In multi-agent social settings, model reliability varies across relationships. Beyond inferring what others will do, an agent must calibrate how confidently those inferences should guide policy selection for each relationship. An agent may maintain a well-validated model of one partner, a fragile model of another, and a model under revision for a third; collapsing these into a single confidence estimate loses information relevant to policy selection. We therefore formalize affective precision as a relationship-specific metacognitive estimate of confidence in the current partner model. Each partner's behavioral evidence updates a local confidence estimate that modulates policy precision during selection, regulating how strongly current beliefs are expressed in policy rather than changing the content of those beliefs. Simulations in a multi-partner graded trust game show that partner-local affective precision influences behavior primarily through policy commitment rather than direct improvement of partner-state inference. Because the mechanism tracks partner-response predictability rather than realized payoff, greater confidence produces sharper policy commitment without necessarily producing higher rewards. Under abrupt shifts in social behavior, confidence accumulated from previously reliable predictions can remain behaviorally active after the relationship changes, showing that confidence revision can lag behind social change. Finally, varying precision gain and priors produce distinct trust-calibration dynamics, showing how confidence accumulation and revision depend on model parameters. Together, these results show how relationship-specific affective precision can distinguish social prediction from social policy commitment.
Figures & tables
Figure 1: Partner-local versus shared precision in the locality probe (30 seeds per condition). Left and middle: absolute active-encounter partial correlations of precision with surprisal and payoff, controlling for encounter count and the alternative signal. Right: cumulative payoff. Error bars: 95% CIs.
Figure 2: Tracked-only ablation in the open graded partner-choice regime. Bars show means across 30 seeds for within-episode βˉk range, policy entropy, and cumulative payoff (95% CIs). Partner-local precision lowers entropy while tracker movement remains comparable. Paired contrasts are reported in the text.
Figure 3: Abrupt betrayal in the partner-choice regime. Left and right panels compare partner-local precision with no-affect; the middle panel compares partner-local precision with tracked-only. Dashed vertical lines mark partner 0’s round-31 stance switch. Lines show means within 10-round bins across 30 seeds; shading: 95% CIs.
Table 2: Endpoint stance transition dynamics. The focal agent’s graded investment interpolates between Bhigh and Blow according to Eq. ( 6 ).
Condition
β tracker and γk
Precision scope
No-affect
off; γk=γbase fixed
—
Tracked-only
q(βk) updated; γk=γbase fixed
per-partner
Partner-local
q(βk) updated; γk from q(βk)
per-partner
Shared- β
pooled q(β) updated; γk from shared β
pooled tracker; separate beliefs
Appendix
Table 3: Precision ablation conditions
Condition
Purpose
Open partner choice
Tracked-only deployment contrast and descriptive partner-type allocation readout from the same trajectories
Locality probe
Precision–surprisal vs. precision–payoff correlations under local/shared β
Abrupt betrayal
Episode-level policy entropy, payoff, and joint type–stance accuracy, with post-switch partner-selection, precision, and entropy trajectories
Profile variation
α /prior manipulations for trust-calibration profiles
Forgiveness
Reengagement and β recovery after repair
Appendix
Table 4: Simulation condition families and purpose
Experiment family
Setting
Rounds
Seeds
Open deployment / allocation
graded partner choice; shared trajectories for Sections 3.2 – 3.3
200
30
Locality probe
graded partner choice
200
30
Abrupt betrayal
scheduled partner-choice betrayal
120
30
Profile program
sweeps, factorial, forgiveness
200
20
Appendix
Table 5: Representative episode lengths and replication counts by experiment family
Metric
Definition
Total payoff
Cumulative focal payoff over episode/window
Policy entropy
Natural-log Shannon entropy of q(π) , averaged over rounds/seeds
Joint type–stance accuracy
Mean simultaneous correctness of inferred partner type and stance
Selection share
Fraction of rounds each partner is chosen
Selection concentration
Gini coefficient over partner-selection distribution
Mean within-episode βˉk range
Average within-episode max–min range of posterior mean βˉk , averaged across seeds and episodes
Appendix
Table 6: Metric glossary
α
Mean within-episode βˉk range
0.05
0.097
0.1
0.126
0.3
0.195
0.5
0.219
1.0
0.277
2.0
0.402
Appendix
Table 7: Mean within-episode βˉk range by precision-gain α (betrayal regime)
Figure 4: Precision-gain sweep across open graded and betrayal environments (20 seeds per α /environment). Points show means with 95% CIs, excluding undefined recovery times. The entropy panel pools both environments (40 runs per α ); early, mid, and late denote rounds 1–50, 51–150, and 151–200. All horizontal axes show gain α on a logarithmic scale. Other metric definitions appear in Table 6 .
Profile
Payoff
Recovery (rounds)
Gini
Mean βˉk range
Trust asymmetry
Anxious-reactive
2285.8
25.1
0.626
0.459
14.87
Hypervigilant
2230.7
21.8
0.609
0.556
6.97
Naive-stubborn
2218.0
15.5
0.567
0.173
4.87
Avoidant-rigid
2166.0
13.8
0.567
0.302
1.47
Default
2244.2
31.4
0.643
0.427
7.79
Appendix
Table 8: Behavioral metrics by profile in the betrayal environment; the default agent is shown for reference. Values are means from 20 seeds per profile, excluding undefined recovery times. The βˉk range is the within-episode max–min range, averaged over partners and seeds.
Figure 5: Gain-prior profile signatures in the betrayal environment. Printed cell values are raw means; color scales each metric separately from the lowest to the highest profile mean. Thus, colors compare profiles within a column, not magnitudes between metrics. Table 8 gives the numerical summary.
Variant
Reengagement
Payoff recovery
Latency (rounds)
Mean βˉk range
cautious-high- α
0.529
1.043
14.5
0.607
cautious-low- α
0.684
1.063
6.7
0.301
default
0.584
1.021
10.3
0.485
naive-high- α
0.577
0.989
3.5
0.604
naive-low- α
0.613
1.010
7.9
0.190
no-affect
0.586
1.042
12.4
–
Appendix
Table 9: Forgiveness and trust-repair metrics (20 seeds per profile). Reengagement is the post-repair partner-0 selection rate; payoff recovery uses the windows in Figure 6 . Latency averages only seeds that reengage (16–20 per profile). The dash denotes the absent no-affect β tracker.
Figure 6: Forgiveness and trust repair (means across 20 seeds per profile). A: partner-0 posterior mean βk ; shading marks betrayal (rounds 81–120), followed by repair. Lower βk indicates higher policy precision; no-affect has no tracker. B: partner-0 selection during repair. C: payoff in rounds 151–200 relative to rounds 50–80; the dashed line at 1 marks baseline recovery.
Social interaction can improve collective learning but also amplify early mistakes. We study this tension when the credibility of social information is generated by the sender's own decision process rather than fixed ex ante. Reinforcement-learning agents make binary choices through a drift--diffusion process that jointly determines choice, decision time, and confidence; decision confidence then becomes social credibility by weighting anticipatory influence and retrospective social learning. Under balanced community exposure, the anticipatory field admits an exact quotient representation. Its local Jacobian is a scalar decision-sensitivity term multiplying the community-coupling matrix, which yields a common-mode amplification threshold and an analytical role for cross-community permeability in damping relative community differences. Monte Carlo experiments show the corresponding non-monotone performance pattern: moderate transmission accelerates correction, whereas strong transmission can lock populations into wrong consensus; low permeability instead sustains disagreement. Ablations reveal a dual role for confidence: credibility-sensitive transmission amplifies social error, while confidence-dependent private learning stabilises it. The model yields testable predictions linking sender confidence to receiver behaviour conditional on accuracy.
Gabriel Bontemps, Abhishek Banerjee
GREDEG-CNRS, Université Côte d’Azur, Nice, France · Blizard Institute, Queen Mary University of London, London, United Kingdom of Great Britain and Northern Ireland (the) · Adaptive Decisions Laboratory, Department of Pharmacology, University of Oxford, Oxford, United Kingdom of Great Britain and Northern Ireland (the)
As language-model agents increasingly work in teams, each agent must decide how much to trust its teammates. Yet we lack a standard way to measure trust between AI agents. We propose a behavioral measure based on costly verification. In a cooperative survival game, checking a teammate's work consumes resources, while trusting a wrong answer can be fatal. Relative to a memoryless version of the same model, reduced verification provides an observable measure of trust. Using this framework, we study trust formation, breakage, and recovery across six frontier model snapshots. When paired with a consistently reliable teammate, four snapshots (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, and Gemini 3.1 Pro) reduce verification by roughly 60-85%, whereas two smaller snapshots show little or no such adjustment. Failures reverse this discount, but models differ in how they respond. Some concentrate renewed scrutiny on the culprit, while others become more cautious toward the entire team. Recovery is slower than formation, and clustered failures sustain suspicion far longer than the same number of failures spread apart. These differences have practical consequences. Models that form trust verify less, decide more quickly, and achieve higher payoffs in our environment. By contrast, persistent over-verification is associated with indecision rather than safety. Our results show that trust dispositions can be measured before deployment and suggest that calibration, rather than maximal suspicion, should be the central concern in the governance of multi-agent AI systems.
Yujiao Chen
Massachusetts Institute of Technology Cambridge, MA 02139
Language model agents are increasingly used to simulate social interaction, and the resulting transcripts read as though the agents understand one another. We ask whether that appearance rests on a model of the partner's mind or on the surface record of what the partner said. We build a social simulation in which both questions have exact answers: 40 multi-issue negotiations whose hidden preference weights and whose full Pareto frontier are known by construction. Two model families negotiate across 160 dyads, every transcript is frozen before any measurement, and 2880 counterfactual probes then hold the evidence byte identical while moving one factor at a time: the reader's own stake, the partner's tone, an identity label, and the order of recursion. The agents are socially fluent and economically poor. They reach agreement in 96.2% of dyads with 0 protocol failures, yet only 0.7% of deals land on the Pareto frontier, they leave 20.5% of the available joint value unclaimed, and they miss the one issue on which their interests are perfectly aligned in 76.6% of deals; on the frontier and on that aligned issue, a package drawn at random from the set both sides would accept does as well. The probes locate the failure. Swapping only the reader's own payoff sheet, while the partner's words and offers stay identical, moves the inferred top priority by 15.0 percentage points, which is egocentric projection rather than inference, while a tone rewrite moves it by 5.3 percentage points and an identity label by 0.0. Most tellingly, an agent predicts what its partner believes about it 72.5% of the time while that partner's belief is itself correct only 51.2% of the time: the agents track the conversation far better than they track the mind behind it.