Research map / Multi-agent systems and games / LLM agents

Steering the Herd: A Framework for LLM-based Control of Social Learning

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Models controlled sequential social learning where a planner (altruistic or biased) chooses private-signal precision for each agent at a cost, derives optimal Markov policies, and validates them with LLM planner and LLM agent simulations.

Problem

How an algorithmic information mediator (e.g., an LLM) that controls private signal precision can steer sequential social learning and public opinion, and what that does to social welfare, under altruistic or biased objectives.

Approach

Combines dynamic programming over public belief with Bayesian sequential social learning in which the planner chooses each agent's signal precision at a cost; proves convexity of the altruistic value function and characterizes optimal altruistic and biased policies by belief regions, then checks these against LLM-run planners and agents.

Key results

• Biased planner with misaligned goals decreases social welfare by roughly 40-50% in both analytical and LLM settings, via intentionally reducing signal precision.
• LLM planner policies deviate less than 10% from the analytically optimal policy for the majority of belief states in both altruistic and biased settings.
• LLM agents show non-Bayesian belief updating: underreact to belief-consistent signals, overreact to counter-prior signals, and require stronger public belief to enter cascades.

Surprising findings

• For some belief ranges the biased planner intentionally obfuscates agents' signals to steer outcomes.
• Even under transparency constraints (information parity, no lying, full observability), mediation shifts welfare substantially in either direction.

Limitations

• Binary state and binary symmetric signal channel with 0-1 losses; planner has same history as agents.
• Limited human data for validating LLM simulators.
• Assumes Bayesian agents in analytical results; LLM simulation is a single car-purchase scenario.

Experimental setup

Benchmarks: LLM-based car purchase simulation numerical MDP solutions

Metrics: planner utility social welfare change policy deviation from optimal

Baselines: myopic planner analytically optimal planner no-control baseline

Builds on: Bayesian persuasion sequential social learning (Banerjee, Bikhchandani-Hirshleifer-Welch)

Closest papers by meaning

• The Sample Complexity of Online Strategic Decision Making with Information Asymmetry and Knowledge Transportability (ICML 2025 · similarity 0.86)
• Collaborative Mean Estimation Among Heterogeneous Strategic Agents: Individual Rationality, Fairness, and Truthful Contribution (ICML 2025 · similarity 0.86)
• ALSO: Adversarial Online Strategy Optimization for Social Agents (ICML 2026 · similarity 0.86)
• Look-Ahead Reasoning on Learning Platforms (NeurIPS 2025 · similarity 0.86)
• Network Intervention by Polling Strategic Agents (NeurIPS 2026 · similarity 0.85)
• From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.85)