Natural Language Mechanisms via Self-Resolution with Foundation Models
Nicolas Della Penna
TL;DR
The paper introduces Language Model Mechanisms (LMMs) that replace terse reports with natural-language submissions and rely on large-language models as world models to determine outcomes and payments. It formalizes a framework with a $\delta$-sufficient world model and an inter-agent information over-determination condition to guarantee a truthful equilibrium and approximate efficiency, with distinct observable-outcomes and zero-shot settings for reporting. A key contribution is linking LLM-based aggregation to incentive compatibility and efficiency, and demonstrating how LMMs can outperform prediction markets in information-rich, high-dimensional settings. The discussion covers practical deployment, limitations of the strong conditions, and directions for empirical validation and generalization to agents with outcome preferences, emphasizing the practical impact on information aggregation in complex environments.
Abstract
Practical mechanisms often limit agent reports to constrained formats like trades or orderings, potentially limiting the information agents can express. We propose a novel class of mechanisms that elicit agent reports in natural language and leverage the world-modeling capabilities of large language models (LLMs) to select outcomes and assign payoffs. We identify sufficient conditions for these mechanisms to be incentive-compatible and efficient as the LLM being a good enough world model and a strong inter-agent information over-determination condition. We show situations where these LM-based mechanisms can successfully aggregate information in signal structures on which prediction markets fail.