Comprehensive Applicant Profile Score (CAPS)

• CAPS is a holistic college admissions framework that quantitatively models applicant profiles by separating academic performance, essay quality, and extracurricular engagement.
• It integrates modular design with interpretable metrics, using standardized tests, AI-generated embeddings, and rule-based scoring to produce a composite score.
• The framework emphasizes transparency through weighted fusion, SHAP-based explanations, and explicit bonus adjustments to address fairness and bias concerns.

Comprehensive Applicant Profile Score (CAPS) is a multi-modal, modular, and interpretable framework for quantitatively modeling holistic college admissions evaluations. In its explicit formulation, CAPS decomposes an applicant profile into three components—academic performance, essay quality, and extracurricular engagement—then fuses them into a final score intended for transparent, AI-aided decision support rather than purely autonomous selection. The term is introduced in "Quantifying Holistic Review: A Multi-Modal Approach to College Admissions Prediction" (Zeng et al., 12 Jul 2025), where CAPS is presented as a response to opaque, subjective, and inconsistent holistic review procedures.

1. Definition and Conceptual Scope

CAPS is defined as a framework that “quantitatively model[s] and interpret[s] holistic college admissions evaluations” through a structured decomposition of applicant evidence (Zeng et al., 12 Jul 2025). Its central design claim is that holistic review need not remain an opaque heuristic: academic records, essays, and extracurricular activity can be converted into component scores with explicit weighting, modular processing, and post hoc explanation. In the paper’s comparison between traditional review and CAPS, the contrast is stated as “Only GPA, SAT” versus “GPA + Essays + Extracurriculars (EC),” “Opaque Heuristics” versus “Transparent ML-based Model,” and “Manual Review” versus “AI-Aided Modular Scoring” (Zeng et al., 12 Jul 2025).

This conceptualization places CAPS at the intersection of scoring, prediction, and decision support. It is a scoring system because it produces module scores and a final score. It is a predictive framework because it trains models for essay-quality estimation and admissions-tier classification. It is a decision-support tool because the framework emphasizes transparency, SHAP-based explanation, and applicant- or counselor-facing feedback rather than full automation (Zeng et al., 12 Jul 2025).

A common misconception is that CAPS denotes a single opaque scalar replacing holistic review. The original framework is explicitly the opposite: it is built around component-level decomposition into Standardized Academic Score (SAS), Essay Quality Index (EQI), and Extracurricular Impact Score (EIS), with human-readable substructure retained throughout the pipeline (Zeng et al., 12 Jul 2025).

2. Core Architecture and Score Construction

The CAPS architecture is organized around three modules and a late fusion stage.

Module	Primary inputs	Output role
SAS	GPA, SAT, TOEFL, AP_5_Count, Course_Difficulty	Academic strength
EQI	GPT-4o rubric scores, MiniLM essay embeddings, alignment score	Essay quality
EIS	GPT-4o activity scores, tier labels, coherence score	Extracurricular impact

The final fusion mechanism is explicitly given as a weighted combination of the module outputs. The paper defines fused component weights as

$$w_i = \alpha \cdot w_{\text{log},i} + \beta \cdot w_{\text{xgb},i} + \gamma \cdot w_{\text{exp},i}, \quad \alpha+\beta+\gamma=1$$

with default values $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$, followed by

$\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$

and

$$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$

where $x_i$ denotes the standardized module outputs and the bonus may add “up to 12” points for equity considerations (Zeng et al., 12 Jul 2025).

The final reported component weights are comparatively balanced but not symmetric: SAS receives $0.40$, EQI receives $0.31$, and EIS receives $0.29$ (Zeng et al., 12 Jul 2025). This keeps academics as the largest single contributor while preserving substantial influence for essays and extracurriculars.

The “bonus” term is one of the framework’s most controversial elements. The paper gives examples such as “URM, LGBTQ+, rural, green card,” but does not provide an empirical fairness evaluation of the adjustment itself (Zeng et al., 12 Jul 2025). This suggests that CAPS combines learned and rule-based layers rather than relying on a single unified optimization objective.

3. Module Design: SAS, EQI, and EIS

SAS

The Standardized Academic Score quantifies academic ability using GPA, standardized tests, English proficiency, AP performance, and coursework rigor (Zeng et al., 12 Jul 2025). Each feature is z-score standardized,

$z_i = \frac{x_i - \mu}{\sigma}$

then combined through a hybrid weighting scheme that mixes expert priors with PCA-derived weights. The paper specifies PCA coefficients $\alpha_{\text{PCA}} = 1.0$ and $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$0, and a fusion coefficient $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$1, meaning the final academic weights remain close to manual admissions-style priors (Zeng et al., 12 Jul 2025).

The reported fused feature weights are GPA $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$2, SAT $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$3, TOEFL $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$4, AP_5_Count $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$5, and Course_Difficulty $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$6 (Zeng et al., 12 Jul 2025). GPA is therefore the dominant academic feature, with coursework rigor second. SAS is then transformed through a softmax and sigmoid-based scaling to produce a 0–100 academic module score (Zeng et al., 12 Jul 2025).

EQI

The Essay Quality Index is the most model-heavy component. GPT-4o first scores each essay on Content, Language, and Structure, each on a 1–5 scale, while all-MiniLM-L6-v2 generates a 384-dimensional semantic embedding (Zeng et al., 12 Jul 2025). These are concatenated into a 387-dimensional feature vector and passed to an XGBoost regressor predicting a continuous EQI score in $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$7 (Zeng et al., 12 Jul 2025).

The pipeline then applies a prompt-alignment penalty. The paper defines the adjusted score as

$(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$8

where $(\alpha,\beta,\gamma) = (0.3, 0.3, 0.4)$9 is a GPT-4o alignment score, $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$0 is a minimum penalty factor, $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$1 is sigmoid steepness, and $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$2 is the penalty threshold (Zeng et al., 12 Jul 2025). This makes topic adherence part of essay evaluation rather than a separate hard filter.

EIS

The Extracurricular Impact Score is rule-based and LLM-assisted. Each activity receives a GPT-4o score in $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$3 for impact, uniqueness, and leadership, plus a tier label mapped as $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$4, $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$5, $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$6, $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$7, and $\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$8 (Zeng et al., 12 Jul 2025). The activity-level score is

$\text{CAPS}_{\text{raw}} = \sum_i w_i \cdot x_i$9

with default $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$0 (Zeng et al., 12 Jul 2025).

Portfolio-level thematic coherence is then scored and applied as a bounded multiplier:

$$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$1

so coherence can modestly penalize or preserve the average activity score, but cannot overwhelm it (Zeng et al., 12 Jul 2025).

4. Empirical Performance and Interpretability

The empirical evidence in the CAPS paper is mixed in character because different parts of the system are evaluated differently. The strongest quantitative result concerns EQI: on an essay dataset of 200 essays, using an 80/20 split, fixed random_state = 42, and 3-fold cross-validation, the XGBoost regressor achieved $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$2 and $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$3, with best cross-validated negative MSE of $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$4 (Zeng et al., 12 Jul 2025). The abstract rounds this to an EQI prediction $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$5 of $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$6.

For downstream admissions-tier prediction, multinomial logistic regression achieved 75% test accuracy, macro F1 of 0.69, and weighted F1 of 0.74 (Zeng et al., 12 Jul 2025). The XGBoost classifier reached 100% training accuracy with macro and weighted F1 of 1.00, but the paper explicitly cautions that this “may indicate potential overfitting” and does not report a held-out test accuracy for that classifier (Zeng et al., 12 Jul 2025).

Interpretability is a central design goal. SHAP is used to explain the EQI model, and the most important features are reported to be the three GPT rubric scores—EssayContentScore, EssayLanguageScore, and EssayStructureScore—alongside several latent embedding dimensions such as EssayEmbedding_19, EssayEmbedding_375, and EssayEmbedding_319 (Zeng et al., 12 Jul 2025). GPT-4o is then used again to convert SHAP signals into “targeted, actionable feedback,” creating a pipeline in which the same system both scores and explains.

A further misconception is that CAPS has already been validated on real institutional admissions data. The paper is explicit that the applicant dataset is “synthetic yet realistic,” intended to emulate U.S. holistic review, and that future work includes “incorporating real institutional data” (Zeng et al., 12 Jul 2025). The evidence therefore supports feasibility and internal coherence more strongly than external validity.

5. CAPS in Relation to Broader Applicant-Scoring Research

Several adjacent lines of research address components that the original CAPS paper leaves modular or underdeveloped. They do not define the same framework, but they illuminate how a broader applicant-scoring system might be assembled.

In applicant tracking and resume screening, MLAR provides a three-layer LLM-based pipeline consisting of job requirement extraction, resume parsing, and candidate-job similarity scoring. It formalizes job and resume representations as feature sets $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$7 and $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$8, computes similarity as $$\text{CAPS}_{\text{final}} = \min(100, \text{CAPS}_{\text{raw}} \times 100 + \text{bonus})$$9, uses a similarity score between 0 and 100, ranks candidates in descending order, and selects the top 3 for notification (Younes et al., 14 Jul 2025). The paper explicitly does not define a formal multi-component CAPS, but it already implements a monolithic suitability score that could serve as a resume-matching layer.

For interview evidence, "Listening to the Unspoken" predicts five continuous dimensions—Integrity, Cooperation / Collegiality, Social versatility, Development orientation, and Overall employability / hireability—from three modalities, six interview responses, and a two-level ensemble architecture. Its framework achieved a multi-dimensional average test MSE of 0.1824 and is directly relevant as an interview subscore generator rather than a full applicant profile (Li et al., 30 Jul 2025). CoMAI approaches the same subproblem differently: it uses a centralized finite-state machine coordinating four agents for question generation, security, scoring, and summarization; supports rubric-based multidimensional scoring; keeps the Scoring agent resume-agnostic; and reports 90.47% accuracy, 83.33% recall, and 84.41% candidate satisfaction in a university admissions talent-selection setting (Sun et al., 17 Mar 2026).

For rule-grounded admissibility and fit, automated matchmaking models applicant selection as profile-to-profile comparison rather than cohort-relative ranking. The framework represents each profile as a set of constraints $x_i$0, distinguishes hard from soft constraints, supports composite “choose $x_i$1 of $x_i$2” rules, and computes overall similarity multiplicatively as

$x_i$3

(Olugbara et al., 2015). This yields an absolute programme-fit score that does not require comparison with other applicants.

For contextual fairness, the Adaptive Merit Framework adds a correction layer to a base merit score. Its core rule is

$x_i$4

with operational truncation at zero for high-SES applicants, and conditional admission when the corrected score exceeds the threshold of the last regular admit (Lee, 3 Dec 2025). This is not a full CAPS, but it functions as a fairness-aware adjustment and thresholding layer that could sit on top of any transparent base applicant profile score.

A plausible synthesis is that broader CAPS architectures divide naturally into at least four layers: evidence extraction, component scoring, contextual correction, and final thresholding. That synthesis is not stated by any single paper, but it is strongly suggested by the complementarity among multi-modal admissions scoring (Zeng et al., 12 Jul 2025), ATS matching (Younes et al., 14 Jul 2025), interview modules (Li et al., 30 Jul 2025, Sun et al., 17 Mar 2026), rule-based fit systems (Olugbara et al., 2015), and SES correction frameworks (Lee, 3 Dec 2025).

6. Limitations, Controversies, and Open Problems

The most immediate limitation of CAPS in its original form is empirical. The main applicant-level experiments rely on a synthetic but realistic dataset, and even the essay dataset mixes authentic and synthetic materials (Zeng et al., 12 Jul 2025). This constrains claims about deployment validity, calibration, and robustness under real institutional heterogeneity.

Fairness remains unsettled. CAPS includes a diversity bonus, but the paper does not report subgroup fairness metrics, adverse-impact analysis, or calibration by protected or contextual groups (Zeng et al., 12 Jul 2025). AMF makes the opposite design choice: it argues for a direct, continuous SES correction measured against national reference populations, with administrative verification and audit logs, rather than coarse proxy categories or opaque holistic adjustments (Lee, 3 Dec 2025). This suggests a substantive controversy within CAPS-like systems: whether equity is better implemented as explicit context correction or as a normative bonus layer.

Another unresolved issue concerns which evidence should be quantified at all. The exploratory readiness system for university students uses self-reported variables such as communication, comfort zone, salary expectations, community interest, and openness to change, but it is based on $x_i$5, uses self-report-only inputs, and lacks external outcome validation (Assylzhan et al., 2023). This suggests caution in extending psychosocial scoring into high-stakes admissions without stronger psychometric and fairness validation.

Similarity-based applicant scoring raises a different concern. Turtle Score models technical-fit and work-pattern similarity through GitHub activity, learning analytics, error archives, job shadowing, puzzle performance, and Kaggle signals, then predicts suitable roles with Linear SVC at 87% reported accuracy (Varshini et al., 2022). A plausible implication is that heavy reliance on similarity to historical employees may privilege workforce replication over genuine diversity of potential, especially when the reference population is itself historically skewed.

Interview automation introduces additional governance challenges. The multimodal interview framework emphasizes “holistic” and “fair” assessment but provides no subgroup fairness metrics, while CoMAI reports negligible verbosity bias and strong prompt-injection defense but still lacks full subgroup fairness reporting and broad external validity outside its tested setting (Li et al., 30 Jul 2025, Sun et al., 17 Mar 2026). In CAPS terms, this means that adding interview intelligence can improve evidence richness while simultaneously expanding the surface area for bias, privacy concerns, and contestability disputes.

The broader open problem is therefore not merely how to build a more accurate applicant score. It is how to build a score whose decomposition, data sources, fairness corrections, thresholds, and explanations remain auditable under high-stakes institutional use. CAPS, in its original formulation, establishes a modular vocabulary—SAS, EQI, EIS—for one such architecture (Zeng et al., 12 Jul 2025). The surrounding literature suggests that any fully developed CAPS would also need validated interview modules, rule-grounded admissibility logic, contextual fairness correction, and stronger empirical governance than current prototypes yet provide (Younes et al., 14 Jul 2025, Olugbara et al., 2015, Lee, 3 Dec 2025).