LLM-Powered Agents - Index

Routing Summary

Large language models as the decision engine of simulated agents: generative agents with memory, “silicon samples” of survey respondents, Homo silicus economic subjects, and the validity problems all three share. Anchored by Gao et al. (2023 survey), Park et al. (2023), Argyle et al. (2022), Horton, Filippas & Manning (2023/2026) and Santurkar et al. (2023). Complements the rule-based consumer ABM, calibration, WOM and diffusion notes elsewhere in Agent-Based Modeling.

• Need the big picture / where to start? → LLM-Powered Agents - Overview
• Need the memory stream, retrieval score, reflection and planning loop (Smallville)? → Generative Agents Architecture - Memory, Reflection and Planning
• Need to simulate survey respondents from demographic backstories, or the four fidelity criteria? → Silicon Samples and Algorithmic Fidelity
• Need LLMs as experimental subjects in economics (fairness, dictator games, status quo bias, minimum wage)? → Homo Silicus - LLMs as Simulated Economic Agents
• Need to fit a population of personas to human choice data? → simplex-constrained mixture weights
• Need to decide between LLM agents, rule-based agents, or a hybrid? → LLM Agents vs Rule-Based Agents in ABM
• Need a metric for how well an LLM matches a group’s opinion distribution? → consistency
• Need the threat taxonomy and a validation protocol before trusting silicon data? → validation-and-calibration protocol

Concept Map

Concept	Note	Type	Depends On	Key Result
LLM-empowered ABM paradigm	LLM-Powered Agents - Overview	overview	ABM Methodology; Agent Decision Rules; Heterogeneity	Four LLM abilities (perception, reasoning, adaptation, heterogeneity) and four challenges (environment, alignment/personalisation, action simulation, evaluation); micro vs. macro evaluation
Generative agents	Generative Agents Architecture - Memory, Reflection and Planning	method	Overview; Emergent Phenomena	score = recency ($0.995^{\Delta h}$) + importance (LLM 1–10) + relevance (cosine); reflection at importance sum > 150; party awareness 4%→52%, density 0.167→0.74; full architecture TrueSkill 29.89 vs 21.21, $d=8.16$
Silicon sampling	Silicon Samples and Algorithmic Fidelity	concept	Overview; Poststratification	$\hat{P}(V)={\sum }_b{P}_{LM}(V\mid b)P_{ANES}(b)$; four fidelity criteria; tetrachoric 0.90/0.92/0.94; Cramér’s $V$ mean diff $-0.026$
Homo silicus	Homo Silicus - LLMs as Simulated Economic Agents	concept	Silicon Samples; Decision Rules; Discrete Choice	Five recapitulations; “theory in flexibly executable form”; confusion-matrix view; results need empirical confirmation
Persona mixture calibration	Persona Mixture Calibration of LLM Agents	method	Homo Silicus; ABM Calibration Overview	${\min }_w\Vert {\sum }_k{w}_k{v}_k-v_H{\Vert }^2$ on the simplex; out-of-sample MSE 0.094 vs 0.182
LLM vs rule-based agents	LLM Agents vs Rule-Based Agents in ABM	concept	Overview; Decision Rules; Calibration; Validation	Calibration burden becomes validation burden; “judge and jury” and Lucas-critique arguments; hybrid designs
Opinion alignment metrics	Opinion Alignment Metrics for Language Models	method	Silicon Samples	$\mathcal{A}=1-\mathcal{WD}/(N-1)$; every human group beats every LM on representativeness; steering modest; RLHF modal collapse
Validity and calibration	Validity, Bias and Calibration of LLM-Simulated Populations	concept	all of the above; ABM Validation Challenges; Poststratification	15-threat taxonomy; imputed-context confounding; 10-step validation protocol; prediction-powered inference

Notes

• LLM-Powered Agents - Overview — CONTAINS: LLM-agent definition, classical agent desiderata, Gao et al.’s four abilities and four challenges, micro/macro evaluation, application domains, open problems (scaling, benchmarks, robustness, ethics), relevance to marketing measurement, minimal agent-step sketch.
• Generative Agents Architecture - Memory, Reflection and Planning — CONTAINS: memory stream definition, retrieval scoring formula and constants, reflection algorithm and reflection trees, recursive planning and reaction, environment grounding, ablation results table, emergent diffusion/network/coordination numbers, failure modes, retrieval code sketch.
• Silicon Samples and Algorithmic Fidelity — CONTAINS: LM-as-conditional-distribution, algorithmic fidelity definition, four criteria, silicon-sampling marginal-correction identity and its equivalence to poststratification, Studies 1–3 results, critical reading, worked reweighting example, pipeline sketch.
• Homo Silicus - LLMs as Simulated Economic Agents — CONTAINS: Homo silicus definition, latent vs. explicit social information, five recapitulated experiments with designs and numbers, Charness–Rabin utility, theory-as-instruction argument, Fig. 7 confusion matrix, prediction-powered inference, when simulations are informative, in-silico pilot sketch.
• Persona Mixture Calibration of LLM Agents — CONTAINS: theory-grounded persona definition, six-step mixture algorithm, reported weights per model, out-of-sample MSE, comparison table of calibration strategies, identification/uncertainty caveats with Bayesian extension, worked constrained-least-squares example.
• LLM Agents vs Rule-Based Agents in ABM — CONTAINS: side-by-side comparison table, Gao’s four advantages, Horton’s “judge and jury” and Lucas-critique arguments, cost catalogue, four hybrid design patterns, same-decision two-agent example.
• Opinion Alignment Metrics for Language Models — CONTAINS: OpinionQA construction, human/model opinion distributions, Wasserstein alignment (Eq. 1), representativeness/steerability/consistency definitions, empirical findings, implications for ABM, hand-computed alignment examples, metric code.
• Validity, Bias and Calibration of LLM-Simulated Populations — CONTAINS: 15-row threat taxonomy with source evidence, poststratification limits, social-influence bias, imputed-context confounding in potential-outcomes terms, 10-step protocol, four-outcomes definition, disagreements across sources, brand-tracker checklist, dispersion check code.

External / Cross-Folder Links

• Agent Decision Rules and Bounded Rationality, Logit Purchase Decision Model, Consumer Utility Function Components — the rule-based decision models LLM agents replace or complement.
• Heterogeneity in Agent Models, Population Initialization and Parameter Sensitivity — numeric heterogeneity vs. persona heterogeneity.
• ABM Calibration Overview, Approximate Bayesian Computation for ABMs, History Matching for ABMs, HM-ABC Calibration Framework — classical calibration.
• ABM Validation Challenges — micro vs. macro validation.
• Word of Mouth Mechanisms, Opinion Leaders and Social Influence, Network Topology Effects on Diffusion, Product Adoption and Diffusion Models, Emergent Phenomena in ABM — social dynamics that generative agents produce endogenously.
• ABM in Marketing Strategy — marketing applications.
• Poststratification, Hierarchical Models, Posterior Predictive Checking, Prior Predictive Checking — Bayesian workflow tools for reweighting and checking silicon populations.
• Discrete Choice Models — analysis model for human and silicon choice data.
• Potential Outcomes Framework, Multiple Testing Corrections, Power Analysis and Sample Size — experimental-design concepts used in the validity discussion.
• LLM Expert Elicitation for Bayesian Networks — LLMs as stand-ins for human experts elsewhere in the vault.

Sources

• Gao 2023 - LLM Empowered Agent-Based Modeling Survey — Gao, C., Lan, X., Li, N., Yuan, Y., Ding, J., Zhou, Z., Xu, F. & Li, Y. (2023), “Large Language Models Empowered Agent-based Modeling and Simulation: A Survey and Perspectives,” arXiv:2312.11970.
• Park 2023 - Generative Agents Interactive Simulacra — Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P. & Bernstein, M. S. (2023), “Generative Agents: Interactive Simulacra of Human Behavior,” UIST ‘23, arXiv:2304.03442.
• Argyle 2022 - Out of One Many Silicon Samples — Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J., Rytting, C. & Wingate, D. (2022), “Out of One, Many: Using Language Models to Simulate Human Samples,” arXiv:2209.06899.
• Horton 2023 - Homo Silicus LLMs as Simulated Economic Agents — Horton, J. J., Filippas, A. & Manning, B. S. (2023; v2 2026), “Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?”, arXiv:2301.07543.
• Santurkar 2023 - Whose Opinions Do Language Models Reflect — Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P. & Hashimoto, T. (2023), “Whose Opinions Do Language Models Reflect?”, arXiv:2303.17548.