Should LLM Agents Decide in Social Simulations? — Finite-State vs. LLM Policies

TL;DR

Why it matters. In LLM social simulations the model often does two jobs at once: generating an agent’s content (posts, replies) and choosing its action (read vs. like vs. follow). The second job is rarely audited — yet if the action distribution drifts from the policy the researcher specified, every downstream measurement of diffusion, reach, or engagement drifts with it.
What it proposes. A controlled comparison: the same online-social-network simulation (1,000 agents, 10,000 decisions, 7 actions) run under an explicit finite-state-machine policy versus an LLM action selector, with alignment measured by Jensen–Shannon divergence (JSD).
Headline result. LLM selectors can approximate the reference policy in some configurations but do not preserve it reliably: mean global JSD across nine model×prompt configs is 0.212, no prompting strategy is best across models, and adding “behavioral guidance” can make alignment dramatically worse — GPT-OSS under the guided prompt collapses to selecting follow ~85% of the time (JSD 0.672).
Limit. This benchmarks the decision distribution in isolation; it does not yet trace how a given JSD gap propagates into full-simulation outcomes like cascade size or network evolution.
What to keep. LLM action selection is not a drop-in replacement for an explicit policy. It is, on average, 563× slower than sampling a Markov chain and quietly substitutes a different — and prompt-dependent — decision mechanism. Describing the policy in the prompt does not mean the model follows it.

Background

Agent-based simulation earns its keep by being inspectable: explicit rules, finite-state machines, and Markov transition policies are reproducible, cheap, and their assumptions are written down. In online-social-network (OSN) studies — especially adversarial ones about misinformation reach or campaign amplification — this matters, because small changes in how agents choose actions move the measured outcomes a lot.

LLM-based agents trade that inspectability for flexibility. They interpret context, carry memory, and generate language, and several recent simulators (OASIS, covered earlier in this series; Y Social; MOSAIC) let the LLM also decide what the agent does next. The moment action selection is delegated to a prompt, the behavioral policy becomes a function of the model, the prompt, and the context — none of which is guaranteed to match the researcher’s intent. A growing body of evidence (Chirper.ai analyses, persona-conditioned survey studies) already shows LLM outputs diverging from intended behavior. This paper isolates the question to one variable: holding the agents, network, action space, and context fixed, does an LLM selector reproduce an explicit reference policy, and how does that depend on model and prompt?

Core Idea

OSN Simulation Framework block diagram: an OSN Generator producing friendships and users; an OSN Simulator with Agents Behavior and Content Generation; a Generative Module with Context, Memory and LLM Models; and a Decision-making block offering the FSM baseline or three LLM prompt variants (minimal v1, guided v2, probabilistic v3) — **Fig. 1.**The experimental setup. Everything is held constant — the same synthetic users, friendship graph, action space, and context-construction — except the decision-making block, which is swapped between the FSM baseline and one of three LLM prompting strategies. Action distributions are then compared to the FSM reference via JSD.Source: Buitrago López et al. 2026, Fig. 1

The simulation is a microblogging platform with synthetic profiles, a homophily-generated directed follower graph, and a fixed action space:

$A = \{\text{read},\ \text{like},\ \text{share},\ \text{reply},\ \text{post},\ \text{follow},\ \text{unfollow}\}$

Agents are activated sequentially; at each step the activated agent picks exactly one valid action. The only thing that changes between conditions is the selector. The reference is a finite-state machine, a first-order Markov model whose transition probabilities depend on the agent’s user type. The four conditions are that FSM plus three LLM prompting strategies. Each receives the identical agent profile, current state, valid-action mask, and local context.

The population is deliberately skewed like a real platform: 547 Passive users (54.7%), 221 Socializers (22.1%), 135 Debaters (13.5%), and 97 Advanced users (9.7%). Under the FSM, read dominates every user type — the consumption-heavy shape real OSN traces show.

Method

The three prompts: minimal, guided, probabilistic

The prompts form a ladder of how much of the reference policy is handed to the model:

Base / minimal (v1) — agent metadata, recent state, the valid actions, and short definitions, with just enough framing to avoid two artifacts: framing selection as an optimization problem, and artificially balancing actions. This exposes the model’s default action preferences.
Guided (v2) — adds five behavioral rules distilled from empirical OSN research. Users consume feed content before reacting. Content creation is unequally distributed. Actions differ in effort, from a cheap like to an effortful post. User roles shape engagement. And follow/unfollow are ordinary actions. Calibration rules follow, discouraging the model from over-picking linguistically easy actions like post or reply.
Probabilistic (v3) — hands the model the FSM’s actual numbers: the fixed action order, the current state, the user-type transition probabilities, the contextual mask, and the final normalized probabilities, instructed to use them as strong priors without always taking the argmax.

Measuring drift

Alignment is the Jensen–Shannon divergence between the LLM’s action distribution and the FSM reference, with Laplace smoothing ( $\alpha = 10^{-6}$ ) and base-2 logs so values land in $[0, 1]$ ; lower is closer. Execution time over the 10,000 decisions is reported alongside. Three open-weight models — Llama-3.1-8B, gpt-oss-20b, Mistral-Small-3.2-24B — are served under one local vLLM setup (temperature 0.7, top-p 0.9, top-k 40), so prompt effects can be compared across model families rather than attributed to one implementation.

Experiments

Four stacked-bar panels (FSM, Prompt v1, v2, v3) for GPT-OSS showing action proportions per user type; under the FSM read dominates, but under the guided prompt v2 the follow action takes roughly 85% of decisions across all four user types, collapsing the distribution — **Fig. 2.**GPT-OSS action proportions by user type. The FSM (top) is read-dominated. The guided prompt v2 (third row) is the cautionary case: behavioral 'guidance' pushes the model to select follow ~85% of the time across every user type, the single largest deviation in the study (JSD 0.672). More instruction did not mean more fidelity.Source: Buitrago López et al. 2026, Fig. 4

The central table is alignment per model × prompt (lower JSD = closer to the intended policy):

Model	Base (v1)	Guided (v2)	Probabilistic (v3)	Model mean
LLaMA-3.1-8B	0.359	0.223	0.278	0.287
GPT-OSS-20B	0.113	0.672	0.035	0.273
Mistral-24B	0.045	0.055	0.132	0.077
Best prompt	Mistral	LLaMA	GPT-OSS	—

Three readings. First, no prompting strategy wins across models: the guided prompt is best for LLaMA, the probabilistic prompt is best (and excellent, 0.035) for GPT-OSS, and the minimal prompt is best for Mistral. Whatever closes the gap for one model can open it for another. Second, more guidance is not more fidelity — GPT-OSS’s guided-prompt collapse to follow (0.672, Fig. 2) is the clearest evidence that injected behavioral rules can induce a systematic bias the researcher never intended. Third, Mistral-24B is the most faithful overall (mean 0.077), but it is also the most expensive to run — fidelity and efficiency point at different configurations.

That cost gap is the unambiguous result:

Decision policy	JSD	Time (10k actions)	Slowdown vs FSM
FSM (Markov sampling)	0.000	7 s	1×
Fastest LLM config (LLaMA v1)	0.359	946 s	135×
Best-aligned config (GPT-OSS v3)	0.035	2,990 s	427×
Slowest config (Mistral v2)	0.055	9,360 s	1,337×

Across all nine LLM configurations, action selection averaged 563× slower than the FSM, and even the fastest was 135× slower — all to reproduce a distribution the Markov chain produces exactly, by definition, for free.

Limitations

The study is scoped to the decision distribution in isolation. It establishes that LLM selectors diverge from an explicit policy and quantifies by how much. What it does not yet do is connect a given JSD gap to the simulation-level quantities researchers care about: diffusion depth, engagement, network evolution, narrative amplification. The authors name that as the next step. The reference policy is itself a simplification (a first-order, user-type-conditioned Markov model), so “fidelity” here means fidelity to that abstraction and says nothing about how closely either policy tracks real people. And the findings are specific to three open-weight models under one decoding configuration; larger or proprietary models, or different temperatures, could shift the rankings (the strong prompt-sensitivity of GPT-OSS is a warning that they might).

Sapiens Q Take

Three things to carry into simulation design. First, separate the two jobs an LLM is doing and audit them separately. “Generate the post” and “choose the action” are different responsibilities with different failure modes, and the second is the one that silently rewrites your experiment. If your simulator delegates action selection to a prompt, the action distribution is a free parameter you have not pinned down — measure it against whatever explicit policy you think you’re running, ideally with a divergence metric like JSD rather than eyeballing bar charts. Second, read the GPT-OSS guided-prompt collapse as the general lesson: adding human-written behavioral rules to a selection prompt is an intervention on the policy, never neutral background, and it can concentrate probability on whichever action is easiest to justify under those rules. A prompt that spells out the intended behavior still leaves the actual distribution an empirical question. Third, this is a useful corrective to read next to OASIS in this series: OASIS demonstrates how much social phenomenology emerges when LLM agents act at scale, and this paper is the quiet caveat underneath it — the action policy those emergent results rest on is model- and prompt-dependent, several hundred times more expensive than the explicit alternative, and worth pinning down before attributing an outcome to “the scenario” rather than to the selector. When the research question lives at the level of population dynamics and you need a specified, reproducible action policy, a finite-state machine remains the right primitive to reach for.

References

Original paper: Should LLM Agents Decide in Social Simulations? Comparing Finite-State and LLM-Based Decision Policies
OASIS (LLM-driven OSN simulation at scale): Yang et al. (2024), arXiv:2411.11581 — covered earlier in this series
Underlying OSN simulation framework: Buitrago López et al. (2025), arXiv:2512.22082
Synthetic OSN via homophily: Buitrago López et al. (2026), IEEE TCSS
Y Social (LLM social-media digital twin): Rossetti et al. (2024), arXiv:2408.00818
MOSAIC (content dissemination simulation): Liu et al. (2025), EMNLP 2025
Generative Agents (origin of LLM social simulacra): Park et al. (2023)