S-Score: Metrics, Models, and Methods

Updated 7 July 2026

S-Score is a multi-domain term that encompasses benchmark metrics, evaluation frameworks, and statistical score functions used in different research fields.
In CEScore, Sscore quantifies text simplicity by comparing source and simplified texts through hand-designed measures like TSS, SLS, and ASF, aligning well with human judgments.
Other usages include the SCORE framework for LLM robustness, SzCORE for seizure detection, and traditional score functions in statistical inference and generative models.

Across recent arXiv literature, “S-Score” is not a single standardized statistic. The label appears explicitly as the simplicity component of CEScore for split-and-rephrase, while closely related names such as SCORE, SzCORE, SeMaScore, and Server Quality Score denote distinct constructs for LLM robustness evaluation, EEG seizure benchmarking, ASR evaluation, and tennis analytics; in other papers, “score” refers instead to a likelihood score or density score function (Ajlouni et al., 2023, Nalbandyan et al., 28 Feb 2025, Dan et al., 19 May 2025, Sasindran et al., 2024, Li et al., 8 Feb 2026, Wibisono et al., 2024). The term therefore requires domain-specific disambiguation: some usages are benchmark metrics, some are framework names, and some are statistical objects defined by derivatives of log-densities or log-likelihoods.

1. Terminological scope

Representative usages of “S-Score” and closely related terms in recent work are heterogeneous rather than interchangeable (Ajlouni et al., 2023, Nalbandyan et al., 28 Feb 2025, Dan et al., 19 May 2025, Sasindran et al., 2024, Li et al., 8 Feb 2026).

Term	Domain	Operational role
Sscore	Split-and-rephrase	Simplicity score in $[0,1]$
SCORE	LLM evaluation	Robustness framework reporting mean accuracy, $[min,max]$ , and consistency rate
SzCORE	Seizure detection	Standardized event-based benchmarking framework
SeMaScore	ASR	Segment-wise semantic-and-error evaluation metric
SQS	Tennis	Server-side linear predictor for serve efficiency

The most direct instantiation of the string “Sscore” occurs in CEScore, where it denotes the simplicity component of a split-and-rephrase evaluation suite. By contrast, the LLM paper uses SCORE as an acronym for Systematic COnsistency and Robustness Evaluation, not as a single scalar metric, and the seizure-detection paper states explicitly that it does not define a separate metric called “S-Score” as a primary benchmark score (Nalbandyan et al., 28 Feb 2025, Dan et al., 19 May 2025). In statistical work, “score” frequently has its classical meaning, such as a likelihood score statistic or the density score function $s(x)=\nabla \log \rho(x)$ , which is conceptually different from benchmark-style scoring (Chen et al., 2017, Wibisono et al., 2024).

2. Sscore in split-and-rephrase evaluation

In CEScore, Sscore is the component meant to estimate the simplicity of a split-and-rephrase output. It is intended to answer the question “How much simpler is the output sentence(s) than the source?”, and it reflects whether the simplified text reduces complexity relative to the original. The paper frames human simplicity judgments on a scale from $-2$ to $2$, while the automatic Sscore is normalized to $[0,1]$ (Ajlouni et al., 2023).

Sscore is not computed by a learned neural model; it is a hand-designed statistical function based on the source text $T_c$ and simplified text $T_s$ . It first computes $\text{Stss}=TSS(T_s)$ and $\text{Ctss}=TSS(T_c)$ , then forms a normalized difference $[min,max]$ 0 from the TSS scores of the simplified and complex texts. Its clipping rule is explicit: if $[min,max]$ 1, then $[min,max]$ 2; if $[min,max]$ 3, then $[min,max]$ 4; if $[min,max]$ 5, then $[min,max]$ 6. The output is therefore a single scalar in $[min,max]$ 7, with higher values indicating that the output is judged more simple relative to the original. The paper also notes that scores above $[min,max]$ 8 are considered acceptable, while below $[min,max]$ 9 indicate failure to meet the criterion (Ajlouni et al., 2023).

Its internal structure consists of Sentence Length Score (SLS), Average Sentence Familiarity (ASF), and Text Simplicity Score (TSS). SLS is motivated by the empirical observation that very short sentences below a lower threshold are always judged simple, very long sentences above an upper threshold are always judged complex, and intermediate lengths show roughly linear degradation in simplicity; the empirically set thresholds are 5 words and 25 words, with parameters $s(x)=\nabla \log \rho(x)$ 0 and $s(x)=\nabla \log \rho(x)$ 1. ASF uses Brysbaert et al.’s percent_known values, SUBTLEX-US Zipf frequency values, and syllable counts to capture lexical familiarity. TSS then combines lexical and structural simplicity through

$s(x)=\nabla \log \rho(x)$ 2

$s(x)=\nabla \log \rho(x)$ 3

$s(x)=\nabla \log \rho(x)$ 4

with $s(x)=\nabla \log \rho(x)$ 5 and $s(x)=\nabla \log \rho(x)$ 6 (Ajlouni et al., 2023).

Sscore is explicitly distinct from the other CEScore components. Gscore measures grammaticality, Mscore measures meaning preservation via weighted token overlap, and the overall CEscore is

$s(x)=\nabla \log \rho(x)$ 7

Experimentally, Sscore aligned well with human simplicity judgments: at the sentence level, Pearson was $s(x)=\nabla \log \rho(x)$ 8 and Spearman was $s(x)=\nabla \log \rho(x)$ 9; at the model level, Pearson was $-2$ 0 and Spearman was $-2$ 1. The paper further reports that SARI, SAMSA, SAMSAabl, BLEU, and BERTScore were negatively correlated with human simplicity judgments in that experiment, whereas Sscore showed strong positive alignment. Its stated caveats are that it is a heuristic proxy, its thresholds and weights are empirically chosen, it depends on English lexical resources, sentence length is only a proxy for simplicity, and it is tuned to split-and-rephrase rather than general text simplification (Ajlouni et al., 2023).

3. SCORE for LLM robustness

In LLM evaluation, SCORE denotes Systematic COnsistency and Robustness Evaluation, a framework motivated by the claim that reporting a single accuracy number per benchmark is too narrow and often misleading. The framework repeatedly evaluates the same benchmark under multiple non-adversarial conditions and reports both accuracy ranges and a consistency rate (CR), rather than only a best-case score (Nalbandyan et al., 28 Feb 2025).

The framework operates on three perturbation families. Prompt robustness evaluates ten semantically similar prompt variants. Non-greedy inference robustness uses a fixed prompt but different random seeds at temperature $-2$ 2. Choice-order robustness permutes the order of answer options for multiple-choice tasks while preserving question semantics. Consistency is defined through pairwise similarity aggregation over predictions for each item; for multiple-choice questions, $-2$ 3 is $-2$ 4 if the extracted answer letters match and $-2$ 5 otherwise, while for MATH the answers are compared by symbolic equivalence using sympy. The paper emphasizes that CR measures prediction stability, not correctness, so a model can be consistently wrong or inconsistently correct (Nalbandyan et al., 28 Feb 2025).

The selected datasets are MMLU-Pro, AGIEval, and MATH. MMLU-Pro contains 12,032 questions across 14 subjects and up to 10 answer choices; AGIEval contains 2,340 examples from subsets such as SAT, LSAT, LogiQA, and AQuA-RAT; MATH contains about 5,000 competition-level math problems with difficulty levels 1–5. Accuracy is reported as mean accuracy together with the $-2$ 6 range across variants; for MMLU-Pro the paper uses category-level macro accuracy, whereas AGIEval and MATH use micro accuracy. The setup is fully generative and 0-shot, with MCQ answers parsed from the final letter and MATH answers parsed from boxed expressions and checked symbolically (Nalbandyan et al., 28 Feb 2025).

The empirical motivation is substantial instability under benign perturbations. The paper reports that paraphrasing prompts on MMLU-Pro can shift accuracy by up to $-2$ 7, and reordering answer choices on AGIEval can change accuracy by up to $-2$ 8. In the broader choice-order experiment, fluctuations on MMLU-Pro range from $-2$ 9 to $2$0, and on AGIEval from $2$1 to $2$2. It also reports that higher accuracy does not guarantee higher consistency, model size is not a reliable predictor of robustness, AGIEval tends to yield the highest mean accuracy and consistency, and MATH tends to have the lowest consistency. SCORE is therefore a framework for multi-condition reporting rather than an “S-Score” in the sense of a single canonical scalar. Its stated caveats are that it is non-adversarial, it focuses on MCQ-style objectively scored benchmarks, it is computationally expensive, and it remains exposed to possible training-data contamination (Nalbandyan et al., 28 Feb 2025).

4. SzCORE and event-based seizure benchmarking

SzCORE is introduced as the Seizure Community Open-Source Research Evaluation framework for validating EEG-based seizure detection algorithms. It standardizes EEG dataset organization and file format, compliance with BIDS-EEG and HED-SCORE specifications, and the evaluation methodology, especially event-based scoring. The paper states explicitly that it does not define a separate metric called “S-Score” as a primary benchmark score; the operational metrics are sensitivity, precision, F1-score, and false positives per day, with event-based F1 as the ranking criterion (Dan et al., 19 May 2025).

Its scoring is event-based rather than sample-based. If a reference seizure event and a hypothesis event overlap, the seizure is counted as detected; if a hypothesis event overlaps nothing in the reference set, it is a false positive; if a reference event is not overlapped by any prediction, it is a false negative. The challenge used any overlap as the minimum overlap criterion, a 30 second pre-ictal tolerance, a 60 second post-ictal tolerance, a minimum duration of 90 seconds with event merging, and a maximum event duration of 5 minutes with splitting. Scores were computed per subject by summing TP/FP/FN over that subject’s recordings, and the subject-level metrics were averaged across subjects using an arithmetic mean (Dan et al., 19 May 2025).

The challenge dataset came from the Filadelfia Danish Epilepsy Centre and contained 65 subjects, 398 annotated seizures, and 4,360 hours of EEG from January 2018 to December 2020. Recordings were long-term scalp EEG from the EMU, stored as EDF files with 19 electrodes in the international 10–20 system, referenced to common average and sampled at 256 Hz. Three board-certified neurophysiologists annotated seizures independently, and disagreements were resolved by consensus. The winning submission, Sz Transformer, achieved F1-score $2$3, sensitivity $2$4, precision $2$5, and about 1.34 false positives per day. The paper emphasizes the large discrepancy between developers’ self-reported scores and the independent challenge results, interpreting this as a generalization gap between internal validation and deployment-like evaluation (Dan et al., 19 May 2025).

This usage is important because it prevents a common terminological error: in this literature, SzCORE is the framework, event-based F1 is the ranking score, and there is no separate closed-form “S-Score” beyond the overlap-based event scoring scheme (Dan et al., 19 May 2025).

5. Score functions and score statistics

A different family of usages treats “score” in the statistical sense. In nonparametric score estimation, the target is the score function

$2$6

estimated from i.i.d. samples under the weighted loss

$2$7

Under $2$8-subgaussianity and an $2$9-Lipschitz score, the minimax-optimal rate is

$[0,1]$ 0

and a regularized Gaussian-kernel empirical Bayes estimator attains this rate up to logarithmic factors. The same paper highlights the curse of dimensionality through the scaling $[0,1]$ 1 for fixed target accuracy, and relates the result to sample complexity in score-based generative models (Wibisono et al., 2024).

Score-based generative inference uses the same object differently. In score-based data assimilation, the method learns a score model for state trajectories and exploits the observation that the score of an arbitrarily long trajectory can be decomposed approximately into short-segment scores. Training uses denoising score matching on short windows, while inference combines the learned prior score with a likelihood score at test time, enabling non-autoregressive posterior sampling and zero-shot observation scenarios (Rozet et al., 2023). Manifold Aware Denoising Score Matching makes a related decomposition,

$[0,1]$ 2

where $[0,1]$ 3 is a known base score encoding manifold support and $[0,1]$ 4 is the residual learned by the network. The paper derives analytical $[0,1]$ 5 terms for discrete supports, spheres, and $[0,1]$ 6, arguing that the residual is smaller and easier to learn than the full score (Levy-Jurgenson et al., 2 Mar 2026).

Classical score statistics appear in surveillance and anomaly detection. The $[0,1]$ 7 paper defines a standardized one-dimensional score statistic $[0,1]$ 8 for detecting weak spatio-temporal signals that may change both mean and covariance structure, and develops analytical false-alarm approximations for threshold calibration (Chen et al., 2017). In anomaly detection, Overlap loss moves from fixed anomaly-score targets to discrimination between score distributions of normal and abnormal data. It estimates score densities by KDE, computes an overlap measure using an intersection point $[0,1]$ 9,

$T_c$ 0

and uses this bounded loss to separate score distributions without predefined targets or margins (Jiang et al., 2023).

6. Other named score constructs

SeMaScore is an ASR evaluation metric that combines segment-wise semantic similarity with an explicit transcription-error penalty. After edit-distance-based segment mapping, it computes segment cosine similarity $T_c$ 1, penalizes it by Match Error Rate,

$T_c$ 2

weights each segment by contextual importance

$T_c$ 3

and aggregates to

$T_c$ 4

The paper reports that this metric corresponds well with expert human assessments, signal-to-noise ratio levels, and other NLP metrics, and that it was $T_c$ 5 faster than BERTScore on the TORGO dataset, with average computation times of 1.95 seconds for BERTScore and 0.047 seconds for SeMaScore (Sasindran et al., 2024).

The Server Quality Score (SQS) is a serve-specific tennis metric derived from logistic mixed-effects models fit separately for first and second serves, tournament, and gender. For player $T_c$ 6 and serve type $T_c$ 7, it is the server-side linear predictor combining average speed, speed variability, modal location bin, location entropy, and the server random intercept, with the returner effect set to zero so that SQS corresponds to performance against an average returner. In out-of-sample evaluation, SQS aligned more strongly with serve efficiency than weighted Elo, especially for first serves; for example, on Wimbledon men’s first serves, the reported correlations were $T_c$ 8 for SQS versus $T_c$ 9 for wElo (Li et al., 8 Feb 2026).

In clustering, the paper “A score function for Bayesian cluster analysis” defines a parameter-free score $T_s$ 0 for a partition $T_s$ 1,

$T_s$ 2

balancing a within-cluster variance term against a between-cluster entropy term. The paper proposes it as a parameter-free Bayesian clustering score for selecting the number of clusters in methods such as hierarchical clustering and $T_s$ 3-means (Noble et al., 2019).

Online multiple testing introduces another acronymic usage: SCORE, Sequential Control with Overshoot Refund for E-values, refines e-value-based online FDR procedures by exploiting

$T_s$ 4

The overshoot term $T_s$ 5 is interpreted as refundable evidence, leading to SCORE-LOND, SCORE-LORD, and SCORE-SAFFRON, which strictly dominate their baseline counterparts while preserving finite-sample FDR control (Kuang et al., 28 Jan 2026).

7. Interpretation and recurring issues

A central interpretive point is that “S-Score” is not a universal object. Some usages are scalar output-quality metrics, as in CEScore’s Sscore or SeMaScore; some are framework names whose outputs are deliberately multi-dimensional, as in SCORE for LLM robustness and SzCORE for seizure detection; some are model-based latent summaries such as SQS; and some are derivative-based statistical objects such as $T_s$ 6 or Rao-type score statistics (Ajlouni et al., 2023, Nalbandyan et al., 28 Feb 2025, Dan et al., 19 May 2025, Li et al., 8 Feb 2026, Wibisono et al., 2024, Chen et al., 2017).

Several recurring distinctions follow from that heterogeneity. First, robustness-oriented work often rejects the idea that one number is sufficient: SCORE for LLMs reports mean accuracy together with $T_s$ 7 ranges and consistency rate, and explicitly argues that consistency complements rather than replaces accuracy (Nalbandyan et al., 28 Feb 2025). Second, clinically oriented benchmarking can favor event-level counting over pointwise metrics: SzCORE ranks algorithms by event-based F1 because the task is seizure detection and alerting rather than sample-wise waveform labeling (Dan et al., 19 May 2025). Third, in statistical learning, “score” often denotes a structural quantity that drives inference or sampling rather than a post hoc evaluation number, as in score estimation, score-based data assimilation, and manifold-aware denoising score matching (Wibisono et al., 2024, Rozet et al., 2023, Levy-Jurgenson et al., 2 Mar 2026).

This suggests that cross-paper comparison of “S-score” values is generally not meaningful without domain-specific normalization and without checking the unit of analysis, the aggregation rule, and whether the reported object is a metric, a framework output, or a score-function quantity. Within the cited literature, terminological overlap is common, but operational definitions are local to each method and problem setting.