AXCEL: Dual Research Systems Overview

• AXCEL is an acronym representing two distinct systems: a high-pressure xenon gas TPC for 0νββ decay detection and an LLM-based, explainable inconsistency evaluation metric.
• The TPC system leverages an innovative Electroluminescence Light Collection Cell (ELCC) for near-intrinsic energy resolution and robust 3D track imaging for effective background rejection.
• The LLM evaluation system employs a prompt-based Chain-of-Thought approach to extract, justify, and score factual claims, outperforming traditional metrics with transparent reasoning.

AXCEL is used as an acronym for two unrelated, high-impact research systems: (1) a modular, high-pressure xenon gas time projection chamber (TPC) optimized for neutrinoless double-beta decay searches in particle physics (Ban et al., 2017, Yoshida et al., 2023), and (2) Automated eXplainable Consistency Evaluation using LLMs, a prompt-based metric for evaluating factual consistency in text generation (Sreekar et al., 2024). This entry enumerates both domains for completeness.

1. AXEL: High-Pressure Xenon Gas TPC for 0νββ Searches

AXEL denotes a scalable high-pressure xenon gas TPC detector for searching neutrinoless double-beta decay ($0\nu\beta\beta$) of $^{136}$Xe. The central innovation is the Electroluminescence Light Collection Cell (ELCC) readout, designed to deliver near-intrinsic energy resolution and detailed 3D track imaging, which enables powerful background rejection in rare-event searches.

1.1 Detector Architecture and ELCC Principle

AXEL employs a cylindrical vessel containing xenon gas at pressures up to 10 bar. Ionization electrons generated by charged particle interactions drift under a uniform electric field toward the anode, passing through a modular ELCC plane comprising:

• Copper anode plate (at high negative voltage), PTFE spacer, grounded mesh.
• Cell lattice: typical pitch 10 mm, hole diameter $\sim$5 mm; each hole backed with a VUV-sensitive SiPM (3 × 3 mm² or larger).
• Field lines are focused into holes with near-100% electron collection efficiency for appropriate EL region fields ($E_\mathrm{EL} \geq 2.5$ kV/cm/bar).

The electroluminescence region (EL) produces VUV photons via proportional scintillation without avalanche. The photon yield per electron per cm in xenon is given by:

$$\frac{dN_\mathrm{EL}}{dx} \approx 70\cdot(E/p-1)\cdot p$$

with $E$ electric field (kV/cm), $p$ pressure (bar).

1.2 Energy Resolution Measurements

ELCC readout achieves decoupling of light-collection uniformity from detector volume. Energy resolution is quantified as full width at half maximum (FWHM), scaling approximately as $1/\sqrt{E}$. Prototypes attained:

• 4.0 $\pm$ 0.30% FWHM at 122 keV (10 L, 4 bar Xe) (Ban et al., 2017).
• 0.73 $\pm$ 0.11% FWHM at 1836 keV (180 L, 7.6 bar Xe) (Yoshida et al., 2023).
• Extrapolated resolutions at $Q_{\beta\beta}=2458$ keV: $0.6 \pm 0.03\%$ or $0.7 \pm 0.21\%$ depending on fit model (Yoshida et al., 2023).

Mitigating factors—including fluctuations due to the Fano factor ($F=0.13$–0.14), photon statistics ($g$ detected photons per electron), time variation of EL gain, and channel calibration errors—yield total broadening consistent with observed values.

1.3 Track Topology and Background Rejection

ELCC enables pixelized x–y and drift-time z reconstruction, supporting full 3D event topology for discrimination. Track features:

• MeV electron tracks form two terminal “blobs” (signature of $0\nu\beta\beta$).
• Single-site backgrounds (from $\gamma$’s, $\beta$’s) exhibit distinct topologies.
• Diffusion coefficients ($D_T$, $D_L$) measured and consistent with Magboltz predictions, confirming simulation reliability for event reconstruction.

1.4 Scalability, Upgrades, and Future Directions

The ELCC concept supports modular tiling, enabling scale-up to ton-scale detectors without loss of energy resolution or mechanical uniformity. Upgrades include:

• SiPMs with doubled sensitive area.
• Reinforced mesh/electrode architectures for higher EL fields.
• Wavelength-shifting plates to maximize primary scintillation (S1) collection.
• Extended drift geometry and increased number of ELCC modules.

These advancements target sub-1% FWHM energy resolution at Qββ with robust topological discrimination, validating AXEL for next-generation rare-event experiments.

2. AXCEL: Automated eXplainable Consistency Evaluation using LLMs

AXCEL is a prompt-based, automated metric for assessing the factual consistency of text generated by LLMs. The key contributions are explainability, generalizability across tasks, and strong empirical correlation with human judgment (Sreekar et al., 2024).

2.1 Motivation and Limitations of Prior Metrics

Traditional n-gram metrics like ROUGE and BLEU require reference texts and correlate weakly with human consistency judgments. Reference-free NLI-based metrics (e.g., AlignScore) improve correlation but require domain specific training and lack explanation capabilities. Prior prompt-based LLM metrics (G-Eval, ChatGPT-Eval) are black boxes, producing only global scores and often requiring task-specific prompts.

2.2 AXCEL Methodology and Workflow

AXCEL formalizes consistency evaluation as follows:

• Extract all atomic facts $I(DT)$ from the derived text $DT$.
• For each fact $f_i\in I(DT)$, verify against the source text $ST$, assign score $s_i\in\{1,2,3,4,5\}$, and produce rationale $r_i$.
• Aggregate scores:

$$C(DT,ST) = \frac{1}{|I(DT)|} \sum_{f \in I(DT)} V(f,ST)$$

where $V(f,ST)$ is the fact-level verification score.

AXCEL uses a single Chain-of-Thought (CoT) prompt plus 3–5 in-domain exemplars showing explicit fact extraction, reasoning, and scoring. The output for each fact is a tuple $\langle f, r, s \rangle$; span annotation enables visualization of supported and inconsistent content.

2.3 Quantitative Performance and Benchmarks

AXCEL’s performance is established on three tasks—summarization, free-text generation, data-to-text—using CNN/DM, QAGS, WikiBio, and RAGTruth datasets. Baseline and competitor metrics are recalibrated on Claude-3-Haiku, Claude-3-Sonnet, and Llama-3-8B for fair comparison.

Table: Metric Performance on Summarization (Spearman ρ, Claude-Sonnet)

Metric	Avg ρ
ROUGE-L	14.3
BERTScore	20.8
BARTScore	40.7
AlignScore-NLI	59.2
G-Eval	60.9
A X C E L	66.2

AXCEL outperforms AlignScore by 11.8% and G-Eval by 8.7%. On free-text generation, SelfCheckGPT-NLI scores 74.1, A X C E L (Claude-Sonnet) achieves 83.6 (+12.8% over NLI). Data-to-text ROC-AUC of A X C E L is 79.9 (+29.4% over AlignScore).

2.4 Explainability and Task Generalizability

AXCEL provides fine-grained span annotations: fact-supported sections of derived text (DT) are highlighted in green, hallucinated or contradicted spans in red, with fact-level rationales. A single, invariant CoT prompt applies across all tested tasks, with only the exemplars adapted per domain. Ablations show reasoning exemplars are essential for maximal performance (+17.1% over scores-only).

2.5 Limitations and Prospective Developments

AXCEL depends on manual exemplar annotation and LLMs themselves may hallucinate or mis-extract facts. The current design handles exactly one source and one derived text and does not support multi-context evaluations. Computational cost per DT/ST pair is higher than lightweight NLI metrics but lower than sentence-level prompt methods in some cases.

Further research directions include automatic exemplar generation, extension to zero-shot settings, multi-context consistency, and mitigation of internal verification errors.

3. Interpretative Implications

The convergence of high resolution, modular readouts in AXEL TPCs (Ban et al., 2017, Yoshida et al., 2023), and explainable, prompt-invariant LLM consistency scoring systems like AXCEL (Sreekar et al., 2024), demonstrates an increasing emphasis on scalable, robust verification—whether in fundamental physics experiments or applied NLP evaluation. This suggests a common pattern: modularized approaches enable both precision and extensibility, and embedding explicit reasoning in automated tools improves transparency and reliability.

4. Commonalities and Distinctions

Despite acronymic overlap, AXEL and AXCEL are unrelated in scope and methodology. Both share attributes of modularity, extensibility, and empirical validation against relevant standards (energy resolution and human judgment, respectively). A plausible implication is that the acronym “AXCEL” may be chosen for its association with excellence or acceleration, though this is not claimed in any source.

System	Domain	Key Metric
AXEL (TPC)	Particle Physics	Energy Resolution
AXCEL (LLMs)	NLP Evaluation	Consistency Score

• “Electroluminescence collection cell as a readout for a high energy resolution Xenon gas TPC” (Ban et al., 2017)
• “High-pressure xenon gas time projection chamber with scalable design and its performance at around the Q value of $^{136}$Xe double-beta decay” (Yoshida et al., 2023)
• “AXCEL: Automated eXplainable Consistency Evaluation using LLMs” (Sreekar et al., 2024)