---
kind: concept
slug: opinion-dynamics
title: Opinion dynamics
summary: >-
  Opinion dynamics predicts how a group of agents that hold stances changes them through exchange, and which collective state results. The card covers 32 persona-conditioned LLM agents on a signed graph, fitted by an Ising-style logistic rule with separate concordant, discordant and unsigned couplings. The rule is author-reported at 75-86% one-step balanced accuracy on held-out questions.
tags: [multi-agent, statistical-physics, consensus, llm-agents]
papers: ["2608.16578"]
related: [agent-based-simulation, multi-agent-systems, llm-agent]
status: draft
updated_at: 2026-10-19
---

## Overview

Opinion dynamics asks how a group of agents, each holding a stance, changes
its stances over rounds of contact, and what collective state results: stays
undecided, splits into camps, or agrees. For language-model agents the open
question is whether a compact rule fitted to observed votes can predict the
next vote and the group outcome from initial opinions, personas and the
communication graph alone [arxiv:2608.16578 §1] [arxiv:2608.16578 §2].

The invariant shared by the variants is a logistic update. An agent's next
state is +1 with a probability given by a logit: a persona-question field
plus a weighted sum of its neighbours' current states [arxiv:2608.16578 §4].
The signed tie matrix J sets the sign of each neighbour term. The variants
differ in how that social term is split into coefficients, and in whether
agents update in synchronous rounds or at random times
[arxiv:2608.16578 app. Continuous Extension].

This page rests on one variant card, so the design space shown here is the
one that paper explores. It is not a survey of the field.

## Variants

### Single social coupling

One weight multiplies the whole signed sum, so friendly and unfriendly
neighbours enter with one coefficient, beta [arxiv:2608.16578 §4]. This is the
"discrete update" row of the paper's main comparison. Its one-step balanced
accuracy ranges widely. For example, on subjective questions it is 53.9 for
Llama-3.1-8B in-distribution, and the authors describe the rule as dropping
to near chance in some cells [arxiv:2608.16578 tab. 1] [arxiv:2608.16578 §5.1].
Card: [Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents](/papers/2608.16578).

### Three couplings

The social term is split into a concordant part (positive ties), a discordant
part (negative ties) and an unsigned part (any tie), each with its own weight
beta+, beta- and beta0 [arxiv:2608.16578 §4]. The authors report this rule as
the best in every column of their main table, with one-step balanced accuracy
of 75.4 to 86.3 on held-out questions across four models
[arxiv:2608.16578 tab. 1] [arxiv:2608.16578 §5.1]. The unsigned drive is a
linear combination of the two signed drives, so the three weights are not
jointly identified. The paper fits beta0 first and then holds it fixed
[arxiv:2608.16578 app. Fitting]. Card: [Physics of Agents](/papers/2608.16578).

### Continuous-time mean-field update

Agents revise at independent random times. The state is a mean m_i in
[-1, +1] that moves toward a tanh of the social field at a learned rate eps
per round [arxiv:2608.16578 app. Continuous Extension] [arxiv:2608.16578
app. Fitting]. The paper evaluates this only on asynchronous data from
GPT-4o-mini, with raw accuracy rather than balanced accuracy, where it beats
the discrete three-coupling rule [arxiv:2608.16578 §5.5]. Card: [Physics of
Agents](/papers/2608.16578).

### Five couplings (objective questions)

The concordant and discordant channels are split again by whether the
neighbour's current state matches the ground truth. Four weights, beta+_T,
beta+_F, beta-_T and beta-_F, plus beta0 describe this rule
[arxiv:2608.16578 §5.4]. The paper defines it for objective questions only
[arxiv:2608.16578 tab. 3]. It is a diagnostic for whether correct neighbours
pull harder, not a general predictor. Card: [Physics of Agents](/papers/2608.16578).

### Baselines

Three reference rules sit beside the variants. Persistence copies the current
state. Interaction-free uses only the persona-question field. Mean-field adds
the population mean opinion as one feature and has no graph
[arxiv:2608.16578 app. Baselines]. On held-out data the interaction-free rule
scores between 45.6 and 60.8, so the persona-question field alone explains
little of the transitions [arxiv:2608.16578 tab. 1].

## When to use which

The choice depends on what data exists and what must be predicted.

| Situation | Use | Why |
|---|---|---|
| Synchronous rounds with vote logs, a known graph, and a need for the next vote | Three couplings, discrete | Best one-step accuracy in the paper's tables for four models [arxiv:2608.16578 tab. 1] |
| The graph changes or is unknown at fit time | Refit with care | Held-out graph families kept high one-step accuracy for GPT-4o-mini, for example 95.1 on subjective low-rank graphs [arxiv:2608.16578 tab. 2]. Only one model was tested this way. |
| Multi-step forecasts of group outcome | Three couplings, with rollouts discounted | In the three-coupling row, rollout accuracy is lower than one-step accuracy in every column, for example 77.2 against 86.3 for GPT-4o-mini out-of-distribution, subjective [arxiv:2608.16578 tab. 1] |
| Agents update at random times | Continuous-time variant | It beats the discrete rule on asynchronous data, though only for GPT-4o-mini and with raw accuracy [arxiv:2608.16578 §5.5] |
| Objective questions where accuracy toward truth matters | Five couplings, as a diagnostic | Shows beta+_T above beta+_F in all four models [arxiv:2608.16578 §5.4] |
| A new model with no vote data | Not supported by this card | The fitted couplings differ widely by model, for example beta+ on objective questions ranges from 0.29 (Gemma-3n-E4B) to 2.10 (GPT-4o-mini) [arxiv:2608.16578 tab. 3]. Fit per model. |
| Larger mixed-model groups | Three couplings, with caution | One graph and one sample per agent; 81.9 subjective and 80.0 objective one-step accuracy for 64 agents [arxiv:2608.16578 app. Experiments with Frontier Models and Larger Mixed Groups] |

Because the numbers are author-reported, compare only like with like. Balanced
accuracy, raw accuracy and rollouts are different metrics, and the paper uses
K = 5 votes per round in the main runs but K = 1 in the temperature sweep and
the frontier run [arxiv:2608.16578 §5.3] [arxiv:2608.16578 app. Experiments
with Frontier Models and Larger Mixed Groups].

## Implementation sketch

The common core is one logit per agent, then a logistic link. Variants differ
only in the drive terms. The sketch below uses the three-coupling form. Each
drive is a matrix product on the signed tie matrix J.

```pseudocode
# One-step probability that each agent's next state is +1  [arxiv:2608.16578 §4]
# Input:  states (N,) in {-1, +1}; J (N, N) with entries in {-1, 0, +1}
#         features (N, 16) from persona and question embeddings
# Output: P(next state = +1), shape (N,)

def next_state_probability(states, J, features, weights):
    # Intrinsic field from persona and question  [arxiv:2608.16578 app. Fitting]
    field = features @ weights.field
    # Split the social term by the sign of each tie  [arxiv:2608.16578 §4]
    drive_pos = positive_part(J) @ states
    drive_neg = negative_part(J) @ states
    # Unsigned drive: a mix of the signed drives  [arxiv:2608.16578 app. Fitting]
    drive_abs = abs(J) @ states
    # Fitted coefficients enter the logit as-is, no factor 2  [arxiv:2608.16578 §4]
    logit = (field
             + weights.pos * drive_pos
             + weights.neg * drive_neg
             + weights.zero * drive_abs)
    return sigmoid(logit)
```

Fit the weights by maximum likelihood on one-step transitions. Fix beta0
first, then fit the signed couplings [arxiv:2608.16578 app. Fitting]. Scoring
uses the mean of four group accuracies, which makes a copy-the-state rule
score 50 by design [arxiv:2608.16578 app. Balanced Accuracy].

## Contradictions

None between papers, since this concept has one card. Within the paper, the
saved notebook output differs from its Table 1 in some cells, and the
message scheme differs for two of the four models. This page cites the
paper's values [arxiv:2608.16578 tab. 1] [code: res/4_prediction.ipynb, cell 14dfde1e].
The card's Paper vs code section lists the details.

## Papers

- [Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents](/papers/2608.16578): an Ising-style logistic rule with concordant, discordant and unsigned couplings, fitted to votes of 32 LLM agents on a signed graph.