Libra-Test: Cross-Domain Evaluation

• Libra-Test is a cross-domain evaluation label that encapsulates various testing protocols across astrophysics, AI, blockchain, and nuclear instrumentation.
• Its methodologies integrate calibrated benchmarks, controlled experiments, and statistical analyses to validate experimental and computational systems.
• Applications include dark matter annual modulation searches, safety checks in AI, throughput measures in blockchain, and precision nuclear readiness tests.

Libra-Test is not a single standardized protocol. Across the literature, it denotes—or is used as a natural shorthand for—multiple distinct experimental and evaluative programs associated with works carrying the name “Libra” or “LIBRA.” In some cases the term is explicitly present, as in the Chinese safety benchmark “Libra-Test”; in others, the papers state that no artifact, command, or benchmark with that exact name exists, but the reported methodology forms an identifiable testing battery or validation program. The result is a heterogeneous term spanning NaI(Tl) dark-matter modulation searches, multimodal and reward-model benchmarks, probabilistic-model evaluation workflows, blockchain performance harnesses, local-context bias assessment, radiology report-generation evaluation, and validation of nuclear-instrumentation readiness (Coarasa et al., 2024, Xu et al., 2024, Lowd et al., 2015, Zhang et al., 2019, Zhou et al., 29 Jul 2025, Pang et al., 2 Feb 2025, Chen et al., 29 Jul 2025, Zhang et al., 2024, Sun et al., 2024).

1. Terminological scope and ambiguity

The most important feature of Libra-Test is its semantic instability. In the multimodal paper “Libra: Building Decoupled Vision System on LLMs,” the authors do not explicitly introduce a protocol called “Libra-Test”; the term is instead a natural interpretation of the paper’s “full evaluation battery,” including zero-shot image-to-text tasks, multimodal LLM benchmarks, visual sequential modeling diagnostics, attention-pattern analyses, and ablation studies (Xu et al., 2024). The same negative formulation appears in “The Libra Toolkit for Probabilistic Models,” which states that there is no component, script, or command explicitly named “Libra-Test,” and that testing is performed by running learning and inference commands on held-out data and querying marginals, MAP/MPE, and probabilities of evidence (Lowd et al., 2015).

A similar clarification appears in “Libra: Assessing and Improving Reward Model by Learning to Think,” where the paper does not use the term “Libra-Test,” and the relevant testing component is instead the benchmark called “Libra Bench” (Zhou et al., 29 Jul 2025). By contrast, “Libra: Large Chinese-based Safeguard for AI Content” explicitly introduces Libra-Test as “the first benchmark specifically designed to evaluate the effectiveness of safeguard systems for Chinese content” (Chen et al., 29 Jul 2025).

This distribution of usage suggests that Libra-Test functions less as a stable proper noun than as a family resemblance term. In some fields it denotes a named benchmark; in others it refers to a reproducible evaluation methodology associated with a Libra-branded system.

2. NaI(Tl) annual-modulation searches and the DAMA/LIBRA controversy

In astroparticle physics, Libra-Test most directly refers to the long-running annual-modulation program centered on DAMA/LIBRA and its same-target replications. DAMA/LIBRA searches for the model-independent dark-matter annual modulation signature in NaI(Tl) scintillators using the standard time-dependent form

$R(t,E)=S_0(E)+S_m(E)\cos[\omega (t-t_0)],$

with $T \approx 1$ year and $t_0 \approx 152.5$ days, near June 2 (Bernabei et al., 2010, Bernabei et al., 2022). The signal is required to be cosine-like, annual, phase-peaked near early June, confined to low energies, present in single-hit events only, and absent in multiple-hit events and higher-energy control regions (Bernabei et al., 2014).

DAMA/LIBRA reported, for the cumulative 1.17 ton$\times$yr exposure over 13 annual cycles, a modulation amplitude in 2–6 keV of $(0.0116 \pm 0.0013)$ cpd/kg/keV, a phase of $(146 \pm 7)$ days, and a period of $(0.999 \pm 0.002)$ yr, corresponding to 8.9$\sigma$ confidence level (Bernabei et al., 2010). In a later summary over 22 annual cycles and 2.86 ton $\times$ yr exposure, the free-period/free-phase fit in 2–6 keV gave a modulation amplitude of $(0.01014 \pm 0.00074)$ cpd/kg/keV, phase $T \approx 1$0 days, and period $T \approx 1$1 yr, with 13.7$T \approx 1$2 confidence level (Bernabei et al., 2022).

The central methodological response to DAMA/LIBRA has been same-material replication. ANAIS-112 uses 112.5 kg of NaI(Tl) in nine 12.5 kg modules at the Canfranc Underground Laboratory and was built explicitly to perform a model-independent negative test of the DAMA/LIBRA claim (Coarasa et al., 2024). Its three-year reanalysis, based on 312.53 kg$T \approx 1$3yr effective exposure, incorporated an upgraded analysis chain with low-energy calibration anchored to internal coincidence-tagged lines at 0.90 keVee and 3.2 keVee, improved saturation linearization, a Boosted Decision Tree trained on bulk nuclear recoils from external $T \approx 1$4Cf neutron calibrations, a trigger-rate cut, and explicit time-dependent background modeling in the modulation fit (Coarasa et al., 2024). In the 1–6 keVee window it found $T \approx 1$5 counts keV$T \approx 1$6 ton$T \approx 1$7 d$T \approx 1$8 and reported incompatibility with DAMA/LIBRA at 3.7$T \approx 1$9 confidence level in its dedicated compatibility test; in 2–6 keVee it found $t_0 \approx 152.5$0 counts keV$t_0 \approx 152.5$1 ton$t_0 \approx 152.5$2 d$t_0 \approx 152.5$3 and 2.6$t_0 \approx 152.5$4 incompatibility (Coarasa et al., 2024).

The six-year ANAIS-112 analysis strengthened that result. With 625.75 kg·yr effective exposure, ANAIS-112 reported no modulation and found its data incompatible with DAMA/LIBRA at 4$t_0 \approx 152.5$5 in 1–6 keVee, 3.5$t_0 \approx 152.5$6 in 2–6 keVee, and 4.2$t_0 \approx 152.5$7 in an equivalent nuclear-recoil window under constant quenching-factor assumptions (Amaré et al., 3 Feb 2025). The fit explicitly included detector-wise time-dependent backgrounds through

$t_0 \approx 152.5$8

with nuisance parameters per detector and fixed $t_0 \approx 152.5$9 year, $\times$0 June 2 for direct comparison to DAMA (Amaré et al., 3 Feb 2025).

COSINE-100 performs an analogous same-target test at Yangyang Underground Laboratory with NaI(Tl) crystals and an active liquid-scintillator veto. In its 6.4-year analysis, using 61.3 kg effective mass and 358 kg⋅years exposure, it adopted DAMA-like linear keVee calibration, a 0.7 keV threshold, and a time-dependent background model

$\times$1

finding no evidence of an annual modulation and disfavoring DAMA at greater than $\times$2 in both electron-recoil and sodium nuclear-recoil interpretations (Carlin et al., 2024).

Independent statistical work has separated the question of whether DAMA’s residuals are modulated from the question of whether the modulation is due to dark matter. “Robust model comparison tests of DAMA/LIBRA annual modulation” found that, within DAMA’s published residual-rate data, a cosine model is decisively favored over a constant model by frequentist, information-theoretic, and Bayesian criteria, with Bayes factors up to $\times$3–$\times$4 depending on the binning (Krishak et al., 2019). A separate conventional-hypothesis paper proposed that muon-induced delayed scintillation in NaI(Tl) could mimic the low-energy annual modulation and suggested direct time-correlation searches between muon events and delayed few-keV pulses, as well as low-temperature operation to suppress thermally activated delayed emission (Nygren, 2011).

A frequent misconception is therefore that the dark-matter “Libra-Test” has a single accepted meaning. In practice it includes at least three distinct layers: DAMA/LIBRA’s own modulation search, statistical tests of whether the residuals favor a cosine over a constant, and same-target replication efforts designed to validate or refute a dark-matter interpretation.

3. AI and machine-learning benchmark suites

In machine learning, Libra-Test usually designates an evaluation battery rather than a detector experiment. The decoupled-vision multimodal model “Libra” is evaluated through a broad image-to-text benchmark suite that includes general VQA, image captioning, MLLM-oriented benchmarks, visual sequential modeling diagnostics, attention-pattern analyses, and ablations (Xu et al., 2024). The architecture integrates a routed visual expert and a cross-modal bridge into a pretrained LLaMA2-7B-Chat backbone, uses discrete autoregressive modeling over image and text tokens, and is evaluated with greedy decoding and dataset-specific prompts on VQAv2, OKVQA, GQA, VizWiz, ScienceQA, NoCaps, Flickr, COCO Karpathy, POPE, MME, MMB, SEED, MM-Vet, and MMVP (Xu et al., 2024). Within that testing regime, the paper reports, for example, 77.3% on VQAv2, 59.7% on OKVQA, 123.8 CIDEr on NoCaps, and 1494.7 on MME (Xu et al., 2024).

For reward modeling, Libra Bench is the explicit testing artifact. It is a reasoning-oriented benchmark for pointwise judging accuracy on difficult mathematical reasoning outputs, built from MATH-500 level 5, AIME 2024, and AIME 2025, with responses sampled from DeepSeek-R1, Qwen3-32B, QwQ-32B, DeepSeek-R1-Distill-Qwen-7B, and DeepSeek-R1-Distill-Qwen-1.5B (Zhou et al., 29 Jul 2025). The benchmark totals 3,740 samples from 204 problems and 5 models, is balanced between correct and incorrect samples within each subset, and evaluates judges by accuracy per subset and average over subsets (Zhou et al., 29 Jul 2025). On this benchmark, Libra-RM-32B-MATH achieves 83.4 on MATH-500, 81.5 on AIME 2024, 80.3 on AIME 2025, and 81.7 average accuracy (Zhou et al., 29 Jul 2025).

The term also appears in bias and safety evaluation. The Local Integrated Bias Recognition and Assessment framework, abbreviated LIBRA, measures bias from a local cultural context using datasets sourced from local corpora rather than crowdsourcing and introduces the Enhanced Idealized CAT Score,

$\times$5

to combine bias measurement with a beyond-knowledge-boundary score for local vocabulary (Pang et al., 2 Feb 2025). The framework is instantiated in New Zealand, where the paper reports a dataset comprising over 360,000 test cases in the New Zealand context and emphasizes that local words such as “karani” can be misunderstood by GPT-2 and Llama-3, contaminating standard bias estimates if knowledge-boundary effects are not modeled (Pang et al., 2 Feb 2025).

The Chinese safety benchmark “Libra-Test” is more narrowly defined. It contains 5,720 single-turn query–response pairs, with 1,864 safe and 3,856 unsafe items, grouped under seven harm scenarios: Physical and Mental Health, Privacy and Property, Ethics and Morality, Bias and Discrimination, Illegal Activities and Crime, Hate Speech and Abuse, and Rumors and Misinformation (Chen et al., 29 Jul 2025). Each sample is labeled independently by three human annotators under unified safety rules, with final confirmation by a safety expert (Chen et al., 29 Jul 2025). The benchmark is used to evaluate guard models by Accuracy and per-class F1, and Libra-Guard reaches 86.79% accuracy, outperforming Qwen2.5-14B-Instruct at 74.33% and ShieldLM-Qwen-14B-Chat at 65.69% (Chen et al., 29 Jul 2025).

A further AI use appears in biomedical radiology. The temporal-aware multimodal model “Libra” for chest X-ray report generation is evaluated on paired current/prior images using lexical metrics such as ROUGE-L, BLEU-1/4, and METEOR, clinical metrics such as RadGraph-F1, RGER, CheXpert F1, CheXbert vector similarity, and RadCliQ version 0, and a temporal metric, Temporal Entity F1 (Zhang et al., 2024). On the 2,461-sample frontal-view test set, Libra reports ROUGE-L 36.7, BLEU-1 51.3, BLEU-4 24.5, METEOR 48.9, RadGraph-F1 32.9, RGER 37.6, CheXbert vector 46.9, and RadCliQ0 2.7 (Zhang et al., 2024).

4. Toolkit workflows, systems benchmarking, and apparatus validation

Outside benchmark-centric AI, Libra-Test often means an operational workflow for validating a system. In the Libra Toolkit for probabilistic models, testing proceeds by learning a model, compiling tractable representations when appropriate, and querying held-out evidence through consistent command-line programs (Lowd et al., 2015). A canonical workflow is $(0.0116 \pm 0.0013)$8 which supports exact conditional marginals on a compiled arithmetic circuit and approximate inference directly on the Bayesian network for comparison (Lowd et al., 2015). The toolkit spans Bayesian networks, Markov networks, dependency networks, and sum-product networks, supports exact inference via arithmetic circuit variable elimination and exact marginal/MAP inference on ACs and SPNs, and includes Gibbs sampling, loopy belief propagation, mean field, max-product, and iterated conditional modes for approximate evaluation (Lowd et al., 2015).

In blockchain systems, Libra-Test refers to a throughput and latency benchmark harness for the pre-Diem Libra blockchain. The experimental study deploys Libra in release mode through Libra_swarm on Ubuntu 18.04, uses 12 concurrent clients, varies validators over $\times$6, and sweeps workloads from 1,000 to 50,000 transactions (Zhang et al., 2019). Throughput is computed as

$\times$7

using sender sequence numbers and local timestamps (Zhang et al., 2019). The study reports that the Libra blockchain can only process about one thousand transactions per second at most, with measured values of 1030.7 TPS at 1 validator for P2P transfers, 643.82 TPS at 4 validators for P2P transfers, 744.22 TPS at 4 validators for a Do-nothing script, and about 350 TPS at 16 validators under heavy load (Zhang et al., 2019). The dominant bottleneck is consensus rather than execution (Zhang et al., 2019).

In experimental nuclear physics, LIBRA denotes the Lifetimes and Branching Ratios Apparatus at FRIB, and the associated Libra-Test is its readiness-validation program. The instrument extends the particle x-ray coincidence technique to measure average lifetimes in the $\times$8–$\times$9 s range and decay branching ratios of resonances populated by EC/$(0.0116 \pm 0.0013)$0 decay (Sun et al., 2024). The validation program combines theoretical shell-model calculations, Geant4 Monte Carlo simulations, and source-performance tests (Sun et al., 2024). Measured performance includes LEGe X-ray resolution of about 0.24 keV FWHM at Cu/Zn $(0.0116 \pm 0.0013)$1, X-ray efficiency around 7% at 8–9 keV, XtRa $(0.0116 \pm 0.0013)$2 efficiencies of about 0.3% at 1 MeV, and successful delayed-coincidence extraction of the 237Np 59.5-keV state half-life at approximately 68 ns (Sun et al., 2024). In the simulated 60Ga case, the proton-gated Cu/Zn $(0.0116 \pm 0.0013)$3 ratio yields an inferred average proton-decay lifetime of $(0.0116 \pm 0.0013)$4 fs (Sun et al., 2024).

5. Recurring methodological structure

A plausible synthesis is that Libra-Test is best understood as a “validation bundle” (Editor’s term): a reproducible combination of data selection, calibration or prompting, metric definition, nuisance control, and comparative inference.

In the dark-matter literature, the validation bundle is anchored by same-target material, explicit time-dependent background modeling, and direct comparison of modulation amplitudes under fixed period and phase (Coarasa et al., 2024, Amaré et al., 3 Feb 2025, Carlin et al., 2024). In reward modeling, it is anchored by balanced correct/incorrect subsets, pointwise correctness labels, and per-subset accuracy rather than pairwise preference (Zhou et al., 29 Jul 2025). In Chinese safeguard evaluation, it is anchored by unified safety rules, multi-annotator adjudication, and response-level safe/unsafe classification with Accuracy and per-class F1 (Chen et al., 29 Jul 2025). In local-context bias measurement, it is anchored by triplet construction, distributional comparison using Jensen–Shannon divergence, and explicit modeling of beyond-knowledge-boundary vocabulary through $(0.0116 \pm 0.0013)$5 and EiCAT (Pang et al., 2 Feb 2025). In the blockchain study, it is anchored by workload sweeps, validator-count sweeps, and layer-isolation scripts that distinguish consensus from execution costs (Zhang et al., 2019). In the probabilistic-model toolkit, it is anchored by held-out evidence queries and exact-versus-approximate inference comparisons (Lowd et al., 2015).

This suggests a deeper commonality: Libra-Test rarely names a single metric. It usually denotes an evaluation design in which the object under test is decomposed into interpretable subsystems—background evolution versus modulation amplitude, inner-modal versus cross-modal attention, consensus versus execution, knowledge failure versus bias, correctness judgment versus free-form reasoning—and then reassembled into a comparative report.

6. Limitations, controversies, and future directions

The first limitation is terminological. Several papers explicitly state that no artifact called Libra-Test exists in their system, so any encyclopedia treatment must preserve that distinction rather than retroactively standardizing the name (Xu et al., 2024, Lowd et al., 2015, Zhou et al., 29 Jul 2025). The term is therefore descriptive, not canonical, in a substantial part of the literature.

The second limitation is domain-specific controversy. In NaI(Tl) dark-matter searches, the principal dispute is no longer whether DAMA’s published residuals are modulated—they are statistically well described by an annual cosine—but whether that modulation is attributable to dark matter or to an unmodeled detector or environmental effect (Krishak et al., 2019, Nygren, 2011). ANAIS-112 and COSINE-100 now provide same-target null results that are incompatible with DAMA at roughly the $(0.0116 \pm 0.0013)$6–$(0.0116 \pm 0.0013)$7 level, but ANAIS emphasizes that quenching factors, non-proportionality, and related nuclear-recoil response systematics remain important in detailed cross-experiment comparisons (Amaré et al., 3 Feb 2025).

The third limitation concerns benchmark scope. Libra Bench is confined to mathematical reasoning; its own discussion identifies extension beyond math as future work (Zhou et al., 29 Jul 2025). The local-context LIBRA framework depends on definition verification for nonstandard vocabulary and notes that JSD values are not directly comparable across masked and causal LLMs (Pang et al., 2 Feb 2025). The Chinese Libra-Test benchmark is single-turn and does not report multi-turn safety dynamics, inter-annotator agreement statistics, or per-scenario confusion matrices (Chen et al., 29 Jul 2025). The multimodal Libra model inherits LLM weaknesses, including hallucinations and limited support for routed attention in common acceleration libraries (Xu et al., 2024). The radiology Libra evaluation is limited to frontal chest X-rays on MIMIC-CXR and does not yet integrate multiple priors, prior reports, or broader longitudinal EHR context (Zhang et al., 2024). The blockchain benchmark was conducted in a LAN, apparently on a single physical server, and does not report CPU, memory, I/O, or tail-latency statistics (Zhang et al., 2019). The nuclear-instrumentation LIBRA program has demonstrated feasibility and readiness, but its astrophysical impact still depends on the forthcoming 60Ga beam experiment (Sun et al., 2024).

Taken together, these uses show that Libra-Test is best treated as a cross-domain label for rigorous empirical adjudication. In one branch of the literature it is a model-independent test of a long-standing dark-matter anomaly; in another it is a benchmark suite for judging, safety, bias, multimodal understanding, or temporal medical reporting; in a third it is a practical validation workflow for inference engines, blockchains, or nuclear apparatus. The unifying theme is not the object being tested, but the insistence on explicit protocols, controlled comparisons, and quantitatively reported uncertainty.