Exploring Causal Mechanisms with Generative Agent-Based Models

Method

• RePair calibrates a responsive baseline world, operationalizes mechanisms as natural-language rules, compares each rule with a matched baseline, and examines behavioral traces.The matched comparison holds the model, agent profiles, world description, initial conditions, and interaction schedule fixed.Source: §1 · pp. 1–2, Figure 1 · p. 2

RQ1

• Across four worlds, 12 of 23 candidate rules had 90% intervals entirely in their predicted direction.Each world contained at least one rule with an established effect in its predicted direction.Source: §1 · p. 2

RQ2

• The leading rule was recovered in every resample from eight configurations onward, and 99.98% of resamples preserved the ten rule-pair orderings distinguished by the full sample at 268 configurations.Agreement on every rule’s effect conclusion was 80.4% at 40 configurations and 94.2% at 268, because near-zero rules remain uncertain.Source: §1 · p. 2
• Rule effects can change with implementation: inequity-aversion wordings moved the outcome in opposite directions, and sentence position could change a rule’s effect.More configurations improve precision, while variation across models, wordings, and positions establishes the range of implementations to which a conclusion applies.Source: §1 · p. 3

RQ3

• Behavioral traces distinguish mechanisms with similar aggregate effects and can show when an outcome changes without the proposed social process supporting it.Discussion increased without more extreme group opinions in the polarization world, while cascade-world contribution effects appeared before agents observed partners.Source: §1 · p. 3

Contribution

• Generative ABMs can help screen and refine causal explanations before human experiments or observational evaluation, but do not replace external validation.Causal claims remain bounded to the specified simulation.Source: §1 · p. 3

2.1 EMERGENCE AND GENERATIVE EXPLANATION IN AGENT-BASED MODELS

• ABMs explain collective phenomena by showing how local rules and repeated interactions produce macro patterns, though similar outcomes can arise from different rules or parameterizations.Generative explanations show how a pattern can be “grown” from specified agents and interactions.Source: §2.1 · p. 3

2.2 EMERGENT SOCIAL BEHAVIOR IN LLM-AGENT SIMULATIONS

• LLM-agent simulations produce recognizable macro patterns, but the appearance of a pattern does not identify which candidate rule produced it.RePair shifts evidence to measured intervention effects, their stability across configurations, and traces of proposed actions, interactions, and timing.Source: §2.2 · p. 4

2.3 INTERACTIVE TOOLS FOR SYSTEMATIC EXPERIMENTATION

• Interactive simulation tools support rule manipulation and output comparison, while RePair addresses generative ABMs where interventions shape many agents’ decisions and repeated interactions.The evidence appears at both behavioral and collective levels.Source: §2.3 · p. 4

3 REPAIR: A METHOD FOR EXPLORING CAUSAL MECHANISMS

• The method treats a candidate mechanism as a local rule in a world specifying agents, observations, actions, interactions, and simulation time.The method overview frames the workflow as testing how hypothesized individual behavior produces a collective outcome within the simulation.Source: §3.1 · pp. 4–5, §3 · p. 4

3.2 STEP 1: CALIBRATE AND FIX THE BASELINE WORLD

• A usable baseline supports the actions and interactions required by the mechanism and leaves the outcome room to move in both directions, without requiring the target pattern to appear already.Researchers verify runs, vary one operation setting at a time with structural settings fixed, and freeze the calibrated settings before testing rules.Source: §3.2 · pp. 5–6, Figure 3 · p. 6

3.3 STEP 2: OPERATIONALIZE CANDIDATE MECHANISMS

• Candidate mechanisms become natural-language profile rules, and comparing multiple checked wordings helps assess whether an estimated effect extends beyond one sentence.The anchor preserves who holds the attitude, what the attitude is, and what it concerns; wording uncertainty measures sensitivity of the pooled estimate, not a way to select a preferred effect.Source: §3.3 · pp. 6–7, Figure 4 · p. 7

3.4 STEP 3: ESTIMATE AND COMPARE RULE EFFECTS

• Matched comparisons hold model, prompts, initial conditions, and scheduled interactions constant while adding the candidate rule only to the rule condition.The signed paired difference is averaged across configurations in the metric’s own units; pairing does not eliminate stochastic variation or rule-context interactions.Source: §3.4 · pp. 7–8

3.5 STEP 4: ASSESS EXPLANATORY SUPPORT FROM THE BEHAVIORAL PATHWAY

• Behavioral records test whether the named behavior occurs and whether interactions and outcomes follow a proposed pathway, helping distinguish a missing outcome effect from an unsupported behavioral link.Trace compatibility does not show that the proposed process is necessary or uniquely responsible; testing later social transmission may require an additional intervention on the information channel.Source: §3.5 · p. 8

4 EVALUATION DESIGN

• Four worlds grounded in established social-science models or experiments cover residential sorting, repeated contribution decisions, group discussion, and cooperation across changing partners.They provide complementary tests rather than four equivalent replications.Source: §4 · p. 8
• The residential-segregation study demonstrates the full workflow by testing whether matched interventions distinguish the direction, magnitude, and consistency of eight candidate-rule effects.The world is calibrated so households can act on relocation preferences and the departure-rate gap can move in either direction.Source: §4 · p. 8
• The public-goods study uses 268 configurations across four models to assess how effect conclusions and rankings change as configurations accumulate and with wording or sentence position.Agreement with the full-sample analysis measures stability, not accuracy against a known true ranking.Source: §4 · pp. 8–9
• Across all four studies, behavioral records distinguish changes in registered outcomes from changes in underlying behaviors and test whether proposed behaviors precede collective change.The focused polarization and cascade tests examine whether change begins before or after the proposed social interaction.Source: §4 · p. 9

5 OPERATIONALIZATION

• The study asks whether calibrated worlds support measurable comparisons among rules that could explain differential residential sorting.Candidate explanations include preference for same-group neighbors, expectations that similar neighbors will leave, and attachment to home and street.Source: §5 · p. 9

5.1 WORLD AND OUTCOME

• The primary outcome is the white-minus-Black departure-rate gap, measured in percentage points per period against the same simulated world without a candidate rule.The simulated block represents mid-century America; balanced white and Black arrivals mean differences in departure rates drive changes in its racial composition.Source: §5.1 · p. 9

5.2 CALIBRATING THE BASELINE WORLD

• Calibration retained the one-action setting because relocation decisions appeared in all six baselines, compared with one of six under the open-ended action budget.The world was checked for valid actions, stable group labels, executed moves, and an outcome able to move in both directions; four models passed screening.Source: §5.2 · pp. 9–10

5.3 REGISTERING AND TESTING CANDIDATE RULES

• Each candidate rule is appended to every household’s profile, and its effect is compared with a newly collected matched baseline across twelve configurations.The eight rules are divided between predictions to widen and narrow the gap; the initial 30% Black population share is not a numerical minority threshold imposed on white households.Source: §5.3 · p. 10, Table 1 · p. 11
• All four gap-amplifying rules widen the departure-rate gap; out-group prejudice has the largest mean effect but varies more across configurations than minority-status aversion.Own-group attraction primarily increases white departures, while expectation of flight has the smallest mean effect and is associated with earlier departures: period 7.1 versus 10.6 in the baseline. Integration preference narrows the gap to within one percentage point; place attachment has a smaller estimate whose interval reaches zero.Source: §5.3 · pp. 10–12, Figure 5 · p. 11
• Intergroup contact and collective efficacy do not establish effects on the gap, and traces help distinguish their behavioral patterns.Under intergroup contact, households direct 34% of core actions toward neighbors versus 43% in the baseline; collective efficacy lowers both groups’ departure rates by similar amounts.Source: Figure 5 · p. 11, §5.3 · p. 12

6.1 WORLD SETUP AND CALIBRATION

• The public-goods outcome is contribution frequency, but the calibrated baseline does not show the sustained cooperation reported in the human punishment condition.Contribution frequency declines from 88.6% in the first three periods to 79.2% in the last three, despite the available punishment stage.Source: §6.2 · p. 13, §6.1 · p. 13

6.2 RULE EFFECTS AND FULL-SAMPLE RANKING

• Money maximization is the strongest suppressor (CP = 2.90), while conditional cooperation is the strongest amplifier (CP = 0.59); inequity aversion establishes no effect and negative reciprocity acts opposite its registered prediction.Money maximization lowers contribution frequency by 78.6 percentage points; conditional cooperation raises it by 12.0 points. Table 4 reports the full-sample effects and rankings.Source: §6.2 · pp. 13–14, Table 4 · p. 14
• Behavioral responses and contribution outcomes can diverge: negative reciprocity adds 15.6 punishment acts while contribution frequency falls by 2.7 points.Across rules, effects differ by outcome; the leading rule is the same within each admitted model, while conclusions for the other four rules vary by model.Source: §6.2 · p. 14

6.3 STABILITY ACROSS CONFIGURATIONS

• The leading rule is recovered in every resample from eight configurations onward, and agreement on the full-sample ordering of distinguished rule pairs reaches 99.98% at 268 configurations.The full sample is a reference for stability, not a known true ranking.Source: §6.3 · pp. 14–15, Figure 6 · p. 15
• Recovering every rule’s effect conclusion takes more configurations than recovering the leading rule: all five conclusions match the full sample in 94.2% of resamples at 268.Disagreement primarily reflects inequity aversion, whose full-sample effect is near zero and whose resampled interval can cross the decision boundary.Source: §6.3 · p. 15

6.4 SENSITIVITY TO WORDING AND POSITION

• Rule effects can depend on sentence position and wording: conditional cooperation loses 7.9 percentage points when moved toward the profile opening, and inequity aversion loses 10.9.At the full wording pool, every rule is below the study’s wording-uncertainty target, but inequity aversion and removing the shadow of the future need nearly all wordings; inequity-aversion wordings can move the outcome in opposite directions.Source: §6.4 · pp. 15–16, Figure 7 · p. 16

7 EXPLANATORY SUPPORT

• The additional tests model group discussion and changing partners, comparing five candidate rules per world across five registered wordings.The cascade world fixes the table schedule and information channel; participants see prior-table contributions only between periods.Source: §7 · p. 16, §7.2 · p. 18

7.1 GROUP POLARIZATION

• The polarization world calibrates a four-person discussion with ratings from 0 to 3, leaving room for shifts in either direction without requiring baseline polarization.The baseline mean shift is −0.09 scale points across 24 configurations; the widest initial spread instead moved consensus 0.58 points toward the center.Source: §7.1 · pp. 16–17
• In the polarization world, conformity pressure lowers the opinion shift by 0.23 scale points and is the only rule with an established effect; none establishes increased polarization.Value expression raises discussion to 2.5 times baseline while ratings remain near the middle, separating discussion activity from opinion change.Source: §7.1 · p. 17

7.2 COOPERATIVE CASCADES

• In the public-goods cascade world, behavioral contagion raises contributions by 1.76 credits, descriptive norms by 0.92, and social loafing lowers them by 1.80; all three effects are established.Indirect reciprocity and the raised threshold do not establish effects.Source: §7.2 · p. 18
• All three supported cascade-world effects are present before participants see others’ contributions, so they do not establish that cooperation spreads across groups.Later social transmission remains possible, but requires a separate intervention on social information.Source: §7.2 · p. 18

7.3 FINDINGS ACROSS THE FOUR WORLDS

• Across all four worlds, 12 of 23 candidate rules have 90% intervals entirely in their registered direction, and each world has at least one established effect.Three worlds show established effects in both directions; these outcome shifts do not alone establish the proposed social process, and findings apply to the tested baseline worlds and configuration pools.Source: §7.3 · pp. 18–19
• Across the worlds, traces show that more discussion need not produce polarization and that contribution effects may precede exposure to others’ behavior.Outcome effects are evidence of rule effects, not by themselves evidence for the behavioral links proposed to explain them.Source: §7.3 · pp. 18–19

• Natural-language rules produce measurable collective effects across four worlds, but some conclusions stabilize while others change with wording or sentence position.Behavioral records also show that an outcome effect can occur without the social process proposed to explain it.Source: §8 · pp. 19–20
• Calibrate for the research question, then fix the world before rule testing rather than tuning it to reproduce an expected pattern.A baseline need not show the target phenomenon when testing whether a rule produces it, but must leave room for change when testing whether an existing pattern strengthens or weakens.Source: §8 · p. 20
• Treat rule wording and sentence position as design choices, and bound conclusions to the evaluated models and expressions.Some inequity-aversion wordings produce effects in opposite directions, and changing sentence position changes estimated effects.Source: §8 · p. 20
• Match the evidence budget to the intended conclusion and inspect the proposed behavior-to-outcome sequence alongside the registered outcome.A leading rule may stabilize before every effect conclusion does; a compatible trace supports interpretation but does not establish that a process is necessary or uniquely responsible.Source: §8 · p. 20

• The simulations do not establish that agents behave like their real-world counterparts.Results concern changes within the calibrated simulation world.Source: §9 · p. 20
• The four studies test individual mechanisms under fixed interaction structures; combined mechanisms, alternative population structures, and longer-term dynamics remain open.Source: §9 · p. 20
• Generative ABMs may reproduce assumptions and biases encoded in models, prompts, and world designs, especially when agents represent people or social groups.Simulation results can refine hypotheses but should not be treated as direct evidence about represented entities.Source: §9 · p. 20

• RePair calibrates and fixes a simulation world, operationalizes candidate mechanisms as natural-language rules, estimates effects through matched comparisons, and examines behavioral evidence.Source: §10 · p. 21
• Across four worlds, the method distinguishes candidate mechanisms by collective effects, assesses comparison reliability across implementations, and reveals when an outcome effect does not support the proposed process.Source: §10 · p. 21
• The authors present RePair as a structured way to screen, compare, and refine micro-to-macro explanations.Source: §10 · p. 21

A THE MODELS CONSIDERED

• Admission required a valid baseline with room for the outcome to move in both directions; in segregation, candidates also had to meet a pooled-departure calibration threshold.The segregation screen first checked execution, then required four additional runs whose pooled departures averaged about seven per baseline.Source: A THE MODELS CONSIDERED · p. 26
• The world-specific screen admitted four models each for segregation, public goods, and cascade, and three for polarization.Admitted models differed across worlds; twelve were screened for segregation and seven for each other world.Source: A THE MODELS CONSIDERED · p. 26
• Every result reported in the paper rests on models admitted by the screen for that world.Source: A THE MODELS CONSIDERED · p. 26