ALICE: Multidisciplinary Research Insights

• ALICE is a multifaceted term denoting a heavy-ion experiment at CERN’s LHC, sophisticated detector upgrades, automated eFPGA redaction flows, diverse machine learning methods, and an archival astronomy program.
• The heavy-ion experiment employs precision tracking and PID technologies, achieving key results such as low pT tracking performance, D-meson suppression (RAA≈0.2), and enhanced vertex resolution through ITS upgrades.
• Advanced computing frameworks like O² and Hyperloop leverage GPU acceleration for real-time processing, while ALICE’s ML and archival imaging initiatives optimize adversarial training and high-contrast detection techniques.

Across the cited literature, ALICE denotes several unrelated research entities. The dominant usage is A Large Ion Collider Experiment, the dedicated heavy-ion detector at the CERN Large Hadron Collider; the same acronym is also used for a Run 3 analysis-train infrastructure within that experiment, an automated flow for eFPGA redaction, multiple machine-learning frameworks, and an archival high-contrast-imaging program. The shared acronym therefore designates a family of technically distinct systems rather than a single lineage (Schukraft, 2011, Quishpe et al., 2021, Tomajoli et al., 2022, Liang et al., 2020, Piao et al., 20 Mar 2026, Choquet et al., 2015).

1. ALICE as a heavy-ion experiment at the LHC

ALICE, expanded as A Large Ion Collider Experiment, is the dedicated heavy-ion detector at the CERN LHC. Its primary goal is to study the properties of strongly interacting matter under extreme energy densities, where a deconfined Quark–Gluon Plasma is expected to form. The experiment addresses collective phenomena, parton energy loss in the medium, heavy-flavour and quarkonium production and suppression, and bulk thermodynamic properties such as temperature and chemical potentials (Schukraft, 2011).

The detector architecture is optimized for very high-multiplicity heavy-ion collisions while preserving broad low-momentum reach. The central barrel, embedded in a 0.5 T solenoidal field, comprises the Inner Tracking System (ITS), Time Projection Chamber (TPC), Transition Radiation Detector (TRD), Time-Of-Flight (TOF), High-Momentum Particle Identification Detector (HMPID), and the PHOS and EMCAL calorimeters; a forward muon spectrometer covers quarkonia and heavy-flavour decays at forward rapidity. This arrangement gives ALICE excellent charged-particle tracking down to very low transverse momentum and powerful, redundant PID over a broad momentum range (Schukraft, 2010, Zampolli, 2012).

The first Pb–Pb campaign at the LHC established the experiment’s heavy-ion performance at unprecedented collision energy. In the first Pb–Pb run at $\sqrt{s_{NN}}=2.76\ \mathrm{TeV}$, ALICE recorded $\simeq30$ million minimum-bias Pb–Pb collisions. In the 0–5% most central events, the charged-particle density reached $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$. The inclusive charged-hadron nuclear modification factor showed a minimum of $\simeq0.14$–$0.2$ at $p_T\approx5$–$7\ \mathrm{GeV}/c$, then rose toward higher $p_T$; integrated $v_2$ increased by $\sim30\%$ from RHIC to LHC in mid-central collisions; and prompt D-meson suppression in Pb–Pb reached $\simeq30$0–$\simeq30$1, while forward-rapidity $\simeq30$2 yielded $\simeq30$3, flat versus centrality (Schukraft, 2011).

A central feature of ALICE physics is precision PID. The experiment combines specific energy loss in the ITS and TPC, time-of-flight, transition radiation, Cherenkov imaging, and electromagnetic calorimetry. The PID program spans charged hadrons, leptons, photons, and light nuclei, with the TPC achieving $\simeq30$4 $\simeq30$5 resolution for tracks with 159 clusters, TOF delivering $\simeq30$6 time resolution, and HMPID extending $\simeq30$7 separation to several GeV/$\simeq30$8 (Zampolli, 2012).

ALICE measurements also played a central role in the discussion of collectivity outside large nuclei. Pb–Pb results established the simultaneous presence of strong high-$\simeq30$9 suppression, mass-ordered anisotropic flow, and hydrodynamic behavior, while high-multiplicity p–Pb collisions revealed a “double ridge,” nonzero $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$0 and $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$1, and identified-hadron mass ordering similar to Pb–Pb. This supports the view that the acronym ALICE, in high-energy physics, is inseparable from QGP phenomenology, rare-probe reconstruction, and low-$dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$2 precision (Räsänen, 2016).

2. Tracking, alignment, and detector-upgrade program

The original ALICE Inner Tracking System was a six-layer silicon tracker composed of two layers of Silicon Pixel Detectors, two layers of Silicon Drift Detectors, and two layers of Silicon Strip Detectors. Its spatial alignment involved 2198 sensor modules and about 13,000 alignment parameters, with target precision well below 10 micron in some cases. The main track-based alignment method used the Millepede global approach, supplemented by an iterative local method; with about $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$3 charged tracks from cosmic rays, residual misalignments of order 7–20 $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$4m in the most precise directions were achieved (Collaboration, 2010).

This alignment program was foundational for early tracking and vertexing performance. In the commissioning and early-physics period, ALICE used cosmics and pp data to bring the ITS subsystems to a common geometry at the few–10 $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$5m level, verify the $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$6 material budget by $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$7-conversion tomography, and demonstrate impact-parameter and primary-vertex resolutions close to design. That performance enabled reconstruction of strange-baryon cascades and open-charm decays in early LHC data, together with $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$8-based PID down to $dN_{\mathrm{ch}}/d\eta|_{\eta=0}\simeq1600$9 (Rossi, 2011).

For Runs 3 and 4, ALICE replaced the old ITS with a seven-layer, all-pixel detector based on Monolithic Active Pixel Sensors (MAPS). The new layout consists of three Inner Barrel layers and four Outer Barrel layers. The adoption of ALPIDE chips, with nominal pixel pitch $\simeq0.14$0, reduced the material budget to $\simeq0.14$1 per Inner Barrel layer and $\simeq0.14$2 per Outer Barrel layer. The upgrade targeted factor 3–5 improvements in impact-parameter resolution, tracking efficiency $\simeq0.14$3 down to $\simeq0.14$4, and continuous readout up to 50 kHz in Pb–Pb (Colella, 2019).

Laboratory commissioning of ITS2 established the readiness of the new detector for continuous readout. The detector geometry used radii beginning at 22.4 mm, ALPIDE chips in TowerJazz 180 nm CMOS, and sensor thicknesses of 50 $\simeq0.14$5m in the Inner Barrel and 100 $\simeq0.14$6m in the Outer Barrel. At a threshold of 100 $\simeq0.14$7, masking fewer than 0.009 % of pixels in the Inner Barrel yielded a fake-hit rate of $\simeq0.14$8 hits/pixel/event; bit-error-rate tests showed sensitivities down to $\simeq0.14$9 at 44.9 kHz and $0.2$0 at 247 kHz (Colella, 2020).

The LS2 upgrade generalized this detector renewal to the full experiment. ALICE moved to continuous, dead-time-free readout at rates up to 50 kHz in Pb–Pb and 1 MHz in pp collisions, deployed ITS2, MFT, and a 4-GEM TPC for continuous operation, and integrated these systems with a new online–offline computing farm. In that summary, the ITS2 delivered a pointing-resolution improvement of $0.2$1 in the transverse plane and $0.2$2 in the longitudinal direction at 500 MeV/$0.2$3 (Collaboration, 2023).

Beyond Run 3, ALICE is pursuing further detector evolution. ITS3 replaces the three innermost ITS2 layers with bent monolithic pixel sensors in 65 nm CMOS, targeting $0.2$4 per layer and a factor-of-2 improvement in pointing resolution. The proposed FoCal extends coverage to $0.2$5 for prompt-photon and jet measurements at small $0.2$6, while the ALICE 3 concept envisages a next-generation compact detector with 11 barrel layers, extended PID, and high readout rates for Run 5 and beyond (Ferretti, 2022).

3. Run 3 computing, O$0.2$7, and the Hyperloop train system

Run 3 required ALICE to redesign both event processing and organized analysis. The experiment adopted a two-phase online–offline reconstruction strategy within the Online-Offline Computing System (O$0.2$8): a synchronous stage runs in real time during data taking and performs detector calibration, global TPC track finding, and real-time compression; an asynchronous stage reprocesses the buffered compressed data with final calibrations and full multi-detector reconstruction (Rohr, 2021).

The underlying throughput requirements are unusually severe. For 50 kHz continuous readout of minimum-bias Pb–Pb collisions, each 10 ms time frame contains 500 collisions. The expected raw TPC data rate is $0.2$9, corresponding to $p_T\approx5$0 per event before compression. The synchronous chain targets $p_T\approx5$1, a compression factor of $p_T\approx5$2, achieved through zero suppression, hit selection, predictive coding, and Asymmetric Numeral Systems (ANS) entropy encoding (Rohr, 2021).

GPU acceleration is central to this model. The synchronous farm design uses $p_T\approx5$3 servers, each with $p_T\approx5$4-core CPUs and 8 GPUs, for a total of $p_T\approx5$5 GPUs and $p_T\approx5$6 CPU cores. Full TPC reconstruction—clusterization, tracking, and compression—is offloaded to the GPU, with measured speedups of $p_T\approx5$7–$p_T\approx5$8 relative to one 3.3 GHz AMD Rome core depending on GPU model. In scaling tests, NVIDIA V100, NVIDIA A100, and AMD MI100 all met the 50 kHz Pb–Pb requirement with a 20% margin (Rohr, 2021).

The LS2 system report recasts this into a full experiment-wide architecture. Detector data are grouped into Heartbeat Frames and Time Frames, then transported through a unified O$p_T\approx5$9 software stack with three layers: a FairMQ-based Transport Layer, a Data Model Layer, and a Data Processing Layer (DPL). The online farm consists of 199 FLP nodes and 280 EPN nodes; the TPC path reduces throughput from $7\ \mathrm{GeV}/c$0 raw to $7\ \mathrm{GeV}/c$1 at the FLP→EPN stage and $7\ \mathrm{GeV}/c$2 for final Compressed Time Frames (CTF) (Collaboration, 2023).

Analysis organization underwent a parallel redesign through Hyperloop, the successor to the LEGO trains. In Runs 1 and 2, user analysis tasks—“wagons”—acting on the same dataset were combined into a single train, which a train operator assembled, tested, submitted once to the Grid, and then merged. By 2020, 90 % of all ALICE analyses ran as LEGO trains, yielding 16 000 trains and $7\ \mathrm{GeV}/c$3 Grid jobs in a single year. Hyperloop preserves the single-pass-over-data principle while integrating with the O$7\ \mathrm{GeV}/c$4 analysis framework, MonALISA, LPM, and ALICE Analysis Facilities (Quishpe et al., 2021).

Hyperloop adds a React.js-based web UI, a Java model, a PostgreSQL database for bookkeeping, instantaneous automatic testing, and the production of derived skimmed datasets. Analyzers define an “analysis” in JIRA, wagon tests are dispatched within minutes, successful tests feed an automatic train-composition algorithm, and skimmed outputs are staged to Analysis Facilities for interactive or high-throughput use. The system had already launched $7\ \mathrm{GeV}/c$5 Hyperloop trains on converted Run 2 data, and early results showed a 3–10$7\ \mathrm{GeV}/c$6 improvement in event throughput when comparing AliPhysics (LEGO) with O$7\ \mathrm{GeV}/c$7 (Hyperloop runs) (Quishpe et al., 2021).

4. ALICE in electronic design automation: automatic eFPGA redaction

In hardware security and EDA, ALICE denotes “ALICE: An Automatic Design Flow for eFPGA Redaction.” The flow takes an RTL design in Verilog, a set of protected outputs, and eFPGA architectural parameters, then produces a fabric-redacted RTL in which selected RTL modules are replaced by one or more custom embedded FPGAs. The objective is to protect the intellectual property of hardware designs when fabrication is outsourced to a third-party foundry (Tomajoli et al., 2022).

The methodology has three stages. Module Filtering identifies all RTL modules that influence the chosen outputs and removes those whose I/O or resource requirements exceed the specified eFPGA parameters. Cluster Identification forms all combinable subsets whose aggregate I/O pins and LUT/CLB counts fit within a single eFPGA specification. eFPGA Selection and Integration generates candidate fabrics through a customization tool such as OpenFPGA, scores them, solves a small “knapsack-like” selection problem via branch-and-bound, and rewrites the top-level RTL to instantiate each selected eFPGA (Tomajoli et al., 2022).

The formal problem is posed over modules $7\ \mathrm{GeV}/c$8, with per-module I/O demand $7\ \mathrm{GeV}/c$9 and estimated CLB usage $p_T$0, constrained by $p_T$1. ALICE seeks up to $p_T$2 disjoint clusters satisfying

$p_T$3

while maximizing

$p_T$4

with

$p_T$5

and

$p_T$6

The disjointness constraint enforces that no module is redacted more than once, and every selected module must lie on some data-flow path to one of the protected outputs (Tomajoli et al., 2022).

Fabric generation and physical integration are treated as first-class design tasks. ALICE uses custom fabrics with 4-input fracturable LUTs, 4 LUTs per CLB, local flip-flops, routing resources, and I/O tiles delivering up to 8 GPIO pins each. A dominator-tree analysis identifies the nearest common parent for multi-module clusters, thereby minimizing net lengths when the top ASIC module instantiates each eFPGA macro and rewires the original nets to GPIOs (Tomajoli et al., 2022).

The evaluation uses benchmarks from the CEP, IWLS05, and OpenROAD suites, including DES3, FIR, IIR, SHA256, SASC, USB_PHY, and GCD. Two sample configurations are reported: cfg1: $p_T$7, $p_T$8 and cfg2: $p_T$9, $v_2$0. For DES3, cfg1 found two 8×8 fabrics covering 4 modules, whereas cfg2 found one 14×14 fabric covering 8 modules; for GCD, cfg1 yielded two 4×4 fabrics redacting 2 modules with 52,629 $v_2$1m$v_2$2 total eFPGA area, while cfg2 yielded one 5×5 fabric redacting 3 modules with 54,512 $v_2$3m$v_2$4 (Tomajoli et al., 2022).

The reported limitations are also explicit: all eFPGA instances currently share identical architectural parameters, granularity is restricted to the RTL-module level, security scoring is tied to utilization, the prototype accepts only Verilog, and future work includes heterogeneous fabrics, fine-grained sub-module partitioning, richer security metrics, and co-optimization of fabric size and module selection (Tomajoli et al., 2022).

5. ALICE in machine learning and multimodal evaluation

The acronym ALICE is reused for several unrelated machine-learning frameworks. In “ALICE: Active Learning with Contrastive Natural Language Explanations,” it denotes an expert-in-the-loop training framework that uses active learning to select informative class-pair queries, collects both binary labels and contrastive natural-language explanations, parses those explanations into symbolic constraints, and injects the resulting knowledge into a neural classifier through an explanation-conditioning mechanism. Applied to bird species classification and social relationship classification, the framework outperformed baseline models trained with 40–100% more training data, and adding 1 explanation produced a performance gain similar to 13–30 labeled training data points (Liang et al., 2020).

A second ML usage appears in “Adversarial Training for Commonsense Inference,” where ALICE expands to AdversariaL training for Commonsense InferenCE. This method augments ordinary supervised learning with two embedding-space perturbation terms: one based on the true label, and one based on the model prediction. The combined objective minimizes a supervised adversarial loss plus a virtual-label adversarial loss, both approximated by one-step $v_2$5 perturbations. Fine-tuned RoBERTa_large models evaluated on CosmosQA, MCScript2.0, and MC-TACO consistently outperformed standard fine-tuning as well as single-term adversarial baselines (ADV and SMART); on the test set, the reported ALICE scores were 84.57% on CosmosQA, 92.5% and 93.5% on the commonsense and out-of-domain MCScript2.0 settings, and 56.45% EM / 79.50% F1 on MC-TACO (Pereira et al., 2020).

A third, more recent usage is “ALICE: A Multifaceted Evaluation Framework of Large Audio-LLMs’ In-Context Learning Ability.” Here ALICE is a three-stage framework that progressively removes textual guidance under audio conditioning. Stage 1 retains task description and explicit format instruction; Stage 2 removes the explicit format instruction; Stage 3 removes the task description as well, leaving only audio inputs paired with correctly formatted outputs. The framework evaluates six LALMs—Qwen2-Audio, DeSTA2.5-Audio, BLSP-Emo, Qwen2.5-Omni, Phi-4-Multimodal, and Gemini 2.5 Flash (On/Off)—on four tasks: ASR, SER, GR, and MMAU, under two output-constraint families: Closed-Ended Questions (CEQ) and Chain-of-Thought (CoT) (Piao et al., 20 Mar 2026).

The key metrics are Format Compliance Rate (FCR) and core task performance:

$v_2$6

For ASR, performance is measured by WER; for SER, GR, and MMAU, by Accuracy. The central empirical result is a consistent asymmetry: in-context demonstrations improve format compliance but do not improve, and often degrade, core task performance. In Stage 1, even 1-shot examples substantially boost FCR for weaker models; in Stage 2, FCR drops by 7–29 pp when format instructions are removed; and in Stage 3, FCR may rebound while task performance degrades sharply, especially in CEQ (Piao et al., 20 Mar 2026).

Taken together, these ML uses show that the acronym ALICE does not identify one technique family. It names an active-learning framework, an adversarial regularizer, and an evaluation protocol for LALMs, all methodologically independent.

6. ALICE in astronomical archive mining

In astronomy, ALICE expands to Archival Legacy Investigations of Circumstellar Environments. This is an HST Archival Research program that re-analyzes the NICMOS coronagraphic archive, comprising roughly 400 stars observed between 1997 and 2008 with the NIC2 channel, to search for previously undetected debris disks and faint companions (Choquet et al., 2015).

The program uses modern high-contrast post-processing in a Reference-star Differential Imaging (RDI) framework, specifically LOCI and KLIP/PCA, to surpass first-generation analyses in contrast and inner working angle. In LOCI, a linear combination of reference PSFs is optimized locally to minimize speckle residuals; in KLIP, the covariance of a PSF library is diagonalized and the first $v_2$7 principal components are used to model and subtract the stellar PSF (Choquet et al., 2015).

Candidate identification is based on PSF-subtracted residuals cross-correlated with a Tiny-TIM synthetic NICMOS PSF. A source is logged if it appears in both roll images with consistent astrometry and photometry. The raw signal is the cross-correlation peak, while the noise is measured in an annulus of radii 6–10 pixels; the resulting detection statistic is

$v_2$8

ALICE keeps even moderate-SNR sources to maximize completeness, then assesses confidence and completeness through injection–recovery tests framed as hypothesis tests under $v_2$9 and $\sim30\%$0 (Choquet et al., 2015).

The statistical treatment relies on false-alarm rate, completeness, ROC curves, and AUC. By repeating injections over contrast–separation grids, ALICE derives completeness maps in $\sim30\%$1 space. The program reports that it has reprocessed $\sim30\%$2 of the NICMOS exoplanet-program targets in three large surveys, recovering 237 point-source detections in those surveys and 304 across the entire archive. Of these, $\sim30\%$3 are high-SNR detections with SNR $\sim30\%$4, whereas many candidates lie at SNR $\sim30\%$5–$\sim30\%$6, separations 0.3–3″, and contrasts $\sim30\%$7–15 (Choquet et al., 2015).

The reported trade-off between reliability and completeness is explicit. A detection threshold of SNR = 3 yields false-alarm rates of only a few percent while delivering completeness $\sim30\%$8 for sources drawn from the sample. The AUC approaches unity at wide separations and low contrasts, but drops toward random performance for $\sim30\%$9 at $\simeq30$00. In this sense, the astronomical ALICE is a statistically explicit archival survey infrastructure rather than a detector, learning algorithm, or hardware-security flow (Choquet et al., 2015).