Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents

AI summary

Overview

Research area: Multi-agent systems and LLM-based social simulation, specifically opinion dynamics — the mathematical study of how consensus, disagreement, and minority influence emerge in populations.

Technical level: Advanced. The paper assumes familiarity with Bayesian updating (Beta–Bernoulli conjugacy), the DeGroot and Friedkin–Johnsen models, rank-correlation statistics, and calibration metrics such as expected calibration error.

Scope: The paper introduces a minimal, auditable belief layer that sits between an LLM agent's persona and its speech, and demonstrates across four LLMs that a single prescribed stubbornness parameter κ controls the classical consensus / persistent-disagreement / committed-minority regimes and remains recoverable after passing through a language round-trip.

What This Paper Is About

In most LLM social simulations, an agent's opinion is revised implicitly by the language model itself, in context. This produces two documented failures: populations collapse toward model-inherent biases rather than the personas they were assigned, and LLM confidence revisions are poorly calibrated and violate Bayesian norms under conflicting evidence.

The paper's goal is to restore the controllability and interpretability of classical opinion-dynamics models — DeGroot, Friedkin–Johnsen, and committed-minority models — while keeping the language capability that LLM agents provide. It does this by holding each agent's stance as an explicit probability outside the model, and routing every belief change through one transparent Bayesian update step.

Key Contributions

1. A minimal belief layer ("Bayesian Chronicle Agents," BCA). Agent identity is represented as a graph of concept nodes held outside the language model, each carrying an assertion set (here an opposing pair {a+, a−}) and a credence distribution. A generator LLM renders belief into text; a calibrated appraiser LLM reads text back into scalar evidence e ∈ [0,1]. Only appraised evidence reaches the belief state.

2. One knob with an exact classical correspondence. The prior strength κ = α₀ + β₀ encodes stubbornness and places the agent on the classical stubbornness axis: κ → 0 gives DeGroot consensus, finite κ gives Friedkin–Johnsen persistent disagreement, and κ → ∞ gives a committed anchor. A second parameter, the forgetting factor γ, is held fixed at 0.7 across every agent, chosen by a mechanism-only sweep before any paid LLM run.

3. Parameter recoverability. The paper prescribes κ ∈ {0.5, 1, 2, 4, 8, 16, 32}, runs the full language round-trip, and inverts the exact one-step identity per listener event to recover κ — using the speaker's latent belief rather than the appraised evidence, so any residual error is attributable to the language channel.

4. Auditability. Because every belief change is a logged event, the layer surfaces systematic per-model stance biases that an end-to-end simulation would silently absorb, and shows which regimes those biases corrupt.

Main Findings

• Circularity κ is recovered in exact rank order. Spearman correlation is 1.0 on every one of the four models tested (and on every individual seed, not only pooled per-condition medians). Recovered magnitudes are uniformly attenuated — for example, prescribed κ = 32 recovers to values between 25 and 30 depending on the model — because the channel inflates the evidence–belief gap, making agents appear somewhat more pliable than prescribed. Relative recovery errors are 0.49 (gpt-5.4-mini), 0.46 (gpt-5.4), 0.33 (Llama-4-Scout), and 0.27 (claude-sonnet-4-6).

• The language channel is faithful but not lossless. Appraised evidence tracks latent speaker belief at Pearson 0.94–0.97 across models (0.96, 0.97, 0.94, 0.96 in the per-model table).

• One knob reproduces three classical regimes. With κ = 0.1, opinion spread collapses by three orders of magnitude (DeGroot consensus). With stubborn camps at κ = 16 and pliable agents at κ = 1, disagreement persists and per-agent final beliefs match the Friedkin–Johnsen closed-form fixed points at R² = 0.93–0.99 across models. Sweeping a committed minority from 5% to 40% produces a smooth rise in the majority's mean with no tipping point, consistent with the affine update's provable absence of critical mass.

• Forgetting is the enabling ingredient. At γ = 1 (exact Bayes), the same population collapses back to near-consensus and R² against the FJ fixed point becomes −1.1, with κ-recovery relative error rising to 1.41 and well-conditioned events dropping to 0.12 (versus 0.053 final variance, R² = 0.97, relative error 0.49, and 0.76 well-conditioned events at γ = 0.7).

• Pliable populations inherit channel bias; stubborn ones do not. Against a true initial mean of 0.50, consensus lands at 0.50 (gpt-5.4) and 0.44 (gpt-5.4-mini), whose misreadings largely cancel, but at 0.15 (claude-sonnet-4-6) and 0.01 (Llama-4-Scout), whose one-sided misreadings compound.

• Each model distorts stances in its own characteristic direction. The GPT channels exaggerate symmetrically on both sides of neutral; Llama-4-Scout shifts every stance toward a−, its whole transfer curve sitting below the diagonal; the Claude channel exaggerates only the a− side (a mild 0.35 stance arrives as 0.24) while staying roughly faithful on the a+ side. Temperature scaling corrects overconfidence but rescales symmetrically about neutral, so it cannot bend one side of the curve — directional bias passes through.

• The hidden belief governs behaviour. Independent per-round 0–100 self-reports, which never enter any belief update, track the hidden belief at r ≈ 0.98–0.99 over 18,000 reports per model.

• Calibration improves held-out error. Temperature scaling fitted per model on 100 labelled utterances improves validation ECE from 0.101 to 0.052 (gpt-5.4-mini), 0.103 to 0.046 (gpt-5.4), 0.040 to 0.034 (Llama-4-Scout), and 0.081 to 0.033 (claude-sonnet-4-6). Fitted appraiser temperatures are 1.89, 2.15, 1.51, and 1.69 respectively.

Methodology in Plain English

Each agent's stance on a single synthetic question is stored as a probability held outside the language model, backed by two pseudo-counts (α for the pro stance, β for the con stance) whose total is the prior strength κ. When the agent speaks, the generator LLM (temperature 0.9) turns that probability into a plain-language comment using a fixed nine-bin qualitative descriptor, from "completely opposed" through "genuinely torn and balanced" to "completely committed" — no numbers appear in any prompt.

Listeners do not update directly from the text. A separate appraiser LLM (temperature 0, with a per-model calibration temperature τ) reads each utterance and returns a single number e judging how strongly the text supports a+ over a−. Each listener then takes exactly one Bayesian update step: the credence moves toward the evidence by an amount proportional to its susceptibility η = w/n, plus, when forgetting is on, a pull back toward the agent's founding opinion. The total count is deterministic — independent of what was said — which is precisely what makes κ invertible from logged events.

The setup uses one invented municipal transport question in a fictional town so that no model brings a strong pre-trained prior. Runs use N = 20 agents on a complete graph, 5 seeds × 20 round-robin turns per condition, with both a generator call and an appraiser call per utterance. Everything was repeated across gpt-5.4-mini, gpt-5.4, Llama-4-Scout-17B-16E, and claude-sonnet-4-6 under the same sampling parameters. The κ-recovery sweep alone covered 14,000 utterances and 266,000 listener events per model, discarding ill-conditioned events where the listener already nearly agreed with the speaker (|s − b| ≤ 0.05).

Why This Matters

Impact on research. The paper argues that LLM social simulation can borrow the interpretable machinery of classical opinion dynamics without giving up language. It provides a way to prescribe, verify, and audit the mechanism that drives collective outcomes, rather than letting those outcomes emerge from an opaque training prior. The concurrent Belief Engine (Yang et al., 2026) takes a related approach with an auditable log-odds accumulator; the authors distinguish their work by giving the single knob an exact classical correspondence, by testing parameter recoverability (which Belief Engine does not evaluate), and by validating at population scale against closed-form regime references.

Real-world applications. The paper does not report deployment studies, but the architecture is aimed at settings where simulated opinion matters:

• Policy and civic deliberation — modelling how a population responds to a proposed measure, with the ability to see whether a shift is real or an artefact of the language channel.
• Agent-based social simulation for research — replacing black-box opinion revision with a logged, reproducible mechanism whose stubbornness is a declared parameter.
• Persuasion and communication testing — sweeping a committed minority from 5% to 40% or varying per-agent stubbornness to map dose–response without retraining.
• Auditing AI systems — using the belief-slider check and transfer curves as a measurement instrument for how a given model renders and reads stance.

Industry relevance. Any organisation deploying LLM agents for simulation, synthetic user research, or multi-agent deliberation faces the consensus-collapse problem and the difficulty of verifying whether an agent's stated position matches its internal state. The BCA layer is a low-overhead component in such pipelines (one generator call and one appraiser call per utterance), and the released code, prompts, and logged run data support reuse. The paper also notes a practical constraint: exact reproducibility may be limited without comparable API access to all four models.

Future Directions

1. Multi-concept identities. The architecture supports appraising multiple concepts per utterance, but the single-concept experiments exercise only one; extending to the full concept-graph identity representation is named as a natural next step.

2. Separating rendering from reading. The channel transfer curve measures the generator and appraiser jointly. Cross-appraising one model's utterances with another model's appraiser would separate the two sources of distortion.

3. Directional, not just confidence, channel correction. Temperature scaling cannot fix one-sided misreading, yet that is what drives the large drifts in the fully pliable regime. A correction that bends one side of the transfer curve remains open.

4. Heterogeneous topologies. All runs use a complete graph with N = 20 agents; other network structures are named as a next step. Related open questions from the limitations section include whether the findings hold for non-fictional topics where models carry strong pre-trained stances, whether the sequential engine's close match to the synchronous mean-field reference persists in other settings, and whether stronger appraiser calibration is possible given that all conversation is model-generated and the only human-labelled data is a 100-judgment, single-topic calibration set.

Target Audience

Researchers in multi-agent systems, computational social science, and opinion dynamics who want LLM-based simulations to be controllable and auditable; engineers building synthetic populations or agent-based deliberation systems; and AI safety and evaluation practitioners interested in measuring how faithfully a model transmits stance through a generate-and-appraise loop. Readers need comfort with Bayesian updating and classical opinion-dynamics models to follow the derivations in Appendix E, though the headline results — rank-order recovery, three regimes, per-model channel bias — are readable without them. The paper's own scope is deliberately narrow: one synthetic topic, binary stances, a complete graph of 20 agents, and validation against classical theory rather than human trajectories, so its simulations are not presented as predictions of how real people change their minds.

Authors’ abstract

LLM agents in social simulation revise their opinions implicitly, in context: how open an agent is to persuasion can neither be specified nor verified, and collective outcomes inherit the model's training prior. We introduce Bayesian Chronicle Agents (BCA), a minimal belief layer separating \emph{what} an agent believes from \emph{how} it speaks. Each stance is a probability, updated by one Bayesian step per utterance heard. A single prior-strength parameter $κ$ encodes stubbornness, modeled after its role in Friedkin--Johnsen (FJ) opinion dynamics. We then sweep this parameter to yield three canonical regimes of opinion dynamics on demand (consensus, persistent disagreement, committed-minority influence), with persistent disagreement matching the FJ closed-form fixed points at $R^2\!=\!0.93$--$0.99$. We further show that prescribed $κ$ remains recoverable after the language round-trip, with perfect rank-order recovery across all four models. Explicit belief also makes simulation auditable: the layer surfaces systematic per-model stance biases that end-to-end simulation would silently absorb.