Mastery-Score: Frameworks and Applications

• Mastery-Score is a comprehensive term defining multiple operationalizations of competence, including probabilistic, continuous, and threshold-based approaches.
• It is operationalized through various methods such as Bayesian adaptive models, cognitive diagnosis frameworks, and efficiency ratios to capture fluency and progression.
• The concept informs both educational assessment and systems benchmarks by guiding adaptive learning, performance consolidation, and reinforcement learning optimization.

Searching arXiv for the cited benchmark and related "mastery" literature to ground the article in current arXiv records. Searching arXiv for "(Gu et al., 2023) Domain Mastery Benchmark". Mastery-Score is not a standardized singular metric in the arXiv literature. Instead, the term corresponds to a family of operationalizations of mastery, readiness, or competence that vary by domain and modeling tradition. Depending on the setting, mastery may be represented as a fluency-sensitive expected score, a concept-level mastery probability, a continuous attribute profile, a thresholded assessment outcome, a ratio of successful mastery to practice, a rubric stage for independent performance, a blueprint-weighted benchmark component, or a prompt-level accuracy regime in reinforcement learning. This suggests that “Mastery-Score” is best treated as an umbrella label for criterion-referenced competence measures rather than as a fixed technical formalism (Sapountzi et al., 2021, Shang et al., 2022, Verma, 31 Aug 2025).

1. Terminological scope and major families of definition

Several papers explicitly reject the idea that mastery must be a single scalar. The “Mastery Rubric for Statistics and Data Science” presents mastery as stage placement across 13 KSAs rather than as a summed index (Tractenberg et al., 2023). The PCP literacy intervention likewise uses an 80\% formative threshold to gate progression, but does not define an overall composite mastery score (Srinivas et al., 11 Jun 2025). The multilingual commonsense benchmark “mSCoRe” is not a mastery metric at all, but a benchmark whose title is acronymic rather than metric-defining (Ngo et al., 13 Aug 2025).

The literature therefore separates into a small number of recurring design patterns.

Form	Representative definition	Representative sources
Probabilistic scalar	Expected score or mastery probability	(Sapountzi et al., 2021, Liu et al., 2023, Sales et al., 2017)
Continuous profile	Per-attribute mastery on $[0,1]$	(Shang et al., 2022, Cárdenas-Hurtado et al., 25 Nov 2025)
Thresholded mastery	Fixed cutoff for progression or classification	(Pelkola et al., 2017, Srinivas et al., 11 Jun 2025)
Ratio or efficiency proxy	Mastered relative to practiced	(Dani, 2016)
Rubric stage	Developmental placement by KSA	(Tractenberg et al., 2023)
Blueprint or benchmark component	Weighted topic-level mastery term	(Verma, 31 Aug 2025, Gu et al., 2023)
Optimization status	Prompt mastery via rollout accuracy	(Liao et al., 18 Apr 2026)

This heterogeneity is substantive rather than merely terminological. In educational assessment, mastery scores are usually diagnostic and criterion-referenced. In benchmark design, they summarize topic coverage or task difficulty. In reasoning-model optimization, they mark regimes in which competence should be consolidated rather than re-discovered.

2. Scalar probabilistic formulations

One of the clearest scalar operationalizations appears in Bayesian Adaptive Mastery Assessment. BAMA defines latent accuracy and response-time parameters,

$\theta \in [0,1], \qquad \lambda > 0,$

with

$$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$

Its per-item mastery-related score is

$Z = P \cdot \left(1-\frac{T}{d}\right)^+,$

so correct but slow responses contribute less than correct and fluent responses. The practical mastery score is the posterior expected value

$\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$

where $s_t=(\alpha_t,\beta_t,n_t,\gamma_t)$ is the sequential posterior state. BAMA therefore makes mastery a continuous, fluency-sensitive expectation, with optional ordinal bands ranging from “Wheel-spinning” to “Mastered” (Sapountzi et al., 2021).

TRACED uses a different probabilistic semantics. Its basic object is the student–concept–time mastery probability

$p(u_{i,k}^t = 1),$

where $u_{i,k}^t$ is a binary latent mastery state for student $i$, concept $k$, and time $\theta \in [0,1], \qquad \lambda > 0,$0. The transition model incorporates explicit learning and forgetting: $\theta \in [0,1], \qquad \lambda > 0,$1 The paper treats this concept-level posterior or prior probability as the interpretable mastery score, and uses an LSTM-based approximation to avoid the $\theta \in [0,1], \qquad \lambda > 0,$2 complexity of exact multi-concept inference (Liu et al., 2023).

A third scalar latent formulation appears in work on Cognitive Tutor Algebra I. There, section-level mastery outcomes are modeled with a Rasch equation,

$\theta \in [0,1], \qquad \lambda > 0,$3

where $\theta \in [0,1], \qquad \lambda > 0,$4 is student $\theta \in [0,1], \qquad \lambda > 0,$5’s latent propensity to master worked sections under treatment and $\theta \in [0,1], \qquad \lambda > 0,$6 is section difficulty. This is not a raw mastery rate but a latent mastery propensity, intended to correct for unequal section difficulty and unequal exposure (Sales et al., 2017).

Across these formulations, mastery is not simply “percent correct.” It is a latent or posterior quantity combining evidence, uncertainty, and in some cases fluency or temporal dynamics.

3. Continuous partial mastery and multidimensional profiles

Partial-mastery cognitive diagnosis models replace binary attribute indicators with continuous mastery vectors. In PM-CDMs, the core score is

$\theta \in [0,1], \qquad \lambda > 0,$7

where each coordinate is learner $\theta \in [0,1], \qquad \lambda > 0,$8’s degree of mastery of one latent attribute. The paper emphasizes that the natural mastery output is not a single total score but an attribute-specific profile, typically estimated by posterior means $\theta \in [0,1], \qquad \lambda > 0,$9. This allows partial possession of a skill and can materially change conclusions relative to binary CDMs (Shang et al., 2022).

GaPM-CDM extends the same idea nonparametrically. Its latent variables are

$$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$0

with item response functions modeled as mixtures of monotone nonparametric functions of attributes. The most natural mastery score is the posterior mean

$$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$1

again yielding an attribute profile rather than a single scalar. This preserves interpretability while relaxing strong parametric assumptions inherited from classical CDMs (Cárdenas-Hurtado et al., 25 Nov 2025).

The rubric-based literature pushes this logic further by making mastery explicitly profile-based. In MR-SDS, mastery is represented by stage placement across 13 KSAs using six developmental stages: Beginner, Early Apprentice, Late Apprentice, Early Journeyman, Middle Journeyman, and Late Journeyman. The paper identifies the boundary between Late Apprentice (A2) and Early Journeyman (J1) as the transition to independent practice. There is no equation, no weighted sum, and no psychometric cut-score; the “score” is the learner’s stage on each KSA, especially relative to the developing-versus-independent boundary (Tractenberg et al., 2023).

A plausible implication is that multidimensional mastery models are preferred when the construct is inherently decomposable. They preserve information that would be lost under one-number aggregation.

4. Thresholded, ratio-based, and progression-oriented scoring

A more operational tradition defines mastery by explicit cutoffs. In blended mastery learning for university mathematics, mastery was defined as achieving $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$2 out of $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$3 points or $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$4 out of $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$5 points on a weekly quiz or formative test, with separate notions of initial mastery and eventual mastery. The same study introduced a penalised exercise score that discounted repeated attempts and found that raw exercise points had poor discriminatory power because of a ceiling effect produced by unlimited retries; performance on exercises predicted mastery on formative tests only “to a small extent” (Pelkola et al., 2017).

In ALEKS-based learning analytics, the main mastery proxy was the ratio

$$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$6

interpreted as the efficiency with which practice is converted into actual mastery. This ratio correlated with final exam marks at $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$7, and together with initial assessment score explained $$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$8 of the variance in final exam marks. The reported regression was

$$P \mid \theta \sim \mathrm{Bernoulli}(\theta), \qquad T \mid \lambda \sim \mathrm{Exponential}(\lambda).$$9

The same study treated retained mastery as the relation between topics mastered before a comprehensive test and the post-test retained total, and associated sequential topic selection with better retention than random selection (Dani, 2016).

In Cognitive Tutor Algebra I, section progression itself becomes the relevant mastery unit. Each worked section ends in one of four mutually exclusive statuses—Mastery, Promotion, Reassignment, or Final—and the paper studies how often students advance without mastery and how reassignment before mastery relates to post-test outcomes. Reassignment estimates were consistently negative, including $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$0 for the binary “ever reassigned” indicator, although the paper treats those results as observational rather than cleanly causal (Israni et al., 2018).

A related but more predictive use appears in high-dosage tutoring with MATHia. There, the most important ITS-side mastery feature is Opportunities_Till_Mastery, which becomes the root split in the strongest ITS-only and combined predictive trees. The combined tutor-discourse-plus-ITS model achieved AUC $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$1, compared with $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$2 for talk moves alone and $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$3 for ITS alone, indicating that mastery efficiency interacts with tutoring discourse rather than functioning as a standalone predictor (Abdelshiheed et al., 2024).

These formulations make mastery actionable but also expose a recurring problem: thresholds, retries, and progression rules often measure test-taking process or practice efficiency as much as substantive competence.

5. Benchmark and systems uses

Outside learner modeling, mastery is often embedded in benchmark design. The “Domain Mastery Benchmark” introduces DomMa as “an ever-updating benchmark” for evaluating LLM domain knowledge. DomMa consists of 100,000 questions in both Chinese and English, targets Chinese 112 first-level subject classifications, and sources questions from graduate entrance examinations and undergraduate exams in Chinese colleges. Here mastery is benchmarked through broad and continually updated domain coverage rather than through a single learner-centric score (Gu et al., 2023).

The Exam Readiness Index gives a more formal systems-level definition. ERI is a blueprint-aware composite built from six normalized components, one of which is Mastery: $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$4 This makes mastery a topic-weighted aggregation of topic-level mastery maps $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$5, explicitly tied to the exam blueprint. ERI thereby treats mastery as necessary but not sufficient for readiness; coverage, retention, pace, volatility, and endurance remain separate components (Verma, 31 Aug 2025).

The term is also used in recommender systems. In player-conditional League of Legends champion ranking, the paper defines a combined mastery/familiarity term

$Z = P \cdot \left(1-\frac{T}{d}\right)^+,$6

where $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$7 is direct mastery for a champion the player has used and $Z = P \cdot \left(1-\frac{T}{d}\right)^+,$8 is indirect familiarity transferred from similar champions. Mastery here is neither academic nor psychometric; it is a comfort-and-familiarity prior inside an interpretable ranking system (Heo et al., 18 May 2026).

At the edge of the term’s scope lies “mSCoRe,” a multilingual commonsense reasoning benchmark. The paper is explicit that mSCoRe is not a “Mastery-Score” metric, but a skill-based benchmark with multilinguality, fine-grained reasoning skills, and scalable complexity. This is a useful counterexample because it shows that not every occurrence of “score” or “mastery” defines a scalar mastery measure (Ngo et al., 13 Aug 2025).

6. Mastery as consolidation in reasoning and agent training

In reinforcement learning for reasoning models, mastery shifts from measurement to optimization status. MCPO defines prompt-level rollout precision

$Z = P \cdot \left(1-\frac{T}{d}\right)^+,$9

and treats prompts with $\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$0 as mastered prompts, while prompts with $\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$1 are majority-correct prompts. The method then adds a hinge-KL regularizer only on mastered prompts and changes prompt weighting so that

$\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$2

Mastery is therefore not a score to be reported, but a regime in which policy drift should be bounded and partial correctness should be consolidated into full correctness (Liao et al., 18 Apr 2026).

SCoRe, in “From Correction to Mastery,” likewise does not define a scalar mastery score. Instead, it uses a student-centered training pipeline in which the student generates a trajectory, the teacher intervenes only at the first critical error, corrected trajectories are used for supervised fine-tuning, and short-horizon RL starts from the verified prefix before the first critical error with reward assigned at that key step. Mastery here denotes autonomous problem-solving beyond teacher imitation, not a standalone metric (Lyu et al., 12 Sep 2025).

In software-engineering agents, “mastery” is again operational rather than scalar. “Immersion in the GitHub Universe: Scaling Coding Agents to Mastery” states that it does not define a formal Mastery-Score; the closest proxy is resolved rate on SWE-bench Verified. The reported number is $\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$3, up from $\text{Mastery-Score}_t := \mathbb{E}[Z \mid s_t],$4 for the base model, and is supported by executable verification criteria for task validity and training-trajectory quality (Zhao et al., 10 Feb 2026).

These uses broaden the term considerably. They show that “mastery” can denote an optimization objective, a retention regime, or an end-task competence proxy rather than a learner assessment variable.

7. Validity, limitations, and recurrent controversies

A persistent theme is that mastery is highly definition-dependent. Exam-based benchmarks such as DomMa measure broad domain knowledge through exam questions, but this does not automatically establish professional or research-level competence (Gu et al., 2023). BAMA measures a fluency-sensitive score for a single skill with homogeneous items and simulated validation; the paper itself notes that “true mastery” remains hypothetical and that threshold, time limit, discount factor, and priors require calibration (Sapountzi et al., 2021).

Another recurring issue is inflation through retries. In blended mastery learning, eventual mastery was very high under unlimited reattempts, whereas initial mastery was much lower and more informative; the paper explicitly treats eventual online success as an overstatement of independent mastery (Pelkola et al., 2017). Similar concerns appear in section-based tutoring logs, where promotion, reassignment, and curriculum customization complicate any naive interpretation of mastery rates as pure learning measures (Israni et al., 2018).

Profile-based approaches solve some problems but create others. PM-CDM and GaPM-CDM allow refined partial mastery estimation, yet their outputs are multidimensional and model-dependent, so collapsing them to one scalar can obscure substantively important differences across attributes (Shang et al., 2022, Cárdenas-Hurtado et al., 25 Nov 2025). MR-SDS makes the same point from a curricular perspective: stage placement can vary by KSA, and the paper does not define any aggregation rule across the 13 KSAs (Tractenberg et al., 2023).

Finally, several frameworks explicitly separate mastery from adjacent constructs. ERI distinguishes mastery from coverage, retention, pace, volatility, and endurance, implying that high mastery on content does not guarantee readiness for an exam (Verma, 31 Aug 2025). This suggests that “Mastery-Score” is often most defensible when interpreted locally—per concept, per attribute, per benchmark topic, per section, or per prompt—rather than as a universal summary of competence.