Structuring Collective Action with LLM-Guided Evolution: From Ill-Structured Problems to Executable Heuristics

TL;DR

<3-5 sentence high-level summary> ECHO-MIMIC tackles ill-structured collective-action problems by decomposing them into a sequence of Well-Structured Problems, enabling agents to operate via executable heuristics while policymakers deploy personalized nudges to incentivize alignment with system-wide goals. The framework combines LLM-driven evolutionary search with explicit fitness criteria to discover both local decision rules and global targets, and it couples these with language-based persuasion to drive adoption. Experiments in agricultural landscape management and carbon-aware EV charging demonstrate superior performance over strong baselines and reveal how heuristics and nudges adapt to domain context and agent personas. Together, ECHO-MIMIC offers a scalable, adaptable pathway for domain-agnostic policy design that can incorporate real-world feedback and governance considerations.

Abstract

Collective action problems, which require aligning individual incentives with collective goals, are classic examples of Ill-Structured Problems (ISPs). For an individual agent, the causal links between local actions and global outcomes are unclear, stakeholder objectives often conflict, and no single, clear algorithm can bridge micro-level choices with macro-level welfare. We present ECHO-MIMIC, a general computational framework that converts this global complexity into a tractable, Well-Structured Problem (WSP) for each agent by discovering executable heuristics and persuasive rationales. The framework operates in two stages: ECHO (Evolutionary Crafting of Heuristics from Outcomes) evolves snippets of Python code that encode candidate behavioral policies, while MIMIC (Mechanism Inference \& Messaging for Individual-to-Collective Alignment) evolves companion natural language messages that motivate agents to adopt those policies. Both phases employ a large-language-model-driven evolutionary search: the LLM proposes diverse and context-aware code or text variants, while population-level selection retains those that maximize collective performance in a simulated environment. We demonstrate this framework on two distinct ISPs: a canonical agricultural landscape management problem and a carbon-aware EV charging time slot usage problem. Results show that ECHO-MIMIC discovers high-performing heuristics compared to baselines and crafts tailored messages that successfully align simulated agent behavior with system-level goals. By coupling algorithmic rule discovery with tailored communication, ECHO-MIMIC transforms the cognitive burden of collective action into a implementable set of agent-level instructions, making previously ill-structured problems solvable in practice and opening a new path toward scalable, adaptive policy design.

Figures (13)

• Figure 1: ECHO–MIMIC framework. ECHO uses an LLM+EA search loop to propose personal-level decision heuristics aligned with baseline (stage 2) and global (stage 3) objectives. MIMIC optimizes personalized nudges (e.g., messages/mechanisms/policies) using an LLM+EA search loop to drive collective action. Overall, the framework converts the collective action ISP into system- and agent-level WSPs. Figure uses the farm domain. See Appendix \ref{['app:workflow_components']} for a more detailed workflow.
• Figure 2: Domain specific actions. a) Interventions resulting from ECHO learned baseline heuristics in stage 2 for the farm domain. The interventions match the ground-truth baseline computed from stage 1 closely. For the comparison with ground-truth, ECHO stage 3 predictions, and synthetically generated farm geometries, see Appendix \ref{['app:add_results']}. b) Synthetically generated EV charging spatio-temporal configuration. Five agents are placed in a line, each with their own charging demand, preferences, and carbon intensity. They are allowed to specify usage in four time slots for a week.
• Figure 3: ECHO stage 2 results. a) Accuracy ($1-error$) over generations in stage 2 for the farm domain. Some farms are easier to make progress in (farms 2, 5) compared to others (farms 1, 3, 4), and distinct capabilities emerge in harder farms as generations progress. b) Total fitness ($1/error$) delta for both the domains together, resulting from LLM variation operators, summed across generations for the best performing program at the end. Crossover and mutate have the highest positive cumulative change in fitness. c) Accuracy versus normalized complexity metrics of the heuristics for farm 3 in the farm domain. Increased Halstead metrics are correlated with increased accuracy, upto a point, followed by a decrease. d) Accuracy over generations with and without Halstead instructions for farm 3 in the farm domain. Adding additional Halstead instructions to the prompt provides free gains in accuracy at the expense of interpretability.
• Figure 4: MIMIC nudge discovery and personalization. Accuracy of nudges over generations a) for the EV charging domain with versatile personas and generic instructions. At different points MIMIC learns to use personalized benefits, social proof, and environmental impact framing. b) for the farm domain (hard farms) with persona and nudge type specific instructions. P refers to personas and N refers to nudge types. Personas are either Resistant, Economic, or Social. Nudge Types are either Behavioral or Economic.
• Figure 5: ECHO–MIMIC framework. a) ECHO uses an LLM-evolutionary search loop to propose, score, and select farm-level decision heuristics aligned with baseline (stage 2) and global (stage 3) objectives. (b) MIMIC optimizes personalized nudges (e.g., messages/mechanisms/policies) using an LLM-evolutionary search loop, evaluates nudges using simulated agent responses, and iteratively updates nudges to drive collective action. Illustration uses the farm domain as an example. See Fig. \ref{['fig:app:subroutines']} showing the two subroutines of fitness evaluation and LLM-driven variation.