PaperlayerView exact revision on arXiv

Source-linked AI summary

"Hot-Blooded" vs "Cold-Blooded": Simulating the Behavioral Phenotypes of Childhood Aggression via Generative Agents

Liping Fu

arXiv:2610.11951v1cs.HC

TL;DR

Research on childhood aggression faces ethical and observational limits, while LLM simulations require tests of construct validity beyond surface-level behavior. This study evaluated theory-driven agents and found distinct, reliable subtype patterns across simulations, while noting their greater scenario consistency than human interview data.

  • Problem

    Studying reactive and proactive aggression is constrained by ethical challenges, under-detection, and questionnaire biases, and LLM simulations need construct-validity tests beyond face validity.

    Source: Introduction · p. 3

  • Method

    The study tested theory-driven agents simulating aggression subtypes and a non-aggressive baseline using quantitative indicators, cross-seed reliability, and qualitative narratives.

    Source: Discussion · p. 21

  • Results

    Across 1,920 simulated interactions, aggression subtypes and the non-aggressive baseline showed distinct quantitative patterns, high cross-seed reliability, and qualitative convergence.

    Source: Discussion · pp. 21–22

  • Takeaways & Limitations

    Theory-informed LLM agents may serve as a complementary, safe, controllable, and scalable environment for hypothesis generation, measurement refinement, and preliminary intervention evaluation.

    Source: Conclusion · p. 24

    Agents differentiated all eight scenarios, whereas human interviews discriminated adequately in only five, suggesting persona conditioning may over-amplify subtype boundaries.

    Source: Limitations · p. 23

Abstract

from arXiv · show

This study examines the construct validity of LLM-based generative agents in simulating reactive, proactive, and co-occurring aggression in children. Four distinct agents were instantiated using a theory-driven parameterization grounded in the social information processing model. A total of 1,920 simulation runs were conducted across eight social scenarios, employing a hybrid blind-coding pipeline to extract 32 quantitative behavioral indicators. Results demonstrate robust discriminant validity relative to a non-aggressive baseline, with large effect sizes. High cross-seed reliability confirms that behavioral differentiation is driven by underlying psychological parameters rather than model stochasticity. Qualitative narrative analyses further converged with established empirical literature. Overall, these findings indicate that theory-parameterized LLM agents can accurately reproduce distinct aggression subtypes, offering a scalable, highly controllable framework for hypothesis generation, intervention piloting, and the refinement of psychological measurement tools.

Introduction

Reactive and proactive aggression have distinct social-information-processing mechanisms, but experimental and observational methods face ethical, detection, and reporting constraints. This study tests whether theory-grounded LLM agents reproduce these profiles reliably and distinguishably.

Introduction

  • Existing methods struggle to study provocation ethically, detect covert proactive aggression, and avoid questionnaire recall and social-desirability biases.The study frames construct-specific validation as necessary because prior LLM simulations often rely on surface-level outputs rather than underlying mechanisms.Source: Introduction · p. 3, LLM-Based Agents · p. 4, The Present Study · p. 4
  • Reactive aggression involves hot, impulsive responses and hostile intent attribution, whereas proactive aggression uses aggression instrumentally to pursue goals.The SIP model describes five stages from encoding social cues through enacting behavior; the two subtypes are associated with distortions at different stages.Source: Introduction · pp. 3–4
  • Four child agents represented reactive, proactive, co-occurring, and non-aggressive profiles, with SIP-based parameters guiding role prompts and emotion updates.Parameters were theory-driven ordinal boundaries on a 0-to-1 continuum, not empirical epidemiological estimates; trait descriptions remained invariant across scenarios and trials.Source: The Present Study · p. 5
  • Eight vignettes targeted intent attribution and retaliation, instrumental resource acquisition, and cognitive or affective empathy across distinct social situations.Scenarios 1, 5, and 7 focused on ambiguous intent; scenarios 2, 6, and 8 on resource acquisition; scenarios 3 and 4 on empathy.Source: The Present Study · p. 9, Table 2 · p. 10
  • Each agent completed four-turn interactions across eight scenarios, with three replications and 20 random seeds yielding 1,920 simulation runs.The peer was independently generated as a typically developing eight-year-old classmate, with responses conditioned on the focal agent’s preceding behavior.Source: Procedure · p. 10

Data Analysis

The analysis extracted 32 behavioral indicators and classified them by measurement type so that each could be analyzed appropriately.

Data Analysis

  • 32 quantitative indicators were extracted from generated narratives and grouped as continuous ratings, counts, or derived composites.The table presents the indicator set; categorical variables were excluded because they are nominal.Source: Data Analysis · p. 11, Table 3 · p. 12
  • 26 indicators were 0–10 ratings, four were candy-allocation counts, and two were derived measures: aggressive action and attribution shift.Attribution shift was initial hostile attributional bias minus final accusation; aggressive action combined multiple behavior ratings.Source: Data Analysis · p. 11
  • Continuous ratings were analyzed with parametric tests, counts were summarized descriptively, and composites were interpreted as difference scores.The count indicators captured packs allocated to oneself, Leo, and two teammates.Source: Data Analysis · p. 11

Thirty-two Quantitative Indicators

The 32 indicators span ratings of behavioral and psychological dimensions, reward-allocation counts, and two composite measures.

Thirty-two Quantitative Indicators

Cross-Round Reliability

Cross-round reliability varied by indicator: 21 indicators had excellent ICC reliability, while several counts and temptation intensity performed poorly.

Cross-Round Reliability

  • 21 indicators achieved excellent reliability; affective empathy and prosocial behavior each had ICC = 0.99.Dominance had ICC = 0.97, and care had ICC = 0.96.Source: Cross-Round Reliability · p. 12, Table 4 · p. 12
  • Four indicators had good reliability, three moderate reliability, and four poor reliability; intensity of temptation had ICC = 0.08.The poor-reliability reward counts were packs for Leo (ICC = 0.39), Teammate B (ICC = 0.17), and Teammate C (ICC = 0.11).Source: Cross-Round Reliability · p. 12
  • Reliability used two-way random-effects, absolute-agreement, single-measurement ICCs, with random seed as rater and scenario-by-subtype as subject.ICC values of at least 0.75 were classified as excellent; 0.60–0.74 good, 0.40–0.59 moderate, and below 0.40 poor.Source: Cross-Round Reliability · p. 12

Cross-round reliability of quantitative measures (rater = seed)

Cross-seed reliability varied across indicators and behavioral profiles: 21 indicators were classified as excellent, while stability also varied by subtype and scenario.

Cross-round reliability of quantitative measures (rater = seed)

  • 21 indicators achieved excellent cross-round reliability, led by affective empathy and prosocial behavior (ICC = 0.99 for each).Dominance had ICC = 0.97, and care had ICC = 0.96.Source: Cross-round reliability of… · p. 13
  • ICC values were 0.72 for anticipated guilt, 0.71 for impulsivity, 0.61 for outcome evaluation, and 0.61 for number of packs for self.Source: Cross-round reliability of… · p. 13
  • ICC values were 0.59 for cognitive empathy, 0.47 for concern for consequences, and 0.46 for attribution shift.Source: Cross-round reliability of… · p. 13
  • ICC values were 0.39 for number of packs for Leo, 0.17 for number of packs for Teammate B, 0.11 for number of packs for Teammate C, and 0.08 for intensity of temptation.Source: Cross-round reliability of… · p. 13
  • CV was calculated across 26 continuous 0–10 indicators within each subtype-by-scenario cell; lower CV indicates greater stability across replications.The coefficient of variation was calculated as SD/M.Source: Cross-round reliability of… · p. 13
  • Reactive aggression was the most consistent subtype, while proactive and co-occurring aggression showed greater variability, especially in Scenarios 1, 3, and 4.Reactive-agent CVs ranged from 0.19 to 0.44; proactive-agent CVs reached 1.26 in Scenario 1 and 1.95 in Scenario 4.Source: Cross-round reliability of… · p. 13

Discriminant Validity

The four agents displayed contrasting profiles across aggression, empathy, and prosociality indicators, with eight selected measures chosen to represent core domains and support readable comparison.

Discriminant Validity

  • The eight radar-chart indicators span emotional arousal, instrumental motivation, empathy, and prosociality; the remaining measures were excluded for scenario specificity, measurement type, or instability.The selection also prioritized statistical robustness and visual clarity.Source: Discriminant Validity · p. 14
  • Alex combined elevated anger (M = 7.32) with relatively high affective empathy (M = 8.20) and prosocial behavior (M = 8.55).Alex’s retaliatory motivation was M = 7.43.Source: Discriminant Validity · p. 15
  • Bailey showed low anger (M = 1.36) and affective empathy (M = 0.43), alongside high selfishness (M = 6.87) and dominance (M = 7.52).Source: Discriminant Validity · p. 15
  • Casey combined elevated anger (M = 5.27) and retaliation (M = 7.58) with the highest selfishness (M = 8.22) and lowest care (M = 0.38) and prosocial behavior (M = 0.30).Source: Discriminant Validity · p. 15
  • David scored near baseline on aggressive constructs and near ceiling on care (M = 9.23), affective empathy (M = 9.03), and prosocial behavior (M = 9.50).Source: Discriminant Validity · p. 15

Divergent Validity

Comparisons with the non-aggressive baseline distinguished the aggression profiles on the selected indicators, with proactive aggression’s anger and retaliation as exceptions.

Divergent Validity

  • All baseline comparisons were statistically significant except proactive aggression on retaliation (p = .086) and anger (p = 1.00).The comparisons used ten indicators selected for their theoretical relevance, continuity, and adequate reliability.Source: Divergent Validity · p. 16, Table 4 · p. 17
  • Alex scored higher than David on dominance (M = 7.15 vs. M = 1.37), retaliation (M = 7.43 vs. M = 0.08), and anger (M = 7.32 vs. M = 1.33).Alex also scored lower than David on forgiveness, help, care, prosocial behavior, and both empathy measures.Source: Divergent Validity · p. 16, Table 4 · p. 17
  • Bailey scored higher than David on dominance (M = 7.52) and selfishness (M = 6.87), and lower on prosocial behavior (M = 0.57) and affective empathy (M = 0.43).Bailey did not differ significantly from David on retaliation or anger.Source: Divergent Validity · p. 16, Table 4 · p. 17
  • Casey scored higher than David on dominance (M = 8.38), selfishness (M = 8.22), retaliation (M = 7.58), and anger (M = 5.27).Casey scored lower than David on prosocial behavior, affective empathy, care, forgiveness, help, and cognitive empathy.Source: Divergent Validity · p. 16, Table 4 · p. 17

Comparisons of subtypes of aggression to the non-aggressive baseline

Narrative and quantitative results described distinct profiles: reactive aggression paired anger and retaliation with empathy, proactive aggression showed low empathic concern, and co-occurring aggression showed broad deficits in care and prosociality.

Comparisons of subtypes of aggression to the non-aggressive baseline

  • In Scenario 3, Alex expressed affective empathy alongside uncertainty about how to respond, consistent with no significant difference from baseline on affective empathy.Source: Prosocial Behavior · p. 19
  • Proactive aggression showed strong negative effects for prosocial behavior (d = −17.80) and affective empathy (d = −13.39), with near-zero anger (d = 0.02) and retaliation (d = 0.57).Bailey’s narratives included little empathic concern and attention to peers’ reactions.Source: Prosocial Behavior · p. 19
  • Co-occurring aggression showed very large positive effects for dominance (d = 8.03) and selfishness (d = 7.94), and negative effects for prosocial behavior (d = −18.37), affective empathy (d = −12.96), and care (d = −11.11).Casey’s narratives depicted immediate escalation and dismissal of others’ distress.Source: Prosocial Behavior · p. 19
  • David’s narratives showed attention to others’ feelings and helping behavior, alongside highest care, prosocial behavior (M = 9.50), and affective empathy (M = 9.03).Source: Prosocial Behavior · p. 19
  • Reactive aggression showed strongest positive effects for retaliation (d = 8.27) and anger (d = 4.18), with smaller negative effects for prosocial behavior (d = −1.15) and empathy.The narrative also depicted emotional arousal and hostile attribution in Scenario 1.Source: Prosocial Behavior · p. 18, Figure 3 · p. 20

Scenario Discrimination Power

Discrimination varied by scenario: forgiveness and prosocial behavior had the largest reported ranges, while empathy in Scenario 8 had the smallest, though it exceeded the scale midpoint.

Scenario Discrimination Power

Discussion

Across 1,920 simulated interactions, aggression subtypes and the non-aggressive baseline showed distinct quantitative patterns aligned with the SIP model, with high cross-seed reliability and convergent qualitative narratives.

Discussion

  • Across 1,920 simulated interactions, aggression subtypes and the non-aggressive baseline showed distinct quantitative patterns aligned with the SIP model.Cross-seed reliability was extremely high for most core dimensions, and qualitative excerpts supported the patterns.Source: Discussion · p. 21
  • The reactive agent showed high anger and retaliation while retaining empathy and prosocial behavior comparable to baseline, consistent with a “hot-blooded” profile.The authors interpret this dissociation as reflecting emotional-regulation and social-cue interpretation deficits rather than lack of empathy (Dodge & Crick, 1990; Euler et al., 2017).Source: Discussion · pp. 21–22
  • The proactive agent combined high dominance and selfishness with near-zero anger and deficits in affective empathy, while the co-occurring agent showed the most pervasive and severe deficits.The co-occurring profile combined both aggression subtypes and was described as distinct rather than merely additive.Source: Discussion · p. 22
  • Neutral, theory-blind coding of persona-conditioned agent behavior was used to mitigate confirmation bias in the hybrid blind-coding pipeline.Source: Discussion · p. 22
  • High cross-seed reliability supports reproducibility, while thousands of consistent trials offer a scalable way to examine rare or extreme behavioral profiles.Source: Discussion · p. 22
  • Neutral scenarios were less effective than affective contexts at eliciting aggression, and agents over-discriminated between subtypes under neutral conditions compared with human interviews.Human interviews did not differentiate subtypes in neutral scenarios, where socially acceptable responses may constrain disclosure (Fung et al., 2026; Fung et al., 2023).Source: Discussion · pp. 22–23
  • Generative agents could support vignette design, in silico intervention piloting, and synthetic benchmarks that supplement—but do not replace—real-world trials.The authors also propose extending the framework to multi-agent simulations of peer bullying and other group behaviors.Source: Implications · p. 23

Limitations

The authors identify boundaries in model-family generalizability, alignment with human scenario responses, and ecological validity; they conclude that simulation complements rather than replaces empirical research.

Limitations

  • The study used only DeepSeek-V3, limiting generalizability across LLM architectures and leaving consistency across model families to be tested.Source: Limitations · p. 23
  • Agents differentiated all eight scenarios, but human interviews showed adequate discrimination in only five, suggesting persona conditioning may amplify discrete subtype boundaries.The authors call for benchmarking against larger, multi-modal human datasets to calibrate agent sensitivity.Source: Limitations · p. 23
  • The study relied on hypothetical vignettes, which may not capture the real-time, emergent complexity of dynamic social interactions.Proposed future assessments include interactive virtual interviews, standardized psychometric surveys, and longitudinal multi-agent sandboxes.Source: Limitations · pp. 23–24

Conclusion

  • The authors frame these agents as a safe, controllable, scalable complement for hypothesis generation, measurement refinement, and preliminary intervention evaluation—not a substitute for traditional empirical research.Source: Conclusion · p. 24
Loading 2610.11951v1…