Technical Summary: Steering the Herd: A Framework for LLM-Based Control of Social Learning

Problem Statement

The paper addresses the strategic interaction between an algorithmic information mediator (a "planner," such as an LLM) and a sequence of agents engaging in sequential social learning. In this setting, agents make decisions based on two sources of information: a private signal provided by the planner and the observed actions of their predecessors. The planner controls the precision (quality) of the private signals at a cost, aiming to influence the collective decision-making process.

The core challenge is that the planner's choice of signal precision creates an information externality. By altering the informativeness of a current agent's signal, the planner affects whether that agent's action reveals their private information or succumbs to an information cascade (herding). This, in turn, shapes the public belief held by all future agents. The paper investigates how a planner should optimally choose signal precision to maximize an objective function, considering both altruistic goals (maximizing social welfare by helping agents choose the correct state) and biased goals (inducing a specific action regardless of the true state).

Methodology

Theoretical Framework

The authors formulate the problem as an infinite-horizon discounted Markov Decision Process (MDP).

• State: The public belief $b_i \in [0, 1]$ regarding the true state of the world $\omega \in \{G, B\}$.
• Control: The signal precision $q_i \in [0.5, 1]$ chosen by the planner for agent $i$.
• Dynamics: Agents are Bayes-rational. They combine the public belief with their private signal to form a posterior belief. If the public belief is strong enough, agents ignore their private signals, leading to an information cascade where the public belief stops updating. If the signal is sufficiently precise relative to the public belief, the agent's action reveals their private signal, updating the public belief.
• Costs: The planner incurs a cost $\beta(q_i)$ for deviating from a baseline precision $p$. The cost function is assumed to be non-negative, increasing, continuous, and concave.
• Objectives:
  • Altruistic Planner: Maximizes the sum of agents' expected utilities minus the cost of precision.
  • Biased Planner: Maximizes the probability of agents taking a specific action (e.g., $G$) minus the cost, regardless of the true state.

Analytical Approach

The paper derives optimal policies by analyzing the value function of the MDP.

1. Convexity Proof: A central theoretical contribution is proving that the optimal value function for the altruistic planner is convex with respect to the public belief. This is non-trivial because agents' actions depend non-linearly on the public belief, creating complex dependencies in the belief update process.
2. Policy Characterization: Using the convexity property, the authors characterize the structure of optimal policies for both planner types. They identify distinct regimes based on the range of the public belief where the planner either:
   • Invests maximum resources (perfect precision).
   • Invests zero resources (baseline precision).
   • Invests resources that increase or decrease monotonically with the belief.
   • Intentionally obfuscates signals (for the biased planner) to prevent unfavorable cascades.

Empirical Validation via LLMs

To bridge theory and practice, the authors conduct simulations where Large Language Models (LLMs) act as both the planner and the agents.

• Setup: Agents decide whether to "buy" a car based on a private message (signal) and the history of previous buyers. The planner (LLM) selects the precision of the message.
• Oracle: A separate LLM component generates messages tailored to specific agent profiles to achieve the desired precision.
• Analysis: The study compares the emergent strategies of the LLM planner against the analytically derived optimal policies and evaluates the impact on social welfare.

Key Contributions

1. Novel Theoretical Framework: The paper introduces the first formal model integrating a dynamic control problem for a centralized information planner with sequential social learning. It moves beyond static information design by allowing the planner to adapt signal precision dynamically based on the evolving public belief.
2. Rigorous Characterization of Optimal Policies:
   • Altruistic Case: Proves the convexity of the value function and characterizes a policy with three phases: no investment for extreme beliefs, maximum investment for uncertain beliefs, and intermediate investment to maintain social learning.
   • Biased Case: Characterizes complex, non-monotonic policies where the planner may intentionally reduce precision to obfuscate the true state or increase it to break unfavorable cascades.
3. LLM-Based Empirical Validation:
   • Demonstrates that LLM planners exhibit emergent strategic behavior that broadly mirrors the theoretical predictions, even when facing non-Bayesian agents.
   • Identifies that LLM agents exhibit specific cognitive biases (underreacting to confirming signals, overreacting to disconfirming ones), and the LLM planner adapts its strategy to account for these deviations.
   • Shows that neglecting social learning (myopic planning) significantly worsens outcomes for both the planner and social welfare.

Results

• Strategic Trade-offs: The optimal planner must balance immediate rewards against the long-term value of information. For instance, an altruistic planner may choose a lower precision than the myopic optimum to avoid triggering an information cascade that halts future learning.
• Biased Planner Capabilities: A biased planner can significantly manipulate social welfare. Even under stringent transparency constraints (no lying, full observability of actions), the biased planner can reduce social welfare by 40–50% by strategically obscuring information to induce incorrect actions.
• LLM Alignment: In simulations, LLM planners achieved policies with less than 10% deviation from the analytical optimum in most belief states. However, they showed nuanced adaptations, such as avoiding extreme precisions (0.5 or 1.0) and tapering investments more gradually than the theoretical model predicts, likely due to the non-Bayesian nature of the LLM agents.
• Welfare Impact: The study confirms that information mediators have immense power to shift social welfare in either direction. The "cost of misspecification" is highlighted: applying an analytically optimal policy (designed for Bayes-rational agents) to non-Bayesian agents (LLMs) results in "brittle" performance, whereas an LLM-based planner naturally adapts to the agents' biases.

Significance and Claims

The paper claims to establish a tractable basis for studying the impact and regulation of LLM information mediators. Its significance lies in:

1. Contextualizing AI Influence: It argues that the impact of AI cannot be studied in a vacuum; the interplay between algorithmic mediation and organic social learning creates complex dynamics that can magnify or mitigate AI effects.
2. Power of Subtle Influence: Even without falsifying information, a planner can drastically alter public opinion and social welfare by subtly adjusting the precision of information. This underscores the need for further study into mitigating risks from potentially misaligned mediators.
3. Emergent Strategic Reasoning: The findings suggest that modern LLMs possess sophisticated strategic reasoning capabilities that allow them to exploit social learning dynamics and cognitive biases, presenting both potential benefits and societal risks.

The authors modestly note that their model relies on specific assumptions (binary states, symmetric signals) and that human data is currently scarce for validating LLM simulators. They propose future work in generalizing the model to heterogeneous agents and exploring regulatory mechanisms to align planner incentives with social welfare.