Wolf-pack algorithm with multiple leaders flags deceptive statements in executive background checks

A new hybrid artificial intelligence framework borrows the cooperative hunting logic of wolf packs to hunt a very different kind of prey: strategically crafted lies in the written statements of executive job candidates. The system, described in Discover Artificial Intelligence, combines a redesigned swarm-intelligence optimizer with a dual-channel text representation to flag statements that documentary evidence later contradicted. On a proprietary corpus of 1,710 statement units drawn from 286 anonymized executive background-check cases, the complete framework achieved an F1 score of 0.8733 (95% CI: 0.861–0.886) and an area under the ROC curve of 0.9012, significantly outperforming standard Grey Wolf Optimization and BERT fine-tuning baselines with adjusted p < 0.001.

The core algorithmic innovation is called the Multi-Leadership Wolf Search Algorithm, or ML-WSA, and it targets a well-known weakness of the original Grey Wolf Optimization method. Standard GWO mimics a wolf pack’s hierarchy by letting the three fittest solutions—the alpha, beta and delta wolves—guide every other search agent. That elite-guidance scheme is simple and fast, but it concentrates the whole population’s movement around a handful of leaders. When those elites cluster prematurely in a local optimum, the entire search stagnates. ML-WSA replaces this fixed hierarchy with a dynamic leadership structure: the population is partitioned according to its spatial distribution, and a local best solution is elected from each region. The leader set therefore changes as the population geometry evolves, keeping several promising regions of the high-dimensional feature space active simultaneously.

Leader influence is also individualized rather than globally shared. Each search agent weighs guidance from the regional leaders using coefficients determined jointly by leader fitness and spatial proximity, so agents in different areas receive different directions from the same leader set. This creates what the authors describe as a state-dependent guidance graph, in which information flow depends on population geometry, leader quality and agent location. Two supporting mechanisms round out the design: a cosine-based nonlinear convergence schedule replaces the conventional linear decrease of the exploration-exploitation factor, and a Cauchy mutation operator—triggered when population diversity, measured by average Hamming distance in the binary feature-selection space, falls below a threshold of 0.2—injects occasional large jumps that help the pack escape local optima thanks to the Cauchy distribution’s heavy tails.

On the complete CEC2022 single-objective benchmark suite of 12 functions at 10 and 20 dimensions, ML-WSA achieved the lowest overall average rank of 1.63 among recent GWO variants, ahead of MSGWO at 2.38, although the advantage was not uniform across all 24 function-dimension cases. In the deception task itself, an optimizer-controlled comparison under identical feature representations and budgets showed ML-WSA improving F1 by roughly 3.17 percentage points over MSGWO, the strongest competing variant, while reaching the convergence criterion in 89 iterations versus 96. The extra leadership machinery carries measurable overhead: average runtime per optimization run was 68.6 seconds, an 11.0% increase over standard GWO, and peak memory rose from 1,048 to 1,218 megabytes, though faster convergence reduced fitness evaluations from about 6,400 to 4,700.

The text representation feeding this optimizer is deliberately two-channelled. A semantic channel uses a pre-trained BERT model, taking the hidden state of the [CLS] token as a 768-dimensional sentence embedding that captures contextual and discourse-level patterns. A statistical channel encodes 42 theory-linked linguistic features grounded in cognitive-load, reality-monitoring and interpersonal deception theories: pronoun use, hedging, sentiment polarity, passive constructions, sentence complexity and related surface cues. The two vectors are concatenated and passed through a learnable nonlinear fusion layer. ML-WSA then performs binary feature selection over the fused representation, encoding each candidate feature subset as a wolf’s position and scoring it with a fitness function that balances the F1 score on inner validation folds against subset compactness, with the best trade-off found when the selected subset was about 35% of the original dimensionality.

Class imbalance posed a serious obstacle: only 187 of the 1,710 statement units were labeled as evidence-supported misrepresentation, against 1,523 corroborated statements. Conventional classifiers tend to bias toward the majority class in such settings, so the researchers introduced a semantic-constrained oversampling strategy that generates synthetic minority samples only where embedding-space similarity confirms semantic consistency, reducing the risk of noisy or contradictory synthetic text. Ablation results underscore how much this matters: removing the oversampling component dropped recall from 0.8663 to 0.7623, an absolute reduction of 10.40 percentage points, and shifted deceptive-statement prediction probabilities toward the 0.3–0.5 range, below the default decision threshold of 0.5.

The ablation study also quantified each component’s contribution. Removing the regional multi-leadership mechanism caused the largest optimizer-specific drop, 5.77 percentage points of F1, and slowed feature-selection convergence, with final fitness falling from about 0.8960 to 0.8365. Adaptive leader weighting, nonlinear convergence control, Cauchy mutation and the statistical channel accounted for smaller reductions of 3.72, 3.10, 2.21 and 2.66 percentage points respectively, and the authors note these conditional effects are not additive because components interact. Without any feature selection, a BERT-only configuration reached an F1 of 0.8064, rising to 0.8231 with the statistical channel and to 0.8733 with ML-WSA selection. Across 30 random initializations, the F1 coefficient of variation was just 1.37%, and 47 features were selected in at least 80% of runs, indicating stable core feature sets.

The framework was stress-tested in several directions. Cross-dataset evaluation on three heterogeneous public deception benchmarks—Deceptive Opinion Spam, a fake news corpus and the Bluff dataset—yielded a macro-average F1 of 0.8067, which the authors present as evidence of robustness under domain shift rather than independent validation in the executive setting, since those corpora differ in communicative purpose, audience and verification mechanisms. Adversarial testing with TextFooler, which substitutes synonyms while preserving meaning, showed F1 declining by 7.0% at a 20% word-substitution budget and 12.4% at 30%, with the statistical channel providing partial resistance because features like sentence-length distributions and syntactic complexity survive word-level attacks better than semantic embeddings. A subgroup fairness audit found non-native English speakers initially received lower recall and F1; class-balanced reweighting and threshold calibration improved non-native F1 from 0.8426 to 0.8574 and narrowed the equal-opportunity difference from 0.063 to 0.041, though a disparity remained.

Interpretability was addressed through post hoc SHAP and LIME explanations, which identified the most stable transparent predictors as uncertainty hedge count, first-person pronoun ratio, negative emotion score, average sentence length, passive voice frequency and hedging word density. These outputs are intended to help investigators understand why a statement was prioritized, not to serve as automatic evidence of deception. The authors are explicit that the system is a risk-assessment and prioritization tool for human-supervised screening, not an autonomous judge of honesty: the deceptive label denotes evidence-supported misrepresentation established through a three-stage verification protocol with inter-rater agreement of Cohen’s kappa = 0.82, and psychological intent cannot be directly observed from archival records. They warn specifically about automation bias, recommend that model outputs never directly trigger adverse employment decisions, and call for investigator override authority, documented audit trails and local validation before deployment.

Limitations remain substantial. The executive corpus derives from a single archival source of 286 cases with limited organizational diversity, and the original agency records cannot be released because of confidentiality and re-identification risks, though anonymized shareable materials and implementation code are publicly available. The fairness analysis covered only language-background groups, not sex, age or other demographic dimensions, and the text-only design ignores vocal, facial and interactional cues. External validation on independent executive datasets from different organizations, industries and countries, along with prospective studies of how investigators actually use the model’s flags, will be needed before the framework could be considered mature for routine hiring or promotion screening. For now, the results suggest that reorganizing how a swarm of digital wolves shares leadership can meaningfully sharpen the search for the rare, carefully worded statements that deserve a human investigator’s closest attention.

Subject of Research: A multi-leadership wolf search optimization model for text-based deception-risk assessment in executive background checks

Article Title: Wolf-pack intelligence with multi-leadership Wolf search hybrid model for text-based deception-risk assessment in executive background checks

Article References: Wang, Y., & Zhang, Y. (2026). Wolf-pack intelligence with multi-leadership Wolf search hybrid model for text-based deception-risk assessment in executive background checks. Discover Artificial Intelligence, 6(1), Article 1620. https://doi.org/10.1007/s44163-026-02575-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02575-4

Keywords: deception detection, Grey Wolf Optimization, swarm intelligence, executive background checks, natural language processing, BERT, feature selection, class imbalance, machine learning, text classification, algorithm design, risk assessment