LLMs can play roles in simulations—negotiators, coordinators, or rumor spreaders. Treated as boundedly rational, they help prototype policies and probe mechanisms that are hard to observe in the wild.
Where LLM Agents Shine
- Scenario exploration: rapid A/B testing of rule changes or incentives.
- Language-grounded behavior: norms, persuasion, and instructions are textual.
- Human-in-the-loop design: analysts can steer, audit, and correct.
Design Patterns
Hybrid ABM + LLM
Use an ABM for state updates and constraints; delegate decisions or messages to LLM agents. Keep prompts structured (roles, goals, memory) and add critic modules for self-checks.
Transferring “Theory of Mind”
Large models often show stronger strategic reasoning. We can distill guidance into smaller models via examples or reward shaping, retaining efficiency while improving cooperation.
Evaluation & Pitfalls
- Stochasticity: set seeds and replicate runs.
- Overfitting to prompts: rotate instructions; test out-of-distribution cases.
- Goal misalignment: add guardrails and explicit preferences.
Example: In a repeated public-goods game, add an “information modulation” rule (share aggregated feedback, not identities). Many LLM agents cooperate more under this framing.
Note: LLMs are not ground truth. Use them to augment theory and data—not replace either.