Cognitive networks reconstruct mindsets about STEM subjects and educational contexts in almost 1000 high-schoolers, University students and LLM-based digital twins

Francesco Gariboldi; Emma Franchino; Edith Haim; Gianluca Lattanzi; Alessandro Grecucci; Massimo Stella

cs.CLFeb 16, 2026

Cognitive networks reconstruct mindsets about STEM subjects and educational contexts in almost 1000 high-schoolers, University students and LLM-based digital twins

Authors: Francesco Gariboldi, Emma Franchino, Edith Haim, Gianluca Lattanzi, Alessandro Grecucci, Massimo Stella

Organizations: Department of Psychology and Cognitive Science, University of Trento, Corso Bettini, 31, Rovereto, 38068, TN, Italy · Department of Physics, University of Trento, Via Sommarive, 14, Povo, 38123, TN, Italy · Department of Psychology and Communication, University of Bari, Via Scipione Crisanzio, 42, Bari, 70122, BA, Italy

Abstract

Attitudes toward STEM develop from the interaction of conceptual knowledge, educational experiences, and affect. Here we use cognitive network science to reconstruct group mindsets as behavioural forma mentis networks (BFMNs). In this case, nodes are cue words and free associations, edges are empirical associative links, and each concept is annotated with perceived valence. We analyse BFMNs from N = 994 observations spanning high school students, university students, and early-career STEM experts, alongside LLM (GPT-oss) "digital twins" prompted to emulate comparable profiles. Focusing also on semantic neighbourhoods ("frames") around key target concepts (e.g., STEM subjects or educational actors/places), we quantify frames in terms of valence auras, emotional profiles, network overlap (Jaccard similarity), and concreteness relative to null baselines. Across student groups, science and research are consistently framed positively, while their core quantitative subjects (mathematics and statistics) exhibit more negative and anxiety related auras, amplified in higher math-anxiety subgroups, evidencing a STEM-science cognitive and emotional dissonance. High-anxiety frames are also less concrete than chance, suggesting more abstract and decontextualised representations of threatening quantitative domains. Human networks show greater overlapping between mathematics and anxiety than GPT-oss. The results highlight how BFMNs capture cognitive-affective signatures of mindsets towards the target domains and indicate that LLM-based digital twins approximate cultural attitudes but miss key context-sensitive, experience-based components relevant to replicate human educational anxiety.

Figures & tables

Sample	Total	High anxiety	Low anxiety	Excluded
Experts	59
High Schoolers (South)	159
Physics Undergraduates	10
GPT-oss Physics Undergraduates	10
High Schoolers (North)	62	30	27	5
GPT-oss High Schoolers	62	28	29	5

Table 1: Groups and subgroups (by anxiety level) used to create behavioural forma mentis networks.

Study 1: STEM subjects	Study 2: Educational and research contexts (actors and places)
Mathematics	Professor
Statistics	Teacher
Physics	Scientist
Science	School
Computer Science	University
Research

Table 2: Target words analysed in each experiment, chosen as central to the main domains of interest of the two studies.

Figure 1: Examples of a behavioural forma mentis network ( $a$ ), a semantic frame (of the node Teacher , $b$ ) extracted from it, and an emotional flower ( $c$ ).

Sample	$N_{v}$	$N_{e}$	$l_{G}$	$d$	Hubs (top 1% degree in full network)
Experts	1489	2861	4.51	10	Science (65), Physics (54), System (53), School (51), Mathematics (47), University (47), Life (46), Network (44), Complex (38), Art (38), Chemistry (34), Model (31), Theory (31), Water (30), Biology (30)
H-S South	4149	10207	4.06	7	Physics (113), Mathematics (100), Machine (86), Discussion (82), Time (79), Together (79), School (77), Art (67), Renewal (65), Water (64), Maximum (63), Science (63), System (63), Program (62), Tree (61), Teacher (61), Self organization (61), Rock (60), Behavior (58), Transmission (58), Domination (58), Law (57), Graph (56), Society (56), Organization (55), Person (55), Chemistry (54), Situation (54), Gas (54), Rational (54), Skin (54), Study (53), Action (53), Universe (53), Mathematically (52), Discovery (52), Work (52), Life (52), Graduated (51), Populations (51), Cell (51), Come on (50), Substance (50), Quality (50), Technology (50)
H-Anx H-S	1207	2079	3.55	5	Anxiety (74), School (74), Mathematics (72), Life (70), Discovery (66), Research (64), Stress (63), Physics (62), Future (62), Art (62), Failure (58), Teacher (58), Grade (58)
L-Anx H-S	1189	1987	3.61	5	Challenge (68), Physics (64), Mathematics (62), Future (62), School (61), Research (59), Experiment (59), Data (58), Art (57), Life (57), Discovery (56), Stress (55), Science (55)
GPT H-Anx H-S	520	753	4.16	7	Average (37), Teacher (31), Degree (28), Behavior (27), Development (27), Scientist (27)
GPT L-Anx H-S	540	773	4.12	7	Average (42), Teacher (35), Behavior (30), Tests (30), Degree (29), Development (28)

Table 3: Sample-wise network features of the entire network and for the target semantic frames.

Figure 2: Jaccard similarity values between Mathematics and Anxiety semantic frames across samples.

Figure 3: Semantic frames and emotional flowers of Mathematics .

Figure 4: Semantic frames and emotional flowers of Statistics .

Figure 5: Semantic frames and emotional flowers of Computer science .

Figure 6: Semantic frames and emotional flowers of Physics .

Figure 7: Semantic frames and emotional flowers of Science .

Figure 8: Semantic frames and emotional flowers of Teacher .

Figure 9: Semantic frames and emotional flowers of Professor .

Figure 10: Semantic frames of the node Scientist . "Male" and "Female" refer to the Italian form of the word.

Figure 11: Semantic frames and emotional flowers of Research .

Figure 12: Semantic frames of School (top) vs University (bottom).

group	keyword	k	$\bar{x}-\hat{\mu}_{0}$	$Z$	Cohen’s $d$	$\bar{x}$	Cliff’s $\delta$
H-S South	Professor	17	0.65	2.62	0.67	3.65	0.41
GPT H-S H-anx	Computer Science	12	0.61	2.20	0.59	3.58	0.38
H-S South	Computer Science	15	0.47	1.81	0.48	3.46	0.31
H-S South	Research	13	0.47	1.66	0.49	3.48	0.29
H-S South	University	30	0.40	2.18	0.40	3.38	0.26
Experts	School	51	0.38	2.58	0.39	3.36	0.25

Table 4: Frames showing significantly higher concreteness (

|Z|>1.6449

\alpha=0.1

), ordered by decreasing

\bar{x}-\hat{\mu}_{0}

group	keyword	k	$\bar{x}-\hat{\mu}_{0}$	$Z$	Cohen’s $d$	$\bar{x}$	Cliff’s $\delta$
GPT Phy	Science	12	-0.48	-1.69	0.50	2.51	0.26
GPT H-S H-anx	Science	20	-0.46	-2.05	0.44	2.56	0.24
GPT Psy H-anx	Science	32	-0.44	-2.44	0.44	2.54	0.23
GPT H-S L-anx	Science	19	-0.40	-1.88	0.40	2.58	0.20
GPT Psy H-anx	Statistics	28	-0.39	-2.16	0.40	2.58	0.18
GPT Psy L-anx	Science	38	-0.32	-1.97	0.31	2.68	0.15

Table 5: Frames showing significantly lower concreteness (

|Z|>1.6449

\alpha=0.1

), ordered by decreasing

\bar{x}-\hat{\mu}_{0}

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	48	107	0.06	1.91	Mathematics (47), Physics (11), Science (10)
H-S South	101	473	0.08	1.91	Mathematics (100), Physics (32), Calculation (25), Numbers (24), Mathematically (22), Number (22)
H-Anx H-S	73	212	0.05	1.92	Mathematics (72), Physics (20), Numbers (17), Problem (17)
L-Anx H-S	63	183	0.06	1.91	Mathematics (62), Physics (18), Analysis (15), Numbers (14), Data (14)
GPT H-Anx H-S	25	33	0.03	1.89	Mathematics (24), Equation (8)
GPT L-Anx H-S	19	25	0.05	1.85	Mathematics (18)

Appendix

Table 6: Network features for the semantic frame of Mathematics .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	26	38	0.04	1.88	Statistics (25), Mathematics (5), Number (5)
H-S South	29	59	0.08	1.85	Statistics (28), Mathematics (12)
H-Anx H-S	37	68	0.05	1.90	Statistics (36), Analysis (12)
L-Anx H-S	41	91	0.07	1.89	Statistics (40), Analysis (14), Mathematics (12)
GPT H-Anx H-S	15	16	0.02	1.85	Statistics (14)
GPT L-Anx H-S	23	25	0.01	1.90	Statistics (22), Average (4)

Appendix

Table 7: Network features for the semantic frame of Statistics .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	27	40	0.04	1.89	Computers (26), Science (7)
H-S South	16	31	0.15	1.74	Computer science (15)
H-Anx H-S	51	90	0.03	1.93	Computer science (50), Programming (14), Innovation (8), Future (8)
L-Anx H-S	45	83	0.04	1.92	Computer science (44), Programming (12), Laboratory (8)
GPT H-Anx H-S	19	18	0.00	1.89	Computer science (18)
GPT L-Anx H-S	22	23	0.01	1.90	Computer science (21), Innovation (3)

Appendix

Table 8: Network features for the semantic frame of Computer science .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	55	108	0.04	1.93	Physics (54), Science (15), Mathematics (11)
H-S South	114	435	0.05	1.93	Physics (113), Mathematics (32), Science (24), Physical (18), School (18), Come on (17), Electron (17)
H-Anx H-S	63	157	0.05	1.92	Physics (62), Mathematics (20), Laboratory (13), Experiment (12), Research (12)
L-Anx H-S	65	158	0.05	1.92	Physics (64), Mathematics (18), Science (15), Research (14)
GPT H-Anx H-S	25	28	0.01	1.91	Physics (24), Science (4)
GPT L-Anx H-S	23	23	0.00	1.91	Physics (22), Theory (2), Science (2)

Appendix

Table 9: Network features for the semantic frame of Physics .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	66	156	0.04	1.93	Science (65), Physics (15), System (12), Complex (11)
H-S South	64	242	0.09	1.88	Science (63), Physics (24), Biology (18), Study (17), Mathematics (17)
H-Anx H-S	48	110	0.06	1.90	Science (47), Scientist (14), Discovery (14)
L-Anx H-S	56	162	0.07	1.89	Science (55), Research (20), Scientist (17)
GPT H-Anx H-S	23	30	0.03	1.88	Science (22), University (5)
GPT L-Anx H-S	22	24	0.01	1.90	Science (21), Innovation (3)

Appendix

Table 10: Network features for the semantic frame of Science .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	15	24	0.11	1.77	Research (14)
H-S South	13	25	0.20	1.68	Research (12)
H-Anx H-S	65	168	0.05	1.92	Research (64), Discovery (23), Scientist (15), Laboratory (13)
L-Anx H-S	60	189	0.08	1.89	Research (59), Science (20), Laboratory (17)
GPT H-Anx H-S	11	13	0.07	1.76	Research (10)
GPT L-Anx H-S	12	14	0.05	1.79	Research (11)

Appendix

Table 11: Network features for the semantic frame of Research .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	4	3	0.00	1.50	Professor (3)
H-S South	17	30	0.12	1.78	Professor (16)
H-Anx H-S	3	2	0.00	1.33	Professor (2)
L-Anx H-S	2	1	0.00	1.00	Professor (1), Teacher (1)
GPT H-Anx H-S	17	22	0.05	1.84	Professor (16)
GPT L-Anx H-S	14	15	0.03	1.84	Professor (13)

Appendix

Table 12: Network features for the semantic frame of Professor .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	16	27	0.11	1.77	Teacher (15)
H-S South	62	159	0.05	1.92	Teacher (61), School (23), Student (16), Classroom (12)
H-Anx H-S	2	1	0.00	1.00	Teacher (1), Teacher (1)
L-Anx H-S	2	1	0.00	1.00	Teacher (1), Teacher (1)
GPT H-Anx H-S	32	37	0.01	1.93	Teacher (31), Professor (6)
GPT L-Anx H-S	36	40	0.01	1.94	Teacher (35), Knowledge (3), Curiosity (3), Professor (3)

Appendix

Table 13: Network features for the semantic frame of Teacher .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	10	9	0.00	1.80	Scientist (9)
H-S South	32	98	0.14	1.80	Scientist (31), Study (13)
H-Anx H-S	43	130	0.10	1.86	Scientist (42), Laboratory (17), Scientist (17)
L-Anx H-S	46	105	0.06	1.90	Scientist (45), Science (15), Laboratory (14), Experiment (14)
GPT H-Anx H-S	28	30	0.01	1.92	Scientist (27), Curiosity (4)
GPT L-Anx H-S	26	30	0.02	1.91	Scientist (25), University (4)

Appendix

Table 14: Network features for the semantic frame of Scientist (male) .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	10	9	0.00	1.80	Scientist (9)
H-Anx H-S	37	96	0.10	1.86	Scientist (36), Scientist (17)
L-Anx H-S	45	111	0.07	1.89	Scientist (44), Science (17), Research (16)
Physics	29	40	0.03	1.90	Scientist (28), Laboratory (6), Research (6)
H-Anx Psy	18	31	0.10	1.80	Scientist (17)
L-Anx Psy	24	38	0.06	1.86	Scientist (23), Science (8)

Appendix

Table 15: Network features for the semantic frame of Scientist (female) .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	52	101	0.04	1.92	School (51), University (14), Science (9)
H-S South	78	255	0.06	1.92	School (77), Teacher (23), Mathematics (21), Physics (18)
H-Anx H-S	75	219	0.05	1.92	School (74), Anxiety (19), Stress (14), Mathematics (14)
L-Anx H-S	62	128	0.04	1.93	School (61), Mathematics (13), Research (10), Teacher (10)
GPT H-Anx H-S	3	2	0.00	1.33	School (2)
GPT L-Anx H-S	2	1	0.00	1.00	Average (1), School (1)

Appendix

Table 16: Network features for the semantic frame of School .

Sample	$N_{v}$	$N_{e}$	$C_{i}$	$l_{G}$	Hubs (top 5% of degree distribution)
Experts	48	93	0.04	1.92	University (47), School (14), Science (9)
H-S South	36	110	0.13	1.83	University (35), Study (15)
H-Anx H-S	8	14	0.33	1.50	University (7)
L-Anx H-S	10	23	0.39	1.49	University (9)
GPT H-Anx H-S	25	35	0.04	1.88	University (24), Degree (8)
GPT L-Anx H-S	20	29	0.06	1.85	University (19)

Appendix

Table 17: Network features for the semantic frame of University .

Set	$N_{w}$	Administered to (human / GPT)
1	50	Experts; High Schoolers (South).
		10 core cues in fixed order (italic in Table 19 ) + 40 random cues
		drawn per participant from a 390-word Wikipedia-derived pool (Complex Systems/Physics/Math/Bio/Chem/Psych).
2	51	Psychology undergraduates subsample ( $n=57$ ).
3	40	Psychology undergraduates subsample ( $n=70$ ); all GPT-oss simulated students.
4	41	Psychology undergraduates subsample ( $n=166$ ).

Appendix

Table 18: Cue-word sets and their administration across samples/subsamples. Subgroup size is notated as

n

Set	STEM Disciplines	Education and Research	Evaluation	Learning and Motivation	Mental Health
1 ( $N_{w}=50$ )	Science, Physics , Mathematics , Chemistry , Biology , Statistics, Electricity, Computer, Evolution, Dna, Experiment, Cell, Derivative, Space, System , Graph, Theorem, Trigonometry, Integral, Model, Complex , Chaos, Solid, Number, Hub, Network, Theory	University , School , Conference, Lab, Studies, Teaching, Student, Degree, Mathematician, Prisoner	Analysis, Game, Word, Loop, Market	Language, Person	Life , Brain, Water, Scientific, Cloud, Art
2 ( $N_{w}=51$ )	Algorithm, Biology, Calculation, Demonstration, Equation, Physics, Engineering, Mathematics, Mathematician, Numbers, Performance in mathematics, Statistics, Technology, Math test, Utility of mathematics	Class, Parent, Teacher, Student, Board, Concept map, Notes	Exam	Change of plans, Task, Concentration, Confusion, Curiosity, Distraction, Errors, Set a goal, I commit, Importance, Intelligence, Store, Motivation, Plan, Problem, Plan, Program, Review, Postpone, Expiration, Challenge, Effort, Solution, Study, Success, Keep track	Anxiety, Failure
3 ( $N_{w}=40$ )	Biology, Equation, Physics, Informatics, Mathematics, Statistics, Science, Scientist, Innovation, Neuroscience	Teacher, Professor, University	Exam, Trial, Test, Average, Assessment	Creativity, Curiosity, Fun, Challenge, Model, Know, Development, Passion	Anxiety, Attachment, Well being, Cognition, Emotion, Mind, Personality, Psychology, Psychopathology, Psychotherapy, Therapist
4 ( $N_{w}=41$ )	Biology, Equation, Physics, Mathematics, Numbers, Statistics, Technology, Graphic, Science, Discovery, Neuroscience	Teacher, Student, Professor, School, University	Exam, Exercise, Grade	Task, Knowledge, Creativity, Curiosity, Challenge, Future, Model, Relation	Anxiety, Well being, Cognition, Emotion, Frustration, Mind, Boredom, Personality, Psychology, Therapy, Stress
5 ( $N_{w}=42$ )	Analyses, Equation, Physics, Informatics, Mathematics, Numbers, Statistics, Data, Hypothesis, Laboratory, Logic, Research, Science, Scientist, Scientist, Discovery, Experiment, Innovation	Companions, Teacher, Parents, School	Exam, Grade	Knowledge, Creativity, Problem, Programming, Challenge, Future, Rules	Anxiety, Art, Well being, Counseling, Failure, Embarrassment, Meditation, Panic, Fear, Stress

Appendix

Table 19: All the cue words sets used for the association task with all our samples (or sub-samples).

N_{w}

stands for the number of words present in each set.

Figure 13: Semantic frames of the node Math .

Figure 14: Semantic frames of the node Statistics .

Figure 15: Semantic frames of the node Computer Science .

Figure 16: Semantic frames of the node Physics .

Figure 17: Semantic frames of the node Science .

Cognitive Modeling

Cognitive modeling uses computational theories to explain and predict how people perceive, learn, remember, reason, and decide.

See all 119 papers

Mar 21, 2026

Subjective-Graph LLM Agents for Simulating Uncertainty in Classroom Social Perception

Social actors do not observe a common social world: each individual forms judgments from a partial and potentially distorted view of the surrounding network. We study whether graph-local evidence and credibility-weighted communication can generate persistent distortions in perceived academic standing, even when agents repeatedly receive objective performance signals. We introduce a data-constrained multi-agent framework in which LLM agents operate through individualized subjective graphs that determine peer visibility, evidence access, and interaction opportunities. Agents exchange uncertainty-annotated assessments, evaluate message credibility, and maintain explicit Gaussian belief states updated through Bayesian fusion. We evaluate the framework on 12 middle-school classrooms comprising 482 students, using questionnaire-derived social information and six consecutive examinations. On the Social-Observed subset (n=419), collective ranking error increases from 0.066 \pm 0.008 to 0.124 \pm 0.009 across six epochs despite repeated exam-based anchoring. Ablations associate individualized visibility and LLM-based trust gating with more stable long-horizon behavior, while constrained retrieval primarily safeguards against global-information leakage. Compared with evaluated DeGroot configurations, the proposed framework achieves lower final ranking error; those DeGroot configurations exhibit near-zero terminal opinion diversity. These findings establish subjective-graph LLM agents as a mechanism-oriented framework for data-constrained simulated social perception. Code is available at https://anonymous.4open.science/r/Rashomonomon-0126.

Jinming Yang, Xinyu Jiang, Xinshan Jiao +1

Complex Lab, School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China · Department of Computer Science, City University of Hong Kong, Hong Kong · School of Health and Medical Technology, Chengdu Neusoft University, Chengdu,2026 China

Multi-Agent LLM Systems Opinion Dynamics AI in Education Education Computational Social Science Agent-Based Modeling

Ask PDF

Aug 18, 2026

Interpretable Humans, Alien LLMs: Expert Analysis of Latent Structures in Assessment Responses

The evaluation of large language models (LLMs) relies heavily on human-designed assessments, implicitly assuming that AI and humans employ similar underlying cognitive constructs. Challenging this assumption, we investigate whether the latent factors governing LLM performance carry the same substantive, human-interpretable meaning as the cognitive constructs governing human learners. Using responses from humans and six LLMs across quantitative reasoning and chemistry assessments, we conducted Exploratory Factor Analysis (EFA) separately for both groups. Subject-Matter Experts (SMEs) then blindly evaluated the resulting factor graphs to ascribe pedagogical meaning to the emerged constructs. SMEs successfully interpreted most of the human-derived factors. Conversely, they could not ascribe meaning to any LLM-derived factors in quantitative reasoning and interpreted only half of the LLM factors in chemistry. By combining data-driven EFA with blind expert interpretation, this framework shows that LLMs frequently operate on statistically opaque mechanisms distinct from human reasoning.

Alona Strugatski, Licol Zeinfeld, Jason Cooper +3

Weizmann Institute of Science

LLM Evaluation LLM Interpretability Scientific Reasoning in Language Models

Ask PDF

Apr 30, 2026

Math Education Digital Shadows for facilitating learning with LLMs: Math performance, anxiety and confidence in simulated students and AIs

To enhance LLMs' impact on math education, we need data on their mathematical prowess and biases across prompts. To fill this gap, we introduce MEDS (Math Education Digital Shadows) as a dataset mapping how large language models reason about and report mathematics across human- and AI-like conditions. MEDS involves 28,000 personas from 14 LLMs (from families like Mistral, Qwen, DeepSeek, Granite, Phi and Grok) shadowing either humans or AI assistants. Each record/shadow includes a set of prompts along with psychological/sociodemographic persona metadata and four types of math tasks: (i) open math interview, (ii) three psychometric tests about math perceptions with explanations, (iii) cognitive networks capturing math attitudes, and (iv) 18 high-school math test questions together with their reasoning and confidence scores. MEDS differs from traditional score-only math benchmarks because it integrates concepts of self-efficacy, math anxiety, and cognitive network science besides math proficiency scores. Data validation shows that the sampled LLMs exhibit schema integrity and consistent personas, together with family-specific peculiarities like human-like negative math attitudes, logical fallacies, and math overconfidence. MEDS will benefit learning analytics experts, cognitive scientists, and developers of safer AI tutors in mathematics.

Naomi Esposito, Anthony Tricarico, Luisa Porzio +2

CogNosco Lab, University of Trento, Department of Psychology and Cognitive Science, Trento, Italy

Mathematical Reasoning Benchmarks Generative AI in Education

Ask PDF

Abstract

Cognitive Modeling

Explore similar work

Subjective-Graph LLM Agents for Simulating Uncertainty in Classroom Social Perception

Interpretable Humans, Alien LLMs: Expert Analysis of Latent Structures in Assessment Responses

Math Education Digital Shadows for facilitating learning with LLMs: Math performance, anxiety and confidence in simulated students and AIs