strength measurements performed at the LHC and the Tevatron, with the Higgs likelihood based on experimental results stored in an easily extensible XML database and evaluated from user input given in XML format in terms of reduced couplings or signal strengths (Bernon et al., 2015).

Subsequent versions expanded both the database and the statistical treatment. Lilith-2.0 extended the Run-1 database to include the ATLAS and CMS Run 2 Higgs results for $36/\mathrm{fb}$, replaced the ordinary Gaussian approximation of Lilith-1.1 with variable Gaussian and Poisson likelihoods, and added support for correlation matrices of arbitrary dimension (Bertrand et al., 2020). The proceedings update on global fits emphasizes that the public Lilith-2.1 release uses signal-strength data, while ongoing development aims to include Higgs STXS data and SMEFT parametrizations for more differential global analyses (Nguyen et al., 2023).

This development trajectory situates Lilith as a precision Higgs-fit engine rather than a Monte Carlo generator or spectrum calculator. Its role is inferential: given a model prediction for a $125$ GeV Higgs sector, it evaluates consistency with the accumulated ATLAS and CMS measurements. That role remains stable across versions even as the database, likelihood parameterizations, and interoperability with other packages evolve (Bernon et al., 2015, Bertrand et al., 2020).

2. Statistical formalism, inputs, and computational interfaces

The basic observable in Lilith is the Higgs signal strength,

$$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$

with $X$ denoting a production mode such as ggH, VBF, WH, ZH, ttH, or tH, and $Y$ a decay mode such as $\gamma\gamma$, $ZZ^*$, $WW^*$, $b\bar b$, or $36/\mathrm{fb}$0 (Bernon et al., 2015, Bertrand et al., 2020). In reduced-coupling mode, Lilith encodes deviations from the Standard Model through rescaling factors $36/\mathrm{fb}$1 or $36/\mathrm{fb}$2, and the signal strengths are reconstructed from these reduced couplings together with the modified total width (Bernon et al., 2015, Barducci et al., 2016). Invisible and undetected decays are incorporated through $36/\mathrm{fb}$3 and $36/\mathrm{fb}$4, which suppress visible rates through the total-width rescaling (Bernon et al., 2015, Bertrand et al., 2020).

The likelihood construction is global. In micrOMEGAs-oriented notation,

$36/\mathrm{fb}$5

and when the individual likelihoods are Gaussian this becomes a $36/\mathrm{fb}$6-type test statistic (Barducci et al., 2016). Lilith-2 generalizes the experimental likelihood model from symmetric Gaussians to variable Gaussian and generalized Poisson forms, and can use correlation matrices of arbitrary dimension, including the $36/\mathrm{fb}$7 CMS matrix for $36/\mathrm{fb}$8 combinations (Bertrand et al., 2020). The STXS development keeps the same logic—experimental observables plus covariance information mapped into a global likelihood—but changes the observable basis from inclusive signal strengths to differential Simplified Template Cross Sections (Nguyen et al., 2023).

The software architecture is deliberately lightweight. Experimental inputs are stored in XML files, user models can be provided either as reduced couplings or as direct signal strengths, and the code exposes Python, command-line, and C/C++/ROOT interfaces (Bernon et al., 2015). In micrOMEGAs 4.3, the program writes Lilith_in.xml, calls Lilith externally, and reads back an SLHA-like Lilith_out.slha containing $36/\mathrm{fb}$9, the number of experimental degrees of freedom, a p-value, and the database version (Barducci et al., 2016). In the 2026 FlexibleSUSY interface, normalized effective Higgs couplings are constructed in NORMALIZEDEFFHIGGSCOUPLINGS and IMNORMALIZEDEFFHIGGSCOUPLINGS, invisible and undetected branching ratios are identified from computed widths, and the resulting reduced couplings are passed to Lilith to obtain a global $125$0 and a derived p-value (Kotlarski et al., 12 Mar 2026).

3. Phenomenological applications, validation, and scope conditions

Lilith is used for both model-independent and model-specific Higgs fits. The original papers validated the tool against ATLAS and CMS coupling fits in benchmark planes such as $125$1 and $125$2, showing that the reconstructed likelihood reproduces official contours with good accuracy when the public information is sufficient (Bernon et al., 2015, Bernon et al., 2014). Lilith has also been used in scans of Type I and Type II Two-Higgs-Doublet Models, including alignment-limit studies in which either the lighter or the heavier CP-even Higgs is identified with the $125$3 GeV state (Bernon et al., 2015, Bernon et al., 2015). In those applications, 2HDMC provides the physical couplings and branching ratios, while Lilith supplies the 125 GeV Higgs signal-strength likelihood.

Within broader phenomenological workflows, Lilith is complementary to HiggsSignals, HiggsBounds, and SModelS. micrOMEGAs uses Lilith or HiggsSignals for the measured $125$4 GeV Higgs signal strengths, HiggsBounds for additional Higgs states, and SModelS for direct searches for new particles; the package explicitly notes that Lilith and HiggsSignals perform the same task and only one should be used in a global fit (Barducci et al., 2016). FlexibleSUSY now provides a fully automated chain from a user-defined BSM Lagrangian to Lilith and HiggsTools, including handling of invisible and undetected decay widths, CP-violating couplings, and neutral Higgs reduced couplings extracted from computed partial widths (Kotlarski et al., 12 Mar 2026).

The scope conditions are important. The signal-strength approach assumes the same structure of Higgs vertices as in the Standard Model and the same production modes as in the Standard Model, with rescaled couplings (Barducci et al., 2016). This is why the STXS and SMEFT extensions matter: inclusive $125$5-based fits assume SM-like acceptances and are dominated by overall rates, whereas STXS retains more information on kinematic structure and is better suited to constraining energy-growing SMEFT operators (Nguyen et al., 2023). A common misconception is therefore to treat Lilith as a generic event-level reinterpretation framework; in the cited literature it is instead a reduced-coupling and signal-strength likelihood engine, with STXS-based differential fits still under development (Nguyen et al., 2023).

4. LILITH as a neutron time-of-flight spectrometer

In laser-plasma neutron diagnostics, LILITH denotes a dedicated, multi-detector time-of-flight spectrometer built to characterize laser-driven fusion neutrons on a shot-by-shot basis at $125$6 Hz. In the neutron-source experiment reported in 2025, it is the central neutron diagnostic used to reconstruct neutron energy spectra, angular distributions, total yields, and source stability for the $125$7 reaction in a pitcher-catcher configuration (Stuhl et al., 12 Feb 2025).

The instrument is an array-type neutron TOF spectrometer composed of eight organic scintillation detectors arranged around the catcher target. It is divided into two subsystems: LILITH-M, consisting of four fast plastic EJ-230 detectors at about $125$8, and LILITH-XL, consisting of four EJ-309 liquid scintillators at about $125$9. The eight detectors collectively cover nearly $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$0 in polar angle around the catcher. All detectors are surrounded by $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$1 cm of lead shielding to attenuate X-rays and gamma rays from the laser-plasma interaction (Stuhl et al., 12 Feb 2025).

The detection principle is timing-based neutron-photon discrimination. A photodiode monitoring the laser provides a reference trigger time; prompt photons appear at about $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$2 ns in the medium detectors and about $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$3 ns in the long-flight-path detectors, while neutrons are integrated in windows of $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$4–$$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$5 ns for LILITH-M and $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$6–$$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$7 ns for LILITH-XL. Neutron energies are reconstructed from

$$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$8

after an offline timing adjustment of $$\mu(X,Y) \equiv \frac{\sigma(X)\,\mathrm{BR}(H\to Y)}{\sigma^{\rm SM}(X)\,\mathrm{BR}^{\rm SM}(H\to Y)}\,,$$9 ns to account for the deuteron time of flight from the pitcher to the catcher,

$X$0

The quoted energy resolution is about $X$1 at $X$2 for LILITH-M and about $X$3 for LILITH-XL. Light-output calibration uses $X$4, $X$5, and $X$6, while neutron efficiencies are calibrated with a PuBe source (Stuhl et al., 12 Feb 2025).

Operationally, LILITH established the central performance claims of the source. In the longest continuous run of $X$7 shots at $X$8 Hz, it recorded over $X$9 neutron events in the four M detectors and over $Y$0 neutron events in the four XL detectors, implying an average yield of $Y$1 neutrons/shot and a peak yield of $Y$2 neutrons/shot over a $Y$3-minute interval. At $Y$4 Hz, this corresponds to about $Y$5 neutrons/s and a laser-to-neutron conversion rate of about $Y$6 neutrons/J. The angular distribution is peaked in the forward direction and enhanced backward, with lower yield at $Y$7, and the system operated continuously for several hours per day with an rms stability of about $Y$8 during the longest run (Stuhl et al., 12 Feb 2025).

5. Lilith as a TESS instrument-and-sky simulator

In exoplanet survey methodology, Lilith is a physics-based TESS instrument and sky model that creates raw TESS data including models for the CCDs, readout electronics, camera optics, behavior of the attitude control system, spacecraft orbit, spacecraft jitter, zodiacal light, and the TESS Input Catalog, together with expected instrumental artifacts and systematics such as cosmic rays, CCD readout errors, thermal-induced focus errors, and spacecraft jitter-induced pointing errors (Osborn et al., 2019).

The same simulator incorporates realistic instances of stellar astrophysics, including stellar variability, eclipsing binaries, background eclipsing binaries, transiting planets, and diffuse light. In the study on rapid classification of TESS planet candidates, the relevant simulation campaign is TSOP-301, a full 4-sector end-to-end run of the TESS science processing pipeline. It simulated 4 sectors with 16,000 targets per sector using the TESS Input Catalog version 6; many targets were near the ecliptic pole and were therefore observed for more than one sector. The injected populations were deliberately tuned for machine-learning utility rather than astrophysical occurrence realism: 20% of all targets had planetary transits, and an additional 20% had eclipsing binaries or background eclipsing binaries (Osborn et al., 2019).

Lilith-generated raw images were processed through the full TESS SPOC pipeline, from raw pixel calibration to transit search and Data Validation products, so that the training inputs for the convolutional neural networks were realistic Threshold Crossing Events with full instrumental and astrophysical ground truth (Osborn et al., 2019). This use of Lilith enabled the first deep-learning model capable of classifying TESS planet candidates. On the simulated data, the binary planet/not-planet model achieved $Y$9 average precision and $\gamma\gamma$0 accuracy on planets, with an additional $\gamma\gamma$1 accuracy gain if planets found at wrong periods were included. When applied to real TESS data, $\gamma\gamma$2 of Threshold Crossing Events coincident with currently published TOIs were recovered as planets, $\gamma\gamma$3 more were suggested to be eclipsing binaries, and a further $\gamma\gamma$4 Threshold Crossing Events were proposed as planet candidates (Osborn et al., 2019).

A frequent misunderstanding would be to treat this Lilith as a classifier. The paper distinguishes the simulator from the classifier: Lilith produces the end-to-end pixel-level simulations and ground truth, while the convolutional neural networks operate on the SPOC-derived light curves, centroid series, and scalar features generated from those simulations (Osborn et al., 2019).

6. LILITH as a conceptual modular LLM architecture

In AI and consciousness studies, LILITH is a proposed developmental, modular large-language-model system composed of multiple specialized LLM “brain regions” that communicate through token-based protocols intended to be analogous to chemical signaling in the brain (Farooqi et al., 6 Jul 2025). The architecture is explicitly conceptual rather than implemented. The paper states that the goal is not to propose an incremental performance improvement, but to put forward a framework in which one optimizes for consciousness emergence rather than task performance, while recognizing substantial implementation challenges (Farooqi et al., 6 Jul 2025).

The proposed modules are a thinking region, a memory region, a sensory region, and a brain stem or regulatory region. The sensory region is the only module that directly receives external input. The memory region has exclusive ability to write to a database. The regulatory region is pre-prompted with basic regulatory rules or goals and has special authority to influence the behavior of other regions. Communication is token-based, but the signal types and semantics are not predefined; the intended design is that they emerge through developmental training (Farooqi et al., 6 Jul 2025).

Training is likewise developmental rather than conventional pre-training. The proposal starts from untrained LLM architectures, places them in a simulated environment, and lets them learn through simulated life experiences, social interaction with other LLMs, and evolutionary selection over generations. The paper also proposes using Integrated Information Theory metrics to investigate consciousness emergence at multiple scales, from inter-module communication to larger system organization (Farooqi et al., 6 Jul 2025). This suggests a research program centered on empirical study of integrated information and inter-module signaling dynamics, not a presently deployable architecture.

The main limitation is explicit in the source: no implementation, no detailed optimization algorithm, and no scalable IIT computation pipeline are given. The work is therefore best understood as a conceptual architecture paper that reuses the LILITH name in a domain entirely distinct from both Higgs phenomenology and experimental diagnostics (Farooqi et al., 6 Jul 2025).