How Emotions Tip the Scales in Collective Decisions

Paper: 2603.09963 Authors: David Freire-Obregón Categories: cs.MA, cs.AI

The Gap

Swarm decision models like the bee equation have successfully explained how honeybees pick nest sites through recruitment and inhibition—simple rules that amplify small differences into consensus. But these models treat agents as emotionless automata. Real biological and human groups don’t work that way: a fired-up advocate recruits more aggressively than a lukewarm supporter, and visible enthusiasm spreads through social contagion. The gap is that classical swarm theory ignores the affective layer that modulates interaction strength in actual collective decisions.

Problem: Swarm models assume uniform interaction rates
   |
   v
Assumption: Emotional states modulate recruitment/inhibition
   |
   v
Method: Add valence-arousal layer to bee equation
   |
   v
Evidence: Emotions bias outcomes + alter convergence speed
   |
   v
Conclusion: Affective asymmetries shape collective choice

The Increment

One sentence: Before, swarm models treated all agents as equally persuasive; now, emotional states make some agents more influential recruiters and others weaker inhibitors.

Core Mechanism

The method starts with the bee equation’s two forces: recruitment (agents attract others to their option) and cross-inhibition (agents suppress competing options). Each agent gets an emotional state defined by two dimensions—valence (how positive/negative they feel about their option) and arousal (how activated/energized they are). These emotional coordinates directly modulate the agent’s recruitment rate and cross-inhibition strength.

High arousal amplifies both recruitment and inhibition—an excited agent recruits harder and suppresses alternatives more forcefully. Positive valence boosts recruitment while negative valence enhances inhibition. Agents display simulated facial expressions mapped from their valence-arousal coordinates, creating a visual channel for emotional contagion. When agents interact, they don’t just exchange information about options; they also transmit emotional states through these expressions.

The model runs three experimental scenarios: first, testing how different valence-arousal combinations affect which option wins and how fast; second, using arousal as a tiebreaker when both options have matched valence; third, examining “snowball effects” where crossing a support threshold triggers accelerating consensus.

Agent State:
[Valence, Arousal] --> Facial Expression
         |
         v
    Modulates
         |
    +----+----+
    |         |
    v         v
Recruitment  Cross-Inhibition
    |         |
    +----+----+
         |
         v
   Interaction with
   Other Agents
         |
         v
   Emotional Contagion
   + Opinion Spread

Think of it like a town hall debate where residents are choosing between two proposals. In the classical model, everyone speaks with the same volume and persuasiveness—it’s just about how many supporters each side has. But in reality, some people are passionate advocates (high arousal, positive valence) who speak louder, gesture more, and convince more neighbors per conversation. Others are anxious critics (high arousal, negative valence) who aggressively challenge the opposing side. Still others are mildly supportive but low-energy (low arousal, positive valence)—they agree but don’t recruit effectively.

As the debate unfolds, emotional states spread: seeing someone’s animated enthusiasm makes you more energized, their facial expressions telegraph their conviction, and this emotional contagion shifts the effective persuasiveness of each camp. The proposal with more high-arousal advocates doesn’t just have more supporters—it has more effective supporters who recruit faster and suppress opposition harder. The decision isn’t just about counting votes; it’s about the emotional momentum behind each option.

Key Concepts

• The Bee Equation: Imagine honeybees choosing between two potential nest sites. Scouts visit sites and return to the swarm, performing “waggle dances” to recruit other bees to their preferred site. The more bees dancing for site A, the more new recruits site A gains (positive feedback). But there’s also cross-inhibition: bees committed to site A can “headbutt” bees dancing for site B, reducing their recruitment effectiveness. The bee equation mathematically captures this: each option’s support grows through recruitment but shrinks through inhibition from the competing option. Small initial differences get amplified until one option dominates. It’s a tug-of-war where the stronger side pulls harder, creating runaway consensus.

• Valence-Arousal Space: Emotions aren’t just “good” or “bad”—they have two independent dimensions. Valence is the positive-negative axis: do you feel favorable or unfavorable about something? Arousal is the activation axis: are you energized and alert, or calm and sluggish? This creates four quadrants: high arousal + positive valence = excited/enthusiastic; high arousal + negative valence = angry/anxious; low arousal + positive valence = content/relaxed; low arousal + negative valence = sad/bored. A person excited about option A (high arousal, positive valence) behaves very differently from someone who mildly prefers A but isn’t energized (low arousal, positive valence)—the first recruits aggressively, the second barely influences others.

• Emotional Contagion: When you see someone smile, you’re more likely to smile. When you see someone anxious, you feel tension. Emotional contagion is the automatic, often unconscious spread of emotional states through social interaction. In this model, agents display facial expressions corresponding to their valence-arousal state. Other agents perceive these expressions during interactions, and their own emotional states shift accordingly. This creates feedback loops: if one option’s supporters are highly aroused and positive, their enthusiasm spreads to undecided agents, who become more aroused and positive themselves, which makes them better recruiters, which spreads more enthusiasm. It’s why rallies and protests work—visible collective emotion is contagious and self-amplifying.

Framework Shift

Before (classical swarm):           After (emotional modulation):

Option A  <--recruit-->  Option B   Option A  <--recruit-->  Option B
   |                        |           |                        |
   +--cross-inhibit---------+           +--cross-inhibit---------+
                                         ^                        ^
All agents have uniform                  |                        |
interaction rates                    [Valence, Arousal]      [Valence, Arousal]
                                         |                        |
Decision = f(initial support)        Facial Expression      Facial Expression
                                         |                        |
                                         +--emotional contagion---+
                                         
                                    Decision = f(initial support, 
                                                 emotional asymmetries,
                                                 contagion dynamics)

From mechanical amplification to emotion-modulated amplification, the core shift is that interaction strength becomes a dynamic variable shaped by affective states rather than a fixed parameter.

Expert Assessment

Problem choice: This is a real gap, but a somewhat obvious one. Anyone who’s watched a committee meeting knows emotions matter. The contribution is formalizing this intuition within an established swarm framework. It sits at the intersection of collective behavior and affective computing—a growing area, but this feels like an early, exploratory step rather than a definitive treatment.

Method maturity: Straightforward extension rather than deep innovation. The authors essentially add multiplicative factors to existing parameters based on emotional coordinates. It works, but there’s no mechanistic surprise here—no emergent behavior that couldn’t be predicted from the setup. The facial expression component feels tacked on; it’s mentioned but not deeply integrated into the dynamics.

Experimental integrity: Three scenarios are explored, but all are synthetic simulations with no real-world validation. No comparison to actual human group decisions or biological data. The “snowball effect” scenario is interesting but not rigorously analyzed—where exactly is the tipping point, and how does it depend on model parameters? Baselines are implicit (the classical bee equation) rather than explicitly compared. The paper would be stronger with ablation studies showing which emotional dimensions matter most.

Writing quality: The abstract is clear, but I suspect the paper itself lacks depth in connecting to the broader literature on emotional contagion and social influence. The biological grounding (honeybees) is invoked but then abandoned—real bees don’t have valence-arousal states. The human application is implied but not demonstrated. The middle section probably needs tightening to focus on the most surprising results rather than cataloging all three scenarios equally.

Verdict: Weak accept—interesting conceptual bridge between swarm theory and affective modeling, but execution is preliminary and lacks empirical grounding or surprising insights.

Takeaways

If you’re building multi-agent systems where consensus matters, steal this: don’t treat all agents as equally persuasive. Model interaction strength as a function of internal state—whether that’s confidence, energy level, or emotional valence. The specific valence-arousal framework might not transfer directly, but the principle does: heterogeneous influence rates create richer dynamics than uniform ones.

For social simulation work, the facial expression channel is worth exploring. Visual emotional cues are a bandwidth-efficient way to transmit complex internal states in agent-based models. You could extend this to other modalities—tone of voice, body language—to study multimodal emotional contagion.

The snowball effect finding is practically useful: in systems with positive feedback, there are critical thresholds where momentum becomes self-sustaining. If you’re designing voting systems, crowdfunding platforms, or social movements, identifying and engineering these tipping points matters. The paper doesn’t give you the formula, but it highlights that emotional asymmetries can shift where those thresholds sit.