Can AI with High Reasoning Ability Replicate Human-like Decision Making in Economic Experiments?

Can AI with High Reasoning Ability Replicate Human-like Decision Making in Economic Experiments? Ayato Kitadai Sinndy Dayana Rico Lugo Yudai Tsurusaki Yusuke Fukasawa Nariaki Nishino

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Can AI with High Reasoning Ability Replicate Human-like Decision Making in Economic Experiments? 2025 Group Decision and Negotiation 10.1007/s10726-025-09946-9 kitadai2025-can-ai-with-high-reasoning-ability kitadai2025-can-ai-with-high-reasoning-ability.txt, the project's staged copy 144fe84a11b4b62dcc1162385844e7d3738db57c open-access experimental-economics

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them null-result Quotation withheld because this record is not cleared for redistribution. LLM generative agents in the ultimatum game Quotation withheld because this record is not cleared for redistribution. simulation simulation Quotation withheld because this record is not cleared for redistribution. mechanisms Persona settings act as a proxy that embeds social preferences and bounded rationality into the agents; without them the agents' improved reasoning drives them to the subgame-perfect solution instead of to human-like fairness behaviour Quotation withheld because this record is not cleared for redistribution. mechanisms Randomly generated persona attributes in earlier studies injected noise into agent behaviour, which masked the effect of reasoning ability and prevented it from being evaluated systematically Quotation withheld because this record is not cleared for redistribution. mechanisms Skew in how the agents' offers were distributed across offer values drives the fitted acceptance-rate lines to impossible values above 1.0 or below 0, producing the anomalies seen in the responder-side figures Quotation withheld because this record is not cleared for redistribution. Excluding persona settings in order to isolate reasoning ability is itself a limitation, since personas are known to shape human-like behaviour. what this limits: the transfer to agents with personas: WHILE OUR METHODOLOGICAL CHOICE CLARIFIES THE ROLE OF REASONING ABILITY, IT ALSO REPRESENTS A LIMITATION OF THIS STUDY - the excluded factor is known to matter Quotation withheld because this record is not cleared for redistribution. The sample size of 1000 decisions per condition was chosen pragmatically rather than by a formal power analysis. what this limits: the precision of every estimate: WHILE THIS NUMBER IS NOT DERIVED FROM A FORMAL POWER ANALYSIS, IT WAS SELECTED BASED ON TWO MAIN CONSIDERATIONS Quotation withheld because this record is not cleared for redistribution. Scope is restricted to decision-making and negotiation tasks; market-level and institutional applications are left to future work. what this limits: the scope of the conclusions: BROADER APPLICATIONS TO MARKET-LEVEL MODELING OR INSTITUTIONAL DESIGN REMAIN IMPORTANT DIRECTIONS FOR FUTURE WORK Quotation withheld because this record is not cleared for redistribution. Earlier claims of success are attributed to the coarseness of the evaluation rather than to the agents: aggregate distributions and binary alignment, with no individual-level analysis. what this limits: published findings that generative agents reproduce human behaviour in economic games: this paper states those successes may stem from coarse-grained criteria — aggregate distributions or binary alignment — without individual-level or reasoning-process analysis Quotation withheld because this record is not cleared for redistribution. The expectation the title asks about is not borne out: greater reasoning ability does not directly produce reproduction of the experiments. what this limits: the assumption that a more capable model will simulate an economic experiment better: this paper reports zero-shot prompting with an older model aligning closely with human behaviour, and states improving reasoning ability cannot lead directly to reproduction Quotation withheld because this record is not cleared for redistribution. The paper's own explanation of why earlier work succeeded is a claim about MEASUREMENT RESOLUTION, not about behaviour: prior studies assessed aggregate distributions or binary alignment 'without deeper analysis of underlying reasoning processes or individual-level deviations'. On that account the earlier results are not wrong, they are coarse - which is a harder thing to test. Quotation withheld because this record is not cleared for redistribution. Agent outputs are reported as highly sensitive to design choices - persona, prompt formulation, role instructions - so any single configuration's result is conditional on choices the paper had to make and could not vary at once. Quotation withheld because this record is not cleared for redistribution. Prompt-engineering techniques for eliciting language-model reasoning are imported into the agent-based-modelling tradition of experimental economics, with prompts standing in for experimental instructions to human subjects. Quotation withheld because this record is not cleared for redistribution. the in-context examples RE-DRAWN AT RANDOM from the raw data and the whole simulation re-run several times, alongside runs at 15 and 20 examples eliminates: the particular examples chosen for the prompt as the source of the result - an LLM's behaviour is conditioned on its examples, so a finding from one draw is a fact about that draw · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. generative agent generative agent Quotation withheld because this record is not cleared for redistribution. economic experiments economic experiments Quotation withheld because this record is not cleared for redistribution. behavioral realism behavioral realism Quotation withheld because this record is not cleared for redistribution. reasoning ability reasoning ability Quotation withheld because this record is not cleared for redistribution. theoretical solution theoretical solution Quotation withheld because this record is not cleared for redistribution. persona persona Quotation withheld because this record is not cleared for redistribution. one-shot ultimatum game one-shot ultimatum game Quotation withheld because this record is not cleared for redistribution. history-dependent learning history-dependent learning Quotation withheld because this record is not cleared for redistribution. reputation effects reputation effects Quotation withheld because this record is not cleared for redistribution. subgame-perfect equilibrium subgame-perfect equilibrium Quotation withheld because this record is not cleared for redistribution. theoretical benchmark theoretical benchmark Quotation withheld because this record is not cleared for redistribution. human-AI performance gap human-AI performance gap Quotation withheld because this record is not cleared for redistribution. stringent test bed stringent test bed Quotation withheld because this record is not cleared for redistribution. artificial correlations artificial correlations Quotation withheld because this record is not cleared for redistribution. weakly dominant strategy weakly dominant strategy Quotation withheld because this record is not cleared for redistribution. perceived unfairness perceived unfairness Quotation withheld because this record is not cleared for redistribution. LLM agents LLM agents Quotation withheld because this record is not cleared for redistribution. traditional economic experiments traditional economic experiments Quotation withheld because this record is not cleared for redistribution. additional methodological tool additional methodological tool Quotation withheld because this record is not cleared for redistribution. opaque biases opaque biases Quotation withheld because this record is not cleared for redistribution. harmful stereotypes harmful stereotypes Quotation withheld because this record is not cleared for redistribution. structural inequities structural inequities Quotation withheld because this record is not cleared for redistribution. ground-truth benchmarks ground-truth benchmarks Quotation withheld because this record is not cleared for redistribution. LLM-based simulations LLM-based simulations Quotation withheld because this record is not cleared for redistribution. boundary conditions boundary conditions Quotation withheld because this record is not cleared for redistribution. outputs of generative agents outputs of generative agents Quotation withheld because this record is not cleared for redistribution. simulation design choices simulation design choices Quotation withheld because this record is not cleared for redistribution. Generative AI Generative AI Quotation withheld because this record is not cleared for redistribution. generative models generative models Quotation withheld because this record is not cleared for redistribution. discriminative models discriminative models Quotation withheld because this record is not cleared for redistribution. underlying data distribution underlying data distribution Quotation withheld because this record is not cleared for redistribution. recognition to generation recognition to generation Quotation withheld because this record is not cleared for redistribution. methodological transformation methodological transformation Quotation withheld because this record is not cleared for redistribution. LLMs LLMs Quotation withheld because this record is not cleared for redistribution. key subclass key subclass Quotation withheld because this record is not cleared for redistribution. Transformer architecture Transformer architecture Quotation withheld because this record is not cleared for redistribution. statistical knowledge base statistical knowledge base Quotation withheld because this record is not cleared for redistribution. social norms social norms Quotation withheld because this record is not cleared for redistribution. strategic heuristics strategic heuristics Quotation withheld because this record is not cleared for redistribution. foundation models foundation models Quotation withheld because this record is not cleared for redistribution. general-purpose substrates general-purpose substrates Quotation withheld because this record is not cleared for redistribution. downstream tasks downstream tasks Quotation withheld because this record is not cleared for redistribution. retraining retraining Quotation withheld because this record is not cleared for redistribution. hallucinations hallucinations Quotation withheld because this record is not cleared for redistribution. learned statistical associations learned statistical associations Quotation withheld because this record is not cleared for redistribution. grounded world models grounded world models Quotation withheld because this record is not cleared for redistribution. black boxes black boxes Quotation withheld because this record is not cleared for redistribution. interpretability interpretability Quotation withheld because this record is not cleared for redistribution. prompt design prompt design Quotation withheld because this record is not cleared for redistribution. few-shot prompting few-shot prompting Quotation withheld because this record is not cleared for redistribution. chain-of-thought (CoT) prompting chain-of-thought (CoT) prompting Quotation withheld because this record is not cleared for redistribution. general mitigation techniques general mitigation techniques Quotation withheld because this record is not cleared for redistribution. scaffolding methods scaffolding methods Quotation withheld because this record is not cleared for redistribution. prompting strategies prompting strategies Quotation withheld because this record is not cleared for redistribution. reasoning quality reasoning quality Quotation withheld because this record is not cleared for redistribution. coherence, inference depth, and contextual relevance coherence, inference depth, and contextual relevance Quotation withheld because this record is not cleared for redistribution. contextually coherent, multi-step responses contextually coherent, multi-step responses Quotation withheld because this record is not cleared for redistribution. inferential structure inferential structure Quotation withheld because this record is not cleared for redistribution. observable behavioral property observable behavioral property Quotation withheld because this record is not cleared for redistribution. experimental variable experimental variable Quotation withheld because this record is not cleared for redistribution. pragmatist perspective pragmatist perspective Quotation withheld because this record is not cleared for redistribution. human cognition human cognition Quotation withheld because this record is not cleared for redistribution. behavioral patterns behavioral patterns Quotation withheld because this record is not cleared for redistribution. methodological tools methodological tools Quotation withheld because this record is not cleared for redistribution. well-defined simulation contexts well-defined simulation contexts Quotation withheld because this record is not cleared for redistribution. empirical tools empirical tools Quotation withheld because this record is not cleared for redistribution. evaluation protocols evaluation protocols Quotation withheld because this record is not cleared for redistribution. methodological safeguards methodological safeguards Quotation withheld because this record is not cleared for redistribution. coarse-grained evaluation criteria coarse-grained evaluation criteria Quotation withheld because this record is not cleared for redistribution. aggregate distributions aggregate distributions Quotation withheld because this record is not cleared for redistribution. individual-level deviations individual-level deviations Quotation withheld because this record is not cleared for redistribution. finer-grained behavioral criteria finer-grained behavioral criteria Quotation withheld because this record is not cleared for redistribution. game-theoretic environments game-theoretic environments Quotation withheld because this record is not cleared for redistribution. failures of alignment failures of alignment Quotation withheld because this record is not cleared for redistribution. behavioral Turing tests behavioral Turing tests Quotation withheld because this record is not cleared for redistribution. latent utility function latent utility function Quotation withheld because this record is not cleared for redistribution. perceived social harmony perceived social harmony Quotation withheld because this record is not cleared for redistribution. material self-interest material self-interest Quotation withheld because this record is not cleared for redistribution. Persona design Persona design Quotation withheld because this record is not cleared for redistribution. behavioral tendencies behavioral tendencies Quotation withheld because this record is not cleared for redistribution. causal effects causal effects Quotation withheld because this record is not cleared for redistribution. reasoning-related prompting strategy reasoning-related prompting strategy Quotation withheld because this record is not cleared for redistribution. confounding variables confounding variables Quotation withheld because this record is not cleared for redistribution. central research question central research question Quotation withheld because this record is not cleared for redistribution. normative benchmarks of rationality normative benchmarks of rationality Quotation withheld because this record is not cleared for redistribution. empirically grounded human behavior empirically grounded human behavior Quotation withheld because this record is not cleared for redistribution. simulation framework simulation framework Quotation withheld because this record is not cleared for redistribution. three sequential steps three sequential steps Quotation withheld because this record is not cleared for redistribution. output format output format Quotation withheld because this record is not cleared for redistribution. 1000 independent decisions 1000 independent decisions Quotation withheld because this record is not cleared for redistribution. formal power analysis formal power analysis Quotation withheld because this record is not cleared for redistribution. robustness check robustness check Quotation withheld because this record is not cleared for redistribution. computational cost per generation computational cost per generation Quotation withheld because this record is not cleared for redistribution. statistical resolution statistical resolution Quotation withheld because this record is not cleared for redistribution. feasibility in simulation-based research feasibility in simulation-based research Quotation withheld because this record is not cleared for redistribution. substitutes for human participants substitutes for human participants Quotation withheld because this record is not cleared for redistribution. agent-based modeling agent-based modeling Quotation withheld because this record is not cleared for redistribution. natural language reasoning natural language reasoning Quotation withheld because this record is not cleared for redistribution. prompt engineering prompt engineering Quotation withheld because this record is not cleared for redistribution. Monte Carlo methods Monte Carlo methods Quotation withheld because this record is not cleared for redistribution. experimental instructions experimental instructions Quotation withheld because this record is not cleared for redistribution. human-subject experiments human-subject experiments Quotation withheld because this record is not cleared for redistribution. Anderson-Darling (AD) Anderson-Darling (AD) Quotation withheld because this record is not cleared for redistribution. Kolmogorov-Smirnov (KS) Kolmogorov-Smirnov (KS) Quotation withheld because this record is not cleared for redistribution. overall distributional differences overall distributional differences Quotation withheld because this record is not cleared for redistribution. maximum cumulative distribution differences maximum cumulative distribution differences Quotation withheld because this record is not cleared for redistribution. social context social context Quotation withheld because this record is not cleared for redistribution. stand-ins for real human groups stand-ins for real human groups Quotation withheld because this record is not cleared for redistribution. treating LLMs as direct substitutes for human subjects treating LLMs as direct substitutes for human subjects Quotation withheld because this record is not cleared for redistribution. flawed approach flawed approach Quotation withheld because this record is not cleared for redistribution. methodological scaffolding methodological scaffolding Quotation withheld because this record is not cleared for redistribution. Persona settings Persona settings Quotation withheld because this record is not cleared for redistribution. contextual variables contextual variables Quotation withheld because this record is not cleared for redistribution. human decision-making human decision-making Quotation withheld because this record is not cleared for redistribution. inequity aversion inequity aversion Quotation withheld because this record is not cleared for redistribution. fairness equilibrium fairness equilibrium Quotation withheld because this record is not cleared for redistribution. bounded rationality bounded rationality Quotation withheld because this record is not cleared for redistribution. social preferences social preferences Quotation withheld because this record is not cleared for redistribution. Variability in reasoning ability and persona configurations Variability in reasoning ability and persona configurations Quotation withheld because this record is not cleared for redistribution. agent behavior agent behavior Quotation withheld because this record is not cleared for redistribution. simulation outcomes simulation outcomes Quotation withheld because this record is not cleared for redistribution. LLM agents → traditional economic experiments LLM agents are proposed to augment rather than replace traditional economic experiments. Augments-but-does-not-replace is absent from the eight types; distinct-from would lose the methodological complement relation. Quotation withheld because this record is not cleared for redistribution. ground-truth benchmarks → LLM-based simulations Human data remain indispensable benchmarks for calibrating LLM simulations and identifying their boundary conditions. Benchmark-for/calibrates is absent from the eight types and is not mere operationalisation. Quotation withheld because this record is not cleared for redistribution. outputs of generative agents → simulation design choices Generative-agent outputs are highly sensitive to simulation design choices. Depends-on/sensitive-to is absent; causation would overstate a design dependence with multiple interacting parameters. Quotation withheld because this record is not cleared for redistribution. foundation models → downstream tasks Foundation models are adaptable to downstream tasks without retraining. Adaptable-to is absent from the eight types; neither kindhood nor operationalisation captures transfer to a task. Quotation withheld because this record is not cleared for redistribution. prompt engineering → experimental instructions Prompt engineering serves as an analog to instructions given to human participants. Analogous-to is absent; same-as would erase the distinction between machine prompts and human-study instructions. Quotation withheld because this record is not cleared for redistribution. social context → stand-ins for real human groups Prompts incorporating social context are a necessary condition for advanced LLMs to serve as human-like stand-ins. Prerequisite-for is absent from the eight relation types. Quotation withheld because this record is not cleared for redistribution. Persona settings → social preferences Persona inclusion may act as a proxy for embedding social preferences in the simulation. Proxy-for is absent; operationalised-as would claim a measurement definition the paper does not make. Quotation withheld because this record is not cleared for redistribution. Persona settings → bounded rationality Persona inclusion may proxy bounded rationality in LLM-driven simulations. Proxy-for is absent from the eight types and the paper presents this as a hypothesis. Quotation withheld because this record is not cleared for redistribution. replication of human behaviour - AT AGGREGATE GRAIN: distributions of choices or binary alignment with typical outcomes, which is the criterion under which earlier work succeeded Quotation withheld because this record is not cleared for redistribution. replication of human behaviour - AT INDIVIDUAL GRAIN: underlying reasoning processes and individual-level deviations, which the earlier evaluations did not examine Quotation withheld because this record is not cleared for redistribution. agent quality - split into two properties that trade off against each other: behavioral realism and rationality Quotation withheld because this record is not cleared for redistribution. successful replication - in the prior literature a claim about AGGREGATE DISTRIBUTIONS or binary alignment with typical outcomes; this paper attributes those successes to the coarseness of the criterion rather than to the agents Quotation withheld because this record is not cleared for redistribution. One-shot ultimatum game - a demanding test that removes learning and reputation and supplies a unique equilibrium benchmark Quotation withheld because this record is not cleared for redistribution. Human low-offer rejection - payoff sacrifice explained as resistance to perceived unfairness Quotation withheld because this record is not cleared for redistribution. Generative AI - distribution-learning synthesis rather than classification within existing boundaries Quotation withheld because this record is not cleared for redistribution. GPT-4 as behavioral model - a statistical knowledge base reflecting norms, heuristics, and communication conventions Quotation withheld because this record is not cleared for redistribution. Hallucination - plausible but wrong output arising from association-based rather than grounded modeling Quotation withheld because this record is not cleared for redistribution. Few-shot prompting - annotated examples guiding output style and logic Quotation withheld because this record is not cleared for redistribution. Chain-of-thought prompting - explicit instruction to articulate intermediate reasoning steps Quotation withheld because this record is not cleared for redistribution. Reasoning ability - observable coherent multi-step response structure, not a claim about mental models or logical soundness Quotation withheld because this record is not cleared for redistribution. LLMs as empirical tools - outputs analyzed, compared, and validated like participant data under safeguards Quotation withheld because this record is not cleared for redistribution. Replication success under coarse criteria - aggregate distributions or binary typical-outcome alignment without individual reasoning Quotation withheld because this record is not cleared for redistribution. Latent utility hypothesis - agents may optimize perceived social harmony rather than material self-interest Quotation withheld because this record is not cleared for redistribution. Isolated reasoning manipulation - vary zero-shot, few-shot, and chain-of-thought while holding persona, framing, and temperature outside the focal contrast Quotation withheld because this record is not cleared for redistribution. Core validity question - convergence toward normative rationality versus divergence from observed human behavior Quotation withheld because this record is not cleared for redistribution. Simulation sample - 1000 independent decisions balancing statistical resolution against chain-of-thought cost Quotation withheld because this record is not cleared for redistribution. Prompt engineering - methodological analog of instructions delivered to human experimental participants Quotation withheld because this record is not cleared for redistribution. AD versus KS - overall distributional difference versus maximum cumulative distribution difference Quotation withheld because this record is not cleared for redistribution. LLMs as human stand-ins - currently unwarranted without explicit social context and methodological scaffolding Quotation withheld because this record is not cleared for redistribution. Persona settings - configured cultural, age, and gender context needed because current models do not inherently possess it Quotation withheld because this record is not cleared for redistribution. Persona as behavioral proxy - possible embedding of social preferences and bounded rationality Quotation withheld because this record is not cleared for redistribution. extends → Kitadai et al. (2023) LLM-driven ultimatum-game simulation Quotation withheld because this record is not cleared for redistribution. contradicts → Horton (2023) and Aher et al. (2023) reports of successful behavioural replication by LLM agents Quotation withheld because this record is not cleared for redistribution. builds-on → prior LLM-agent studies that confounded reasoning with persona, payoff framing and temperature Quotation withheld because this record is not cleared for redistribution. reasoning ability —causes→ theoretical solution reasoning ability theoretical solution differs in: WHICH TARGET THE AGENTS CONVERGE ON, and the two targets pull apart: raising reasoning ability moves the simulation TOWARDS the game-theoretic solution and AWAY from the human experimental result. Better reasoning makes the agent a worse model of people - the paper's title asks the question and its abstract answers no · strength of the relation: 0.9 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but higher reasoning moved results closer to the theoretical solution; causes is loose. Quotation withheld because this record is not cleared for redistribution. reasoning ability —contrasted-with→ behavioral realism reasoning ability behavioral realism strength of the relation: 0.8 · note: The record already carries a sense saying agent quality splits into two properties that TRADE OFF. This edge is that trade-off as the paper measured it. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but higher reasoning moved results away from real experimental results; contrasted-with 'behavioral realism' is loose. Quotation withheld because this record is not cleared for redistribution. Generative AI —distinct-from→ discriminative models Generative AI discriminative models the relation itself is separately evidenced · strength of the relation: 1 · note: The paper explicitly contrasts distribution-learning generation with classification and prediction. Quotation withheld because this record is not cleared for redistribution. opaque biases —causes→ harmful stereotypes opaque biases harmful stereotypes the relation itself is separately evidenced · strength of the relation: 0.7 · note: The paper says biases can perpetuate or amplify harmful stereotypes. Quotation withheld because this record is not cleared for redistribution. scaffolding methods —distinct-from→ general mitigation techniques scaffolding methods general mitigation techniques the relation itself is separately evidenced · strength of the relation: 1 · note: The authors explicitly adopt the strategies not as general mitigation but as reasoning scaffolds. Quotation withheld because this record is not cleared for redistribution. prompting strategies —causes→ coherence, inference depth, and contextual relevance prompting strategies coherence, inference depth, and contextual relevance the relation itself is separately evidenced · strength of the relation: 0.9 · note: Prior work is reported to show these strategies enhance all three output properties. Quotation withheld because this record is not cleared for redistribution. perceived social harmony —contrasted-with→ material self-interest perceived social harmony material self-interest the relation itself is separately evidenced · strength of the relation: 0.7 · note: The proposed implicit objective is social harmony rather than material self-interest. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the claim is what Mei et al. (2024) 'proposed'; attributed. Quotation withheld because this record is not cleared for redistribution. treating LLMs as direct substitutes for human subjects —is-a→ flawed approach treating LLMs as direct substitutes for human subjects flawed approach the relation itself is separately evidenced · strength of the relation: 1 · note: The paper directly classifies treating LLMs as substitutes as a flawed approach. Quotation withheld because this record is not cleared for redistribution. contextual variables —causes→ human decision-making contextual variables human decision-making the relation itself is separately evidenced · strength of the relation: 1 · note: Culture, age, and gender are said to significantly influence human decision-making. Quotation withheld because this record is not cleared for redistribution. Variability in reasoning ability and persona configurations —causes→ agent behavior Variability in reasoning ability and persona configurations agent behavior the relation itself is separately evidenced · strength of the relation: 0.7 · note: The paper states this variability can significantly influence agent behavior. Quotation withheld because this record is not cleared for redistribution. Anderson-Darling test Anderson-Darling test Quotation withheld because this record is not cleared for redistribution. Kolmogorov-Smirnov test Kolmogorov-Smirnov test Quotation withheld because this record is not cleared for redistribution. p-value p-value Quotation withheld because this record is not cleared for redistribution. regression regression Quotation withheld because this record is not cleared for redistribution. confidence interval confidence interval Quotation withheld because this record is not cleared for redistribution. Raising the reasoning ability of LLM-driven agents pushed their ultimatum-game behaviour toward the game-theoretic solution and away from the real human experimental data. Quotation withheld because this record is not cleared for redistribution. No prompting condition produced a distribution that statistically matched either the theoretical solution or the human experimental data. Quotation withheld because this record is not cleared for redistribution. On the proposer side the older, weaker-reasoning model with zero-shot prompting was also relatively close to human behaviour, showing that better reasoning does not straightforwardly buy better reproduction of economic experiments. Quotation withheld because this record is not cleared for redistribution. On the responder side, chain-of-thought prompting produced the closest match to the theoretical (rational) distribution of any condition. Quotation withheld because this record is not cleared for redistribution. Against human responder data the ranking inverted: the weakest-reasoning condition, zero-shot gpt-3.5, matched human acceptance behaviour most closely. Quotation withheld because this record is not cleared for redistribution. The authors conclude that LLM agents do not spontaneously behave like humans in economic games without prompts carrying social context, so using LLM-driven multi-agent simulations as substitutes for real human groups is currently unwarranted. Quotation withheld because this record is not cleared for redistribution. At every temperature setting almost all agents offered a 50-50 split of coins. Quotation withheld because this record is not cleared for redistribution. Because every other setting was identical to the earlier study, the authors attribute the difference in results to the presence or absence of persona settings, and conclude that personas are essential for applying LLM-driven simulation to economic experiments. Quotation withheld because this record is not cleared for redistribution. Robustness checks with 15 and 20 in-prompt examples changed the synthesised distributions only slightly and left the comparative pattern across conditions unchanged. Quotation withheld because this record is not cleared for redistribution. reasoning ability reasoning ability Quotation withheld because this record is not cleared for redistribution. generative agent generative agent Quotation withheld because this record is not cleared for redistribution. Platonic Representation Hypothesis Platonic Representation Hypothesis Quotation withheld because this record is not cleared for redistribution. reasoning-related prompting strategy reasoning-related prompting strategy Quotation withheld because this record is not cleared for redistribution. temperature parameter temperature parameter Quotation withheld because this record is not cleared for redistribution. GPT model GPT model Quotation withheld because this record is not cleared for redistribution. acceptance rate acceptance rate Quotation withheld because this record is not cleared for redistribution. persona settings persona settings Quotation withheld because this record is not cleared for redistribution. 1000 — independent agent decisions generated for every combination of role, prompt type and temperature Quotation withheld because this record is not cleared for redistribution. 10 — real experimental data points used as in-prompt examples for few-shot and chain-of-thought prompting Quotation withheld because this record is not cleared for redistribution. KS = 0.479 and AD = 128 — smallest divergence from the human proposer distribution, obtained by the chain-of-thought condition Quotation withheld because this record is not cleared for redistribution. AD = 15.5 — chain-of-thought responder distribution's divergence from the theoretical distribution, the smallest of any condition Quotation withheld because this record is not cleared for redistribution. AD = 20.8 ... AD = 43.7 — divergence from human responder data for few-shot (third best) and for chain-of-thought (worst of the four conditions) Quotation withheld because this record is not cleared for redistribution. approximately 0.7 / approximately 0.4 or less — human acceptance rate at an offer of 50 versus the simulated agents' acceptance rate at the same offer Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1007/s10726-025-09946-9.