Paper deep dive
Empowering Economic Simulation Through Situation-Aware Llm-Driven Generative System
Zhimei Chen, Mu Chen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 88%
Last extracted: 7/9/2026, 6:04:01 AM
Summary
The paper introduces SAMAS (Situation-aware Multi-agent System), a generative economic simulation framework that leverages LLM-driven role-playing agents. It addresses limitations of traditional top-down modeling and existing ABM/RL approaches by integrating long-term macroeconomic awareness (from LLM parameters) with short-term situational awareness (from individual agent trajectories). SAMAS captures both macro-level structural patterns and micro-level dynamic behaviors, achieving superior performance in volatility realism and turning-point prediction compared to rule-based, RL-augmented, and simple LLM-driven baselines.
Entities (12)
Relation Signals (10)
SAMAS → evaluates → Volatility Realism
confidence 95% · Volatility Realism measures the similarity between simulated statistical properties of macroeconomic variables and real-world data.
SAMAS → evaluates → Turning-Point Hit Rate
confidence 95% · Turning-Point Hit Rate represents the model’s ability to capture the timing of economic cycles.
SAMAS → uses → Large Language Models
confidence 95% · SAMAS leverages LLM-driven role-playing agents for perception and decision-making.
SAMAS → integrates → Long-term Awareness
confidence 90% · Long-term Awareness builds upon the historical understanding embedded in LLMs.
SAMAS → integrates → Short-term Awareness
confidence 90% · Short-term Awareness derives from individual-specific trajectories within the recent context window.
SAMAS → outperforms → Rule-based Systems
confidence 90% · SAMAS outperforms all baseline methods in both Volatility Realism and Turning-Point Hit Rate.
SAMAS → outperforms → RL-augmented Systems
confidence 90% · Although the reinforcement learning approach brings certain improvements... it still falls behind the LLM-driven approach.
Large Language Models → provides → Long-term Awareness
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Traditional economic modeling typically follows a TOP-DOWN paradigm, neglecting individual diversity and the complexity of social interactions. To better capture the complexity of societal structure, Agent-Based Modeling (ABM) employs a BOTTOM-UP solution by incorporating micro-level dynamics to generate macroeconomic phenomena. Reinforcement Learning further improves its decision-making ability through tailored reward signals. However, existing ABM systems struggle to generalize beyond predefined scenarios. Recognizing the potential of LLM-driven role-playing in perception and human-like decision-making, we propose SAMAS, which models individual agents with rich macroeconomic understanding embedded in LLMs and economic trajectories experienced in the passing simulation steps. By jointly modeling both macro-level structural patterns and micro-level dynamic behaviors, SAMAS achieves superior performance in volatility realism and turning point prediction.
Tags
Links
- Source: https://arxiv.org/abs/2606.20720v1
- Canonical: https://arxiv.org/abs/2606.20720v1
Trouble viewing inline? Open PDF directly →
Full Text
25,722 characters extracted from source content.
Expand or collapse full text
EMPOWERING ECONOMIC SIMULATION THROUGH SITUATION-AWARE LLM-DRIVEN GENERATIVE SYSTEM Zhimei Chen 1 Mu Chen 2∗ 1 School of Economics, Southwest Minzu University 2 ReLER, AAII, University of Technology Sydney ABSTRACT Traditional economic modeling typically follows a TOP- DOWN paradigm, neglecting individual diversity and the complexity of social interactions. To better capture the com- plexity of societal structure, Agent-Based Modeling (ABM) employs a BOTTOM-UP solution by incorporating micro-level dynamics to generate macroeconomic phenomena.Rein- forcement Learning further improves its decision-making ability through tailored reward signals. However, existing ABM systems struggle to generalize beyond predefined sce- narios. Recognizing the potential of LLM-driven role-playing in perception and human-like decision-making, we propose SAMAS, which models individual agents with rich macroe- conomic understanding embedded in LLMs and economic trajectories experienced in the passing simulation steps. By jointly modeling both macro-level structural patterns and micro-level dynamic behaviors, SAMAS achieves superior performance in volatility realism and turning point prediction. Index Terms— LLMs, Multi-agent Systems, Agent- based Modeling, Social Simulation, Economic Simulation 1. INTRODUCTION Classical economic modeling normally adopts a TOP-DOWN paradigm, based on either a theory-driven [1], or a data-driven method [2]. These TOP-DOWN solutions have served as the foundation of macroeconomic analysis and policy formu- lation. While effective in capturing empirical correlations among macroeconomic variables, the passive introduction of the “rational expectation assumption” [3] neglects the diversity of individuals in society and the complexity of so- cial interaction, leading to an unrealistic social structure. Agent-Based Modeling (ABM) [4, 5] addresses this by simu- lating individual agent behaviors and the interactions among them. By considering micro-level dynamics, ABM gener- ates macroeconomic phenomena in a BOTTOM-UP manner. Reinforcement learning [6] further extends this direction by enabling agents to learn adaptive strategies from interac- Corresponding author: mu.chen@alumni.uts.edu.au NO Situational Awareness (a) Rlue-based System (c) SaMAS Limited Situational Awareness (b) Learning-based System Rich Situational Awareness Since Last Month Unemployment is rising Economic recession appears Agent Group The Great Depression of 1929 The Oil Crisis of the 1970s The Subprime Mortgage Crisis of 2008 The Great Depression of 1929 The Oil Crisis of the 1970s The Subprime Mortgage Crisis of 2008 Agent 1 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Agent 1 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Long-term SA Short-term SA NO Situational Awareness (a) Rlue-based System (c) SaMAS Limited Situational Awareness (b) Learning-based System Rich Situational Awareness Since Last Month Unemployment is rising Economic recession appears Agent Group The Great Depression of 1929 The Oil Crisis of the 1970s The Subprime Mortgage Crisis of 2008 The Great Depression of 1929 The Oil Crisis of the 1970s The Subprime Mortgage Crisis of 2008 Agent 1 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Agent 1 Blue-collar White-collar Business Owner Reduce Spending Keep Essentials Immediate Spending Cut Keep Spending Agent 2 Agent 3 Long-term SA Short-term SA Fig. 1: (a) Existing economic simulation systems adopt a rule-based paradigm, exhibiting no situational awareness. (b) Learning-based simulation systems learn from historical data and acquire limited situational awareness by understanding various economic contexts. (c)SAMAS explicitly leverages underlying macroeconomic knowl- edge stored in LLMs and stores micro situations experienced by each simulation agent, leading to rich situational awareness. tion with the environment via reward signals, demonstrating strong performance in complex social simulation settings. Despite the impressive decision-making capabilities brou- ght by RL-based ABM, they adopt TAILORED reward func- tions to mimic diverse human behavior. As a result, they fall short in generalizing beyond pre-defined scenarios. Ac- tually, humans rely on past experiences to guide current deci- sions. Specifically, the perception of economic environments shapes humans’ risk preferences [7]. For instance, as shown in Fig 1, awareness of historical crises such as the Great De- pression of 1929, the Oil Crisis of the 1970s, and the Sub- prime Mortgage Crisis of 2008 tends to trigger panic when similar economic signals emerge. On the other hand, indi- viduals are easily affected by recent successful financial deci- sions [8, 9], termed RECENCY BIAS. Furthermore, individu- als gradually develop similar economic decision patterns [10], forming group decisions. Specifically, the economic activities of different individuals are highly correlated to their profes- sions, as they share:❶ similar income and consumption struc- tures;❷ homogeneous information propagates among them in arXiv:2606.20720v1 [cs.MA] 16 Jun 2026 Work 0.83 SA 0.12 Save 2.2k Work 0.98 SA 0.03 Save 4.4k Work 0.53 SA 0.23 Save 1.5k Short-term SA Individual Trajectory Interaction Agent & Env. Init. Start Simulation cycle t = 1 Interest Rate at 3%, Inflation Appears Blue collar White collar Business Owner Alice 35ys Income 12.5k Tax 3.0k Bella 42ys Income 15.0k Tax 4.5k Tom 48ys Income 7.9k Tax 1.4k Prompt Encode 2nd Simulation cycle t = 2 Individual Trajectory Interaction Decode Group Opinion Agent Blue-collar Group Agent Agent E Work 0.83 SA 0.12 Save 2.2k Work 0.98 SA 0.03 Save 4.4k Work 0.53 SA 0.23 Save 1.5k Short-term SA Individual Trajectory Interaction Agent & Env. Init. Start Simulation cycle t = 1 Interest Rate at 3%, Inflation Appears Blue collar White collar Business Owner Alice 35ys Income 12.5k Tax 3.0k Bella 42ys Income 15.0k Tax 4.5k Tom 48ys Income 7.9k Tax 1.4k Prompt Encode 2nd Simulation cycle t = 2 Individual Trajectory Interaction Decode Group Opinion Agent Blue-collar Group Agent Agent E Fig. 2: SAMAS equips with long-term and short-term awareness, the former is from the commonsense awareness in LLMs’ parameters, and the latter is from stored individual trajectories of agents in the system. their daily interactions [11]. For example, white-collar pro- fessionals tend to reduce unnecessary spending but maintain essential consumption during recessions to mitigate unem- ployment risks. In contrast, business owners keep spending since they typically hold larger savings. Blue-collar work- ers, however, will immediately cut back on spending due to the lack of savings. Current ABM solutions could not cap- ture these intrinsic microeconomic patterns, which limits their ability to reproduce group-level decision dynamics. The generative nature of Large Language Models (LLMs) has recently made them attractive as role-playing agents, as they can satisfy the core ABM requirement of equipping individual agents with perception and decision-making ca- pabilities. The macroeconomic understanding encoded in an LLM’s parameters provides a rich background context for an individual agent’s long-term awareness, such as prior understanding of financial crises.Nevertheless, humans’ economic decisions not only rely on long-term historical knowledge but also on their recent experiences [12]. Ef- fective decision-making thus requires dynamically balanc- ing long-term macroeconomic understanding and short-term contextual observations. With such insight, we propose a Situation-aware Multi-agent System (SAMAS) for realistic economic simulation.Within the system, two paradigms of economic awareness predominate: Long-term Awareness (LA) and Short-term Awareness (SA). The former builds upon the historical understanding embedded in LLMs, while the latter derives from individual-specific trajectories within the recent context window.The two paradigms provide complementary perspectives, capturing macro-level social consensus and micro-level adaptive behaviors, respectively. SAMAS demonstrates superior Volatility Realism and Hit Rate compared to existing methods. 2. RELATED WORK Traditional Economic Modeling primarily relies on statis- tical and equilibrium-based frameworks to analyze aggregate behaviors and macroeconomic patterns [1, 2], including rep- resentative Dynamic Stochastic General Equilibrium (DSGE) and Vector Autoregression (VAR), serving as representative examples that are widely applied in policymaking. A key limitation of these approaches, however, is their reliance on stylized assumptions such as agent homogeneity and perfect rationality, which substantially oversimplify the complexity of real-world economies. In particular, such assumptions ne- glect individual heterogeneity, bounded rationality, and the social network structures that shape economic outcomes. Agent-Based Modeling explains macro-level outcomes through micro-level interactions [4, 5]. ABM simulates the behaviors of heterogeneous agents and their localized inter- actions, thereby partially addressing the limitations of tradi- tional macroeconomic modeling, which often neglects indi- vidual heterogeneity and complex interaction structures [13]. Nevertheless, simplified ABM frameworks are deficient in situational awareness, which limits their ability to capture the subtle and context-dependent characteristics of human economic decision-making. To mitigate this, recent research has increasingly explored the integration of RL [14] and DL [15, 16, 17, 18, 19] into ABM frameworks to enhance intelligence. In particular, RL empowers agents to iteratively learn strategies through interaction with their environment. Situational Awareness is also correlated with the proposed system. The emergence of large generative models has re- vealed unprecedented potential for enabling agents to per- ceive, reason, and make decisions. Contextual awareness plays a pivotal role in these processes, as it directly informs perception and guides decision-making. For socially situ- ated agents, memory functions not only as a repository of historical information but also as the foundation for cogni- tive consistency, situational understanding, and the develop- ment of long-term strategies. Existing approaches, however, largely rely on either external memory augmentation, such as retrieval-augmented generation (RAG) [20], or reinforcement learning-based mechanisms to control memory access [21]. While effective in short-term interaction tasks, these methods remain inadequate for macroeconomic simulation, which is inherently characterized by long-term temporal evolution. In particular, current methods fall short in modeling multi- timescale dynamics and the complex interaction structures spanning environments, individuals, and groups. 3. SAMAS SAMAS aims to construct a more realistic social environ- ment by leveraging the generative and reasoning capacities of LLMs in role-playing agents. It addresses the simulation pipeline at two complementary levels: (micro) the decision- making processes of individual agents, and (macro) the emer- gent dynamics arising from their collective interactions. 3.1. Overview As shown in Fig 2, consistent with the principles of ABM, the overall economic environment is generated through the interactions of individual role-playing agents, and the overall simulation performance depends on each agent’s output and the summary results among them. The quality of the simula- tion, in turn, depends on both the behavior of each agent and the aggregated outcomes of their interactions. Specifically, SAMAS constructs an economic simulation environment with multiple simulation steps. At each step t = [1, 2, . . . , T], the system comprises N independent agents [Agent p 1 , Agent p 2 , Agent p 3 , ...] as human-like decision- makers: p i =name i , age i , occupation i , income i (1) Each agent is initialized as a representative participant in economic activities with a real social background, to cap- ture the heterogeneity of real-world individuals. To this end, each agent is characterized by a fixed profile that remains un- changed throughout the entire simulation cycle to capture its attributes as an economic participant. Concurrently macroe- conomic environment E (e.g., the first day after the interest rate cut, inflation appears based on macroeconomic indica- tors.) is initialized, with global parameters set to reflect the macroeconomic conditions. In the current simulation cycle, each agent generates a decision (e.g., consumption willing- ness, labor willingness) based on its profile p and the informa- tion received from the environment E t , in accordance with its assigned role. The decisions and interactions of all agents are then aggregated to constitute the macroeconomic context for that period. Subsequently, these updated indicators are used to refresh the global environment to E t+1 , thereby establish- ing a new context for the next iteration. After multiple cy- cles, both individual-level and macro-level data are collected to evaluate the overall behavior of the system. 3.2. Situational Awareness Rich macroeconomic understanding embedded within the network parameters of LLMs provides agents in the simu- lation system with shared commonsense awareness. This awareness, when applied to economic behavior, manifests as risk perception of the environment. For example, historical events such as the “1929 Great Depression” enable individ- ual agents to maintain a long-term shared awareness of the economic environment, driving them to make conservative economic decisions when similar signals reappear. However, in practice, humans not only rely on distant his- torical patterns but also incorporate their own financial sta- tus and the outcomes of their recent economic actions [8, 9], a phenomenon commonly referred to as RECENCY BIAS. SAMAS incorporates this mechanism by summarizing each agent’s past experiences and, based on key factors in its pro- file that influence economic activities, integrating them with current macroeconomic data to generate a risk perception pa- rameter, which represents Short-term Situational Aware- ness. This parameter captures an agent’s holistic interpre- tation of historical analogies, individual circumstances, and macroeconomic signals, serving as a dynamic measure of its risk sensitivity and adaptive behavioral responses. At the end of each simulation cycle, SAMAS collects critical outputs from micro-level activities and encodes them into a JSON file. These statistics are then used to dynamically update the macroeconomic environment. Inspired by the characteristics of human situational memory, after encoding each agent’s historical trajectory, we apply the Ebbinghaus [22] forget- ting curve to gradually attenuate the weights of past memory, thereby ensuring the realism of economic behavior modeling. Moreover, given that individuals in society typically re- ceive information with varying degrees of bias, different groups exhibit distinct distributions of risk preferences and economic decisions [10, 11]. To address this, SAMAS tracks group structures: by aggregating the past experiences of simi- lar groups, SAMAS guides the local situational awareness of each group, ensuring that both individual and collective risk perceptions evolve realistically within the simulation. 4. EXPERIMENT We conduct simulations to evaluate the effectiveness of SAMAS and answer the following questions: (1) Does SAMAS achieve higher simulation realism compared to traditional methods and simple LLM-driven approaches? (in Table 1) (2) Is the effectiveness of SAMAS influenced by the choice of underlying LLMs? (in Table 2)(3) Does the scale of agents affect the performance of the simulation? (in Table 3) 4.1. Experimental Setting In simulation, all agents operate in an economic environment updated by periodic contextual prompts, with simulation time steps defined daily, monthly, or yearly. The agent number is set to 100 and agents are implemented with open-source LLM APIs. We adopt the following settings as baseline: Table 1: Comparison with existing solutions on two selected economic simulation metrics: Volatility Realism ↑ , Turning-Point Hit Rate ↑ (daily/monthly/yearly), and Token usage ↓ (M). Higher is better ↑ ; lower is better ↓ . Algorithm ComponentVolatility Realism (%) ↑ HR-d (%) ↑ HR-m (%) ↑ HR-y (%) ↑ Token usage (M) ↓ RANDOM0.2%0.40.30.5- + Simple Rule-based [23] 51.8%26.841.546.0- + Complex Rule-based [24]60.5%29.544.249.3- + RL-augmented [25]71.2%32.049.855.4- + Simple LLM-driven [26] 75.8%35.654.760.1141.3/15.5/1.5 SAMAS81.5%42.564.870.9167.9/20.2/2.1 Table 2: Sensitivity Analysis on various LLMs. ModelVR (%) ↑ HR (d/m/y, %) ↑ DeepSeekv3 [27]82.345.1/66.0/75.1 Grok-3 [28] 82.543.3/65.3/71.7 GPT-3.5 [29]81.542.5/64.8/70.9 Table 3: Analysis on different agent number. Agent numberVR (%) ↑ HR (d/m/y, %) ↑ 5077.540.4/61.7/67.4 100 81.542.5/64.8/70.9 20081.742.6/64.9/71.0 50082.142.8/65.2/71.4 100082.543.2/65.6/72.1 • Randam: Agent actions are randomly selected, with no reasoning involved and in violation of economic principles. • Simple& Complex Rule-based: Economic actions are selected either uniformly from a small set of predefined rules or from a richer set incorporating conditional logic and context-specific constraints. • RL-based: Agents employ RL to adaptively optimize their policies. By interacting with the environment, they itera- tively improve decision-making strategies. • Simple LLM-driven: LLM-based Agents with decisions guided by textual instructions. Evaluation Metric. We systematically evaluate results of an economic simulation system from two perspectives: • Volatility Realism measures the similarity between sim- ulated statistical properties of macroeconomic variables (e.g., GDP, CPI) and real-world real data. • Turning-Point Hit Rate: represents the model’s ability to capture the timing of economic cycles, particularly peaks (booms) and troughs (recessions). 4.2. Comparison with Other Methods We first compare SAMAS with traditional rule-based meth- ods and reinforcement learning approaches. As shown in Table 1, With random rules, the simulation results are close to zero and fail to capture the real patterns of economic ac- tivity. SAMAS outperforms all baseline methods in both Volatility Realism (VR) and Turning-Point Hit Rate (HR) across daily, monthly, and yearly levels. Conventional ran- dom or rule-driven baselines show limited performance in volatility realism and hit rate, indicating their difficulty in capturing the complex dynamics of economic behaviors. Al- though the reinforcement learning approach brings certain improvements through policy-making advantages, it still falls behind the LLM-driven approach, which inherently benefits from Agent-Based Modeling (ABM). Furthermore, SAMAS incorporates multi-level situational awareness. While this increases token consumption, it achieves a volatility realism of 81.5% and an annual hit rate exceeding 70%. To examine the sensitivity of SAMAS to the choice of different LLMs, we test three variants: DeepSeek and Grok. The results in Table 2 indicate that different LLMs achieve comparable performance in economic simulation. Finally, we evaluate the impact of the number of agents on simulation realism (Table 3). Starting from 50 agents, perfor- mance improves steadily as the number of agents increases, demonstrating the advantage brought by group awareness. 5. CONCLUSION SAMAS investigates the advantages of LLM-based role- playing agent systems in economic simulation by integrating complementary paradigms of long-term and short-term sit- uational awareness. By endowing agents with human-like perceptual and reasoning capabilities, SAMAS facilitates rich micro-level interactions, which in turn give rise to real- istic macroeconomic phenomena. Our findings indicate that: ❶ LLM-driven ABM exhibits strong potential for simulat- ing a wide range of economic environments under diverse conditions;❷ Collective social behavior within groups sig- nificantly shapes individual agent decisions. These insights provide valuable guidance for the design of next-generation social simulation systems, underscoring the importance of multi-scale situational awareness as a prerequisite for achiev- ing both behavioral plausibility and macro-level realism. 6. REFERENCES [1] Frank Smets and Rafael Wouters, “Shocks and frictions in us business cycles: A bayesian dsge approach,” Amer- ican economic review, vol. 97, p. 586–606, 2007. [2] Christopher A Sims, “Macroeconomics and reality,” Econometrica: journal of the Econometric Society, vol. 48, p. 1–48, 1980. [3] John F Muth, “Rational expectations and the theory of price movements,” Econometrica: journal of the Econo- metric Society, p. 315–335, 1961. [4] Leigh Tesfatsion,“Agent-based computational eco- nomics: Growing economies from the bottom up,” Arti- ficial life, vol. 8, p. 55–82, 2002. [5] J Doyne Farmer and Duncan Foley,“The economy needs agent-based modelling,” Nature, vol. 460, p. 685–686, 2009. [6] TabishRashid,MikayelSamvelyan,Chris- tian Schroeder De Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson,“Monotonic value function factorisation for deep multi-agent reinforce- ment learning,” Journal of Machine Learning Research, vol. 21, p. 1–51, 2020. [7] Hannah Schildberg-H ̈ orisch, “Are risk preferences sta- ble?,” Journal of Economic Perspectives, vol. 32, p. 135–154, 2018. [8] Nusret Cakici and Adam Zaremba, “Recency bias and the cross-section of international stock returns,” Jour- nal of International Financial Markets, Institutions and Money, vol. 84, p. 101738, 2023. [9] Vladimir Kotomin and Abhishek Varma, “Debiasing re- cency: Evidence from individual investor stock sales,” Journal of Behavioral Finance, p. 1–17, 2025. [10] Tetiana Hill, Petko Kusev, and Paul Van Schaik, “Choice under risk: How occupation influences prefer- ences,” Frontiers in psychology, vol. 10, p. 2003, 2019. [11] Cass R Sunstein, Echo chambers: Bush v. Gore, im- peachment, and beyond,Princeton University Press Princeton, NJ, 2001. [12] Chupei Wang and Jiaqiu Vince Sun, “Unable to forget: Proactive lnterference reveals working memory limits in llms beyond context length,” in International Confer- ence on Machine Learning (Workshop), 2025. [13] Eric Bonabeau, “Agent-based modeling: Methods and techniques for simulating human systems,” Proceedings of the national academy of sciences, vol. 99, p. 7280– 7287, 2002. [14] Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang,“Seed-grpo: Semantic entropy enhanced grpo for uncertainty-aware policy optimization,” arXiv preprint arXiv:2505.12346, 2025. [15] Mu Chen, Zhedong Zheng, Yi Yang, and Tat-Seng Chua,“Pipa: Pixel-and patch-wise self-supervised learning for domain adaptative semantic segmentation,” in ACM International Conference on Multimedia, 2023. [16] Mu Chen, Zhedong Zheng, and Yi Yang, “Transfer- ring to real-world layouts: A depth-aware framework for scene adaptation,” in ACM International Conference on Multimedia, 2024. [17] Mu Chen, Liulei Li, Wenguan Wang, Ruijie Quan, and Yi Yang, “General and task-oriented video segmenta- tion,” in European Conference on Computer Vision, 2024. [18] Mu Chen, Liulei Li, Wenguan Wang, and Yi Yang, “Dif- fvsgg: Diffusion-driven online video scene graph gener- ation,” in Proceedings of the Computer Vision and Pat- tern Recognition Conference, 2025. [19] Mu Chen, Zhedong Zheng, and Yi Yang,“Pipa++: towards unification of domain adaptive semantic seg- mentation via self-supervised learning,” arXiv preprint arXiv:2407.17101, 2024. [20] Patrick Lewis, Ethan Perez, et al., “Retrieval-augmented generation for knowledge-intensive nlp tasks,” in Ad- vances in neural information processing systems, 2020. [21] Wujiang Xu, Kai Mei, Hang Gao, Juntao Tan, Zujie Liang, and Yongfeng Zhang, “A-mem: Agentic mem- ory for llm agents,” in Advances in Neural Information Processing Systems, 2025. [22] Hermann Ebbinghaus, ̈ Uber das ged ̈ achtnis: unter- suchungen zur experimentellen psychologie, Duncker & Humblot, 1885. [23] Joshua M Epstein and Robert Axtell, Growing artificial societies: social science from the bottom up, Brookings Institution Press, 1996. [24] Leigh Tesfatsion and Kenneth L Judd, Handbook of computational economics: agent-based computational economics, vol. 2, Elsevier, 2006. [25] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson, “Learning to com- municate with deep multi-agent reinforcement learn- ing,” Advances in neural information processing sys- tems, 2016. [26] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao, “React: Synergizing reasoning and acting in language models,” in International Conference on Learning Representa- tions, 2023. [27] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al., “Deepseek- v3 technical report,” arXiv preprint arXiv:2412.19437, 2024. [28] x.ai, “Grok-3 language model,” 2024. [29] OpenAI, “Openai api models: Gpt-3.5,” 2023, De- scribes model versions and usage for GPT-3.5-turbo.