Bayesian Persuasion Simplified

• Bayesian Persuasion is a strategic information design model where an informed sender commits to a signaling scheme, influencing receiver actions through Bayes' rule and adherence to Bayes plausibility.
• The paradigm spans private and public signaling, multi-agent environments, and dynamic as well as learning-based variants, with optimal strategies often derived via concavification techniques.
• Methodologies incorporate computational approaches, differential privacy, and approximation algorithms to address complexity and practical challenges in real-world applications.

Bayesian persuasion is a model of communication between an informed sender and a receiver with a payoff-relevant action in which the sender commits ex ante to a signaling scheme, the receiver updates beliefs by Bayes’ rule after observing a signal, and the sender chooses the information structure to influence the receiver’s action (Lin et al., 6 Jun 2025, Tang et al., 22 May 2025). In posterior space, the sender’s problem is governed by Bayes plausibility, and the sender’s value is the concave closure of the induced posterior payoff at the prior; in private multi-agent environments, the corresponding computational object is the concave closure of the sender’s set function (Cheng et al., 26 Sep 2025, Babichenko et al., 2016). The literature consequently treats Bayesian persuasion as a broad information-design paradigm spanning private and public signaling, multi-receiver and multi-sender settings, mediation, approximate obedience, differential privacy, dynamic commitment, learning-based prediction, and natural-language implementations (Castiglioni et al., 2020, Gradwohl et al., 2020, Arieli et al., 2022, Pan et al., 2024, Li et al., 3 Feb 2025).

1. Canonical formulation and posterior geometry

In a canonical formulation, there is a state $s \in \mathbb{S}$ drawn from prior $\mu_0 \in \Delta(\mathbb{S})$, a sender who observes $s$, a signal space $\mathbbold{\Sigma}$, and an action space $\mathbb{A}$. The sender commits to a signaling scheme

$\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$

the receiver observes $\sigma\in\mathbbold{\Sigma}$, updates beliefs, and chooses an action according to an action rule

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$

Sender and receiver payoffs are $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$, and the sender solves

$$\max_{\varphi}\ \mathbb{E}_{\mu_0,\varphi,\pi}\!\left[r^i(s,a)\right]$$

subject to obedience constraints

$\mu_0 \in \Delta(\mathbb{S})$0

Under the revelation principle, signals can be identified with action recommendations by taking $\mu_0 \in \Delta(\mathbb{S})$1, so the sender chooses a Bayes-correlated equilibrium / obedient signaling scheme (Lin et al., 6 Jun 2025).

A posterior-space representation writes the sender’s expected stage payoff under posterior $\mu_0 \in \Delta(\mathbb{S})$2 as

$\mu_0 \in \Delta(\mathbb{S})$3

and the Bayesian persuasion objective as

$\mu_0 \in \Delta(\mathbb{S})$4

subject to Bayes plausibility,

$\mu_0 \in \Delta(\mathbb{S})$5

This is the standard concavification form: the sender optimizes over Bayes-plausible distributions over posteriors rather than directly over signals (Ko, 3 Aug 2025).

Two geometric vocabularies recur. In one-receiver models, the sender’s value is the concave closure of $\mu_0 \in \Delta(\mathbb{S})$6 at the prior, which explains why partial disclosure can outperform full disclosure (Cheng et al., 26 Sep 2025). In private multi-agent settings, the relevant object is the concave closure of a set function. Writing $\mu_0 \in \Delta(\mathbb{S})$7 for the probability of recommending adoption exactly to subset $\mu_0 \in \Delta(\mathbb{S})$8, the sender solves

$\mu_0 \in \Delta(\mathbb{S})$9

subject to

$s$0

This makes private Bayesian persuasion computationally equivalent to evaluating the concave closure $s$1 (Babichenko et al., 2016).

2. Signaling architectures and strategic variants

Private signaling allows the sender to communicate separately with each agent. In the multi-agent private model with two states $s$2, two actions $s$3, and monotone sender utility $s$4, the sender commits to

$s$5

A key structural reduction permits restriction to straightforward binary policies: each $s$6, each agent follows the recommendation, and in the good state the sender recommends adoption to everyone,

$s$7

In the bad state, agent $s$8 can be “lied to” only with probability at most

$s$9

so the persuasion problem reduces to choosing a distribution $\mathbbold{\Sigma}$0 over recommendation profiles subject to marginal constraints $\mathbbold{\Sigma}$1 (Babichenko et al., 2016).

Public persuasion imposes the opposite architecture: the sender must send the same signal to all receivers. In the general public model with no inter-agent externalities, arbitrary state spaces, arbitrary action spaces, and arbitrary sender utility, direct public schemes recommend an action profile $\mathbbold{\Sigma}$2, and persuasiveness requires

$\mathbbold{\Sigma}$3

for every receiver $\mathbbold{\Sigma}$4, recommended action $\mathbbold{\Sigma}$5, and deviation $\mathbbold{\Sigma}$6. The literature therefore distinguishes $\mathbbold{\Sigma}$7-persuasive bi-criteria approximations from exact persuasion (Castiglioni et al., 2020).

A mediated architecture inserts uninformed intermediaries between sender and receiver: $\mathbbold{\Sigma}$8 For one mediator, the sender’s value is a constrained concavification $\mathbbold{\Sigma}$9, where feasible posterior tuples must be affine dominating with respect to the mediator’s indirect utility. In this case the sender still needs at most $\mathbb{A}$0 signals, and the presence of the mediator is never profitable for the sender. With two mediators, however, the second mediator may improve the sender’s value, and optimal persuasion may require more signals; the standard $\mathbb{A}$1-signal sufficiency can fail (Arieli et al., 2022).

Competition among multiple senders yields a different departure. In the multi-sender game where each sender commits to a distribution over IC signals and the receiver chooses exactly one sender at the interim stage, every Nash equilibrium is fully informative whenever senders are even slightly uncertain about one another’s preferences. The mechanism is that any sender can simulate a rival’s signal and improve it in states where sender and receiver are aligned, so non-full-information outcomes are destabilized by competition for being chosen (Gradwohl et al., 2020).

3. Computation, approximation, and hardness

The algorithmic literature centers on which structural classes of sender utility admit exact or approximate optimization. In private multi-agent persuasion, if $\mathbb{A}$2 is monotone submodular, there is a polynomial-time signaling policy with revenue at least

$\mathbb{A}$3

for any arbitrarily small $\mathbb{A}$4. The construction discretizes marginal probabilities, rewrites the rounded problem as monotone submodular maximization under a partition matroid, and invokes the $\mathbb{A}$5-approximation for matroid-constrained submodular maximization. This factor is essentially tight: for monotone submodular $\mathbb{A}$6, approximating the concave closure within a factor better than $\mathbb{A}$7 is NP-hard. By contrast, when sender utility is anonymous,

$\mathbb{A}$8

the exact optimal revenue and an optimal signaling policy can be computed in polynomial time (Babichenko et al., 2016).

Approximate obedience substantially alters this picture. Under $\mathbb{A}$9-best response, the receiver may choose any action in

$\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$0

and the sender evaluates a signaling scheme under worst-case $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$1-best-response behavior. The revelation principle then fails: signals must encode the entire $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$2-best-response set induced by the posterior, not merely a recommended action. Nonetheless, exact polynomial-time algorithms exist when either the action space or the state space is small, there is a quasi-polynomial-time approximation scheme for the general problem, and there is no polynomial-time exact algorithm in general unless $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$3 (Yang et al., 2024).

Combinatorial receiver actions create another algorithmic frontier. When feasible actions are constrained by uniform, partition, or graphic matroids, or by paths in a graph, constant-factor approximation is NP-hard in several natural cases. For general matroids, however, if the number of states of nature is constant and receiver utility is linear, the optimal signaling scheme can be computed in polynomial time by enumerating all possible receiver best responses through a hyperplane arrangement in posterior space. A relaxed solution concept, CCE-persuasiveness, replaces full obedience by a coarse-correlated-equilibrium analogue and supports polynomial-time approximability under an appropriate oracle condition (Fujii et al., 2021).

Public persuasion is hard up to quasi-polynomial time. In the general public setting, computing an $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$4-persuasive scheme requires at least quasi-polynomial time under the Exponential Time Hypothesis even with binary action spaces, while a matching quasi-polynomial-time bi-criteria approximation algorithm exists and becomes a QPTAS in specific settings (Castiglioni et al., 2020).

4. Commitment, repeated interaction, and dynamic delegation

Commitment is central in the canonical model, but later work shows that its role is contingent on timing, observability, and equilibrium discipline. In long-run persuasion, a long-lived sender commits once to a stationary information structure $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$5, while a sequence of short-lived receivers may become suspicious and compare $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$6 to an uninformative alternative $\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$7 using the Bayes factor

$\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$8

If the sender’s preferred action is non-revealing and strictly better for the sender in every state, receivers never switch away from the BP-optimal scheme; if the sender’s preferred action is revealing, full disclosure may dominate the one-shot BP-optimal structure for sufficiently patient senders and sufficiently low switching thresholds (Ko, 3 Aug 2025).

A distinct line removes commitment altogether. In persuasion without commitment, the sender privately acquires information by choosing and running experiments, cannot commit to a disclosure policy ex ante, and then verifiably discloses only a subset of realized outcomes. Under general conditions, the sender can still attain the full Bayesian persuasion payoff in equilibrium. The key condition is the existence of a punishing action with respect to the action set used in the Bayesian persuasion solution; with transparent motives, such a punishing action always exists. The implementation logic is selective disclosure plus privately chosen experiments, and it can fail if no punishing action exists or if there is too large a probability that the sender cannot run any experiment (Arieli et al., 24 Nov 2025).

Dynamic delegation embeds persuasion inside intertemporal mechanism design. In Bayesian promised delegation, the principal privately observes a shock affecting state transitions and uses looking-forward persuasion and informational burning to support incentive compatibility. A randomized mechanism over stage menu functions and promised continuation values satisfies dynamic obedience and promise-keeping constraints, and the resulting BPD mechanism can achieve the same optimal social welfare as the original mechanism in stationary Markov perfect Bayesian equilibria. A revelation-principle-like regime represents persuasion with belief hierarchies as direct recommendation of future promises (Zhang et al., 2022).

Bilateral commitment reinterprets the standard environment as information bargaining. In that view, the receiver can evaluate the full game structure and condition acceptance on induced payoff pairs

$\varphi:\mathbb{S}\to \Delta(\mathbbold{\Sigma}),$9

so persuasion becomes a bargaining stage followed by realization. A joint commitment is a fixed point

$\sigma\in\mathbbold{\Sigma}$0

of an update map $\sigma\in\mathbbold{\Sigma}$1, and the reformulation isolates informational advantage from first-proposer advantage while highlighting fairness and Pareto efficiency via bargaining objectives such as

$\sigma\in\mathbbold{\Sigma}$2

(Lin et al., 6 Jun 2025).

5. Learning-based, robust, and privacy-constrained persuasion

When the common-prior assumption is implausible or statistically prohibitive, Bayesian persuasion can be reformulated as a learning problem. In persuasive prediction, the sender observes covariates $\sigma\in\mathbbold{\Sigma}$3, predicts an outcome $\sigma\in\mathbbold{\Sigma}$4, and releases a randomized predictor $\sigma\in\mathbbold{\Sigma}$5. The key obedience proxy is decision calibration: $\sigma\in\mathbbold{\Sigma}$6 with calibration error

$\sigma\in\mathbbold{\Sigma}$7

If $\sigma\in\mathbbold{\Sigma}$8 is $\sigma\in\mathbbold{\Sigma}$9-decision calibrated, it satisfies $\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$0-swap regret. The PerDecCal algorithm is oracle-efficient, and in the single-receiver case its utility matches

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$1

with sufficiently many samples (Tang et al., 22 May 2025).

Robust persuasion under uncertain receiver behavior replaces explicit belief modeling by end-to-end prediction of receiver actions. In the conformal framework, the receiver observes a private signal $\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$2, the sender chooses $\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$3, and a neural predictor

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$4

learns $\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$5. Conformal prediction then defines action sets

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$6

yielding the robust objective

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$7

The framework provides exact marginal coverage for the data-generating policy, explicit degradation bounds under policy shift, neural-network approximation and estimation error bounds, and sample complexity

$\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$8

(Bang et al., 9 Nov 2025).

Privacy constraints alter the feasible posterior region itself. In differentially private persuasion, the state is a binary database $\pi:\mathbbold{\Sigma}\to \Delta(\mathbb{A}).$9, and the signaling scheme must satisfy $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$0-DP, $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$1-DP, or Rényi DP. Under pure DP, feasible posteriors lie in a fixed region $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$2, so optimal persuasion is the concave hull of sender value on that region. Under approximate DP, support-wide slack variables enlarge the problem into a constrained concavification in an augmented space. For a common utility-function family and a wide range of privacy levels, there are significant constant utility gaps between $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$3-DP, $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$4-DP, and no privacy; under homogeneous conditions, optimal differentially private signaling schemes are computable in polynomial time (Pan et al., 2024).

A further extension augments information design with direct management of receiver tastes. In Bayesian persuasion under bias management, the principal chooses both an information policy $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$5 and a bias-management policy $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$6. The inner problem yields bang-bang management,

$r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$7

and the outer problem remains a concavification,

$r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$8

but now of an endogenous posterior value function that already incorporates optimal management and posterior-separable information costs. Information and management can therefore be complements, substitutes, or both, depending on the primitives (Ozbek, 11 Feb 2026).

6. Foundations, rationalization, and language-based implementations

Foundational work also studies when observed behavior can be rationalized as persuasion. If an analyst observes only a marginal distribution of actions rather than state-contingent frequencies, BCE-consistency asks whether there exists a Bayes correlated equilibrium whose action marginal matches the data. The characterization combines the obedience approach of information design with the belief approach of Bayesian persuasion using Strassen’s theorem and Hall’s marriage theorem, and it applies to ring-network games and to public-information rationalization in first-order Bayesian persuasion games (Doval et al., 2023).

An axiomatic literature reconstructs persuasion models from menu-choice data. Preferences over menus characterize Bayesian persuasion with known bias under Axioms 1–11, persuasion with uncertain bias under Axioms 1–10, costly Bayesian persuasion under Axioms 1–9, and sequential costly Bayesian persuasion under Axioms 1–8. In the sequential model, the principal first acquires costly information and then can also manage the receiver’s bias after a posterior is realized. The representation yields canonical information and bias-management cost functions recoverable from observable menu-choice data (Higashi et al., 29 Dec 2025).

Natural-language and LLM-based work lifts BP from symbolic state spaces to dialogue. Verbalized Bayesian Persuasion maps BP to a verbalized mediator-augmented extensive-form game and solves it with a generalized equilibrium-finding procedure built on Prompt-PSRO, reinforced by verbalized commitment assumptions, verbalized obedience constraints, and information obfuscation. Its main theoretical guarantee is that VBP returns an $r^i,r^j:\mathbb{S}\times\mathbb{A}\to \mathbb{R}$9-approximate Bayes correlated equilibrium in static BP and an $$\max_{\varphi}\ \mathbb{E}_{\mu_0,\varphi,\pi}\!\left[r^i(s,a)\right]$$0-approximate Bayes-Nash equilibrium in multistage BP (Li et al., 3 Feb 2025).

A complementary empirical program constructs BP environments from human-human persuasion datasets such as Anthropic Persuasion, DDO, Perspectrum, and CMV. In these environments, frontier models achieve substantially higher persuasion gains than smaller models, especially in dynamic settings, and reinforcement learning with PPO or GRPO improves smaller models’ strategic persuasion behavior (Cheng et al., 26 Sep 2025). A related dialogue framework replaces formal pre-commitment by a commitment-communication mechanism in which the sender narrates a type-induced information schema through natural language. Semi-Formal-Natural-Language BP emphasizes credibility and logical coherence, Fully-Natural-Language BP emphasizes emotional resonance and robustness, and both outperform non-BP baselines in reported success rates (He et al., 15 Oct 2025).

Applied work uses BP for platform governance. In a misinformation model, a social media platform predicts a draft post’s misinformation and validation states with noisy classifiers, commits to a signaling scheme $$\max_{\varphi}\ \mathbb{E}_{\mu_0,\varphi,\pi}\!\left[r^i(s,a)\right]$$1, and influences a user’s share/not-share decision. The optimal noisy signaling scheme is characterized by a linear program, classifier-quality monotonicity is characterized by convex-hull inclusion of confusion-matrix rows, and performative dynamics admit convergence and stability results (Hossain et al., 2023).

Bayesian persuasion is therefore best understood not as a single one-shot commitment theorem but as a general theory of strategic information revelation. Its central mathematical invariants are Bayes plausibility, obedience, and concavification or concave closure; its main substantive variables are who commits, what is observed, what counts as obedience, and which additional constraints—competition, mediation, privacy, learning, dynamic credibility, or language—reshape the feasible posterior geometry (Babichenko et al., 2016, Yang et al., 2024, Arieli et al., 24 Nov 2025).