Taking another look at intelligence and personality using an eye-tracking approach

This study examines how personality traits and eye-movement behavior during test-taking relate to and predict fluid intelligence test performance. The core questions were: (1) How are Big Five personality traits associated with fluid intelligence scores (hypothesizing positive links for Openness and negative for Neuroticism)? (2) Do gaze patterns during a speeded intelligence test predict test performance? (3) Are personality traits reflected in gaze patterns, especially those indicative of higher performance? (4) What test-solving strategies can be inferred from gaze data (e.g., constructive matching vs. response elimination) under timed conditions? (5) Do personality traits (and facets) add unique predictive value beyond gaze patterns? The study pairs traditional measures (NEO PI-R; CFT 20-R) with eye-tracking to capture behavioral processes in a timed testing context that resembles applied assessment settings.

Intelligence is commonly conceptualized as general mental capability, often distinguishing fluid intelligence (novel problem solving) from crystallized intelligence (acquired knowledge). Personality captures enduring patterns of thoughts, feelings, and behaviors and is frequently organized via the Big Five (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism). Prior meta-analytic work shows Openness correlates most with fluid intelligence (e.g., r≈.17), with Neuroticism showing small negative correlations (r≈−.10); associations for Extraversion, Agreeableness, and Conscientiousness are weaker and inconsistent. At the facet level, stronger links emerge, particularly for Openness facets (Ideas, Values, Fantasy) and small negative links for several Neuroticism facets. Complementary research on test-taking processes uses eye-tracking to relate gaze to intelligence test performance. For matrix reasoning and similar tasks, two primary strategies have been studied: constructive matching (mentally construct the solution before searching options) and response elimination (switching between the stem and options to rule out distractors). Higher-ability individuals tend to employ constructive matching more. Prior eye-tracking studies often used unspeeded tests; it remains unclear how findings generalize to speeded tests common in applied settings. Gaze patterns may index underlying cognitive processes and, to some extent, can be trained as strategies that influence performance. Links between personality and gaze during intelligence testing are underexplored; Openness might relate to adaptive gaze patterns, whereas Neuroticism could relate to less beneficial gaze. The study addresses these gaps under timed conditions and with machine learning analyses.

Design and sample: Secondary analyses of the TüEyeQ dataset enriched with personality data. Initial N=315 university students with a university entrance qualification; after predefined quality exclusions for eye-tracking (tracking rate <80%, presentation errors, incomplete data, large calibration offsets) and data linkage issues, final N=182 (mean age 23.32, SD 2.89; 71.82% women; ~96% native German speakers). Ethics approval obtained; informed consent provided. Apparatus: SMI RED eye-trackers at 250 Hz on a 17-inch 1920×1080 laptop; 9-point calibration before each task; stable lighting; mouse or touchpad allowed. Measures: Fluid intelligence via CFT 20-R (figural tasks), using Blocks 1 (series continuation), 3 (matrices), and 4 (topological conclusions); Block 2 (classification) excluded due to AOI resolution constraints. Blocks used test manual time limits (3–4 minutes each). Each item had five response options; scores per block: Block 1 and 3 up to 15 each; Block 4 up to 11; combined maximum 41. Personality via German NEO-PI-R (240 items; 5-point Likert), providing Big Five traits and 30 facets with good internal consistency (traits α≈.85–.94; facets α≈.63–.90). Eye-tracking preprocessing: For each item, stimuli were segmented into semantic Areas of Interest (AOIs): question/stem, target, distractors, and other (e.g., clock, progress bar). Scanpaths were temporally normalized to focus on strategic sequence rather than duration, then binned into 50 equal time bins. Gaze similarity and prediction: A global sequence alignment approach (variation of Needleman–Wunsch) computed pairwise scanpath similarity over AOI sequences with simple match/mismatch/gap scoring. For predicting item-level correctness probabilities, a k-nearest-neighbors procedure (k=5) weighted neighbors by scanpath similarity. Item-level probabilities were aggregated to participant-level predicted CFT scores. Cross-validation: Leave-one-participant-out (LOPO) cross-validation for gaze-based models to avoid overlap between training and test participant data at the item level. Personality-based models used 10-fold cross-validation. Machine learning models: Elastic net regression (with CV-tuned regularization) used to predict CFT scores from (a) Big Five traits; (b) 30 facets; and (c) combined inputs (gaze-based predictions plus traits or facets). Feature scaling applied before model fitting. Predicting personality from gaze: The same similarity-based approach averaged across items to predict Big Five traits and, in additional analyses, facets; tested with all items and with only items whose gaze similarity significantly predicted CFT above chance. Strategy extraction from gaze (unsupervised characterization): Iterative selection of the most representative scanpaths per item: using the similarity matrix, repeatedly identify the scanpath most frequently appearing among the top-11 nearest neighbors of others, mark it as representative, remove its neighbors, and repeat three iterations. This yielded up to three representative scanpaths per item, enabling qualitative mapping to constructive matching vs response elimination patterns without manual labeling. Robustness checks: Replications with standard multiple regression for RQ1; RQ1 re-run on full sample without eye-tracking exclusions; and a 90–10 random train–test split repeated 1000 times (following a published protocol), yielding explained variance differences of ≤.01 from main analyses.

- Personality predicting intelligence (RQ1): Big Five traits explained 3.18% of variance in CFT scores using elastic net (10-fold CV). Only Openness and Agreeableness had meaningful positive feature weights (~0.040 and ~0.034) and were significant in robustness multiple regression. Facet model explained 7.67% of variance; nonzero contributors included Openness: Ideas (0.0034), Values (0.0028), Fantasy (0.0013); Neuroticism: Self-Consciousness (0.0019); Extraversion: Assertiveness (0.0013); Agreeableness: Trust (0.0027), Compliance (0.0027). - Gaze predicting intelligence (RQ2): Gaze-based KNN with scanpath similarity (LOPO-CV) explained 35.91% of variance in CFT scores. Early, trivially easy items contributed little variance; mid-to-late items were most informative. - Gaze predicting personality (RQ3): Models explained essentially 0% variance in Big Five traits from gaze (including when restricting to performance-relevant items). At the facet level, only small amounts (>1%) were explained for a few facets with all items: Self-Consciousness (1.14%), Depression (1.84%), Assertiveness (2.41%), Competence (1.16%). Using only performance-relevant items: Assertiveness (4.27%), Self-Consciousness (1.10%), Aesthetics (1.79%), Fantasy (1.76%). - Strategy patterns from gaze (RQ4): Representative scanpaths suggested that sequences resembling constructive matching (e.g., question then directly to target with minimal question-option shuttling) were more successful than response elimination-like patterns (frequent transitions between question and multiple options). Effects were clearest in Block 1, present but less pronounced in Block 3, and most blurred in Block 4 due to higher difficulty and more transitions. Longer initial fixation on the question and fewer question-option transitions tended to associate with higher success. - Combined models (RQ5): Adding Big Five to gaze-based predictions yielded 37.50% explained variance (10-fold CV). Feature weights indicated gaze dominated (≈1.416), with small yet unique contributions from Openness (≈0.018) and Agreeableness (≈0.019). With facets plus gaze, explained variance was 38.02%; gaze weight ≈1.288; contributing facets: Openness (Ideas ≈0.0022; Values ≈0.0001), Agreeableness (Trust ≈0.0004; Compliance ≈0.0015). - Overall: Behavioral gaze data outperformed personality self-reports by more than an order of magnitude in predicting intelligence test performance; personality still provided small, independent added value.

Findings address the research questions as follows. RQ1: Personality relates modestly to fluid intelligence performance, with Openness and, unexpectedly, Agreeableness being the key contributors; at the facet level, Ideas and Values (Openness) and Compliance and Trust (Agreeableness) were most relevant, along with smaller roles for Self-Consciousness and Assertiveness. RQ2: Gaze patterns during a speeded test were strongly predictive of performance, highlighting the value of behavioral process measures. RQ3: Gaze patterns did not align with Big Five traits overall, suggesting that gaze during a constrained, timed cognitive task is largely not personality-loaded; small associations emerged for a few facets (e.g., Assertiveness, Self-Consciousness), particularly when focusing on performance-relevant items. RQ4: Representative scanpaths partially reflected established strategies; constructive matching-like patterns were generally more successful than response elimination-like patterns, though time pressure blurred clear strategy categorization. RQ5: Personality and gaze provided largely independent information; combined models showed that while gaze dominated prediction, personality added small but significant unique variance. The results underscore that objective, state-like behavioral measures collected in the test situation are highly proximal to performance and thus more predictive than broad trait self-reports, yet trait and facet measures still offer complementary insights into individual differences in intelligence test outcomes.

This study bridges research on intelligence, personality, and eye-tracking by demonstrating that gaze behavior during test-taking robustly predicts fluid intelligence performance and that personality—particularly Openness and Agreeableness and their facets (Ideas, Values, Compliance, Trust)—adds small, unique predictive value. Gaze-derived patterns partially map onto known strategies (constructive matching vs response elimination), though time pressure blurs strategies. Integrating behavioral process measures with traditional self-reports enhances prediction and offers mechanistic insight into test-taking. Future work should (a) compare speeded versus unspeeded tests; (b) examine broader personality frameworks and finer-grained facet/item markers; (c) assess generalizability across intelligence measures (including crystallized intelligence and CHC-aligned batteries); (d) incorporate additional behavioral and physiological signals (e.g., mouse trajectories, heart rate, EEG) and temporally sensitive models; (e) analyze trait-like stability of gaze strategies across items and contexts; and (f) explore how strategy training interacts with time constraints.

- Timed testing context without direct comparison to unspeeded versions limits generalizability of gaze–strategy findings. - Single-session eye-tracking; no assessment of day-to-day variability in gaze, cognition, or personality expression. - Convenience sample of young adults with university entrance qualification (predominantly students), not population-representative. - Modest sample size relative to analytic complexity, especially for facet-level models; multiple testing increases risk of Type I errors; findings require replication. - Personality effects on gaze may be suppressed by strict, time-pressured lab tasks; personality–gaze links could be stronger in real-world, dynamic, or high-stakes contexts, or in older samples. - Eye-tracking focused on figural fluid intelligence tasks (CFT 20-R Blocks 1, 3, 4); Block 2 excluded due to AOI constraints; results may not generalize to other test formats or broader CHC abilities. - Strategy extraction remained partially ambiguous under speeded conditions; representative scanpaths approximate but do not definitively label strategies. - Self-report trait measures capture broad, cross-situational tendencies, which are less proximal to immediate task performance than behavioral gaze measures.