GenSIM: LLM-Driven Social Simulation

Updated 9 July 2026

GenSIM is a general-purpose social simulation platform that uses large language model agents to enable large-scale, error-correctable, and reusable multi-agent scenarios.
It abstracts core functions such as agent creation, interaction scheduling, and environment configuration, allowing scenario design through configuration rather than from-scratch implementation.
The platform supports up to 100,000 agents with distributed computing and incorporates robust feedback loops via PPO and SFT to improve long-term simulation reliability.

to=arxiv_search.search 天天大奖彩票站=recent arguments={"query":"(Tang et al., 2024) GenSim A General Social Simulation Platform with LLM based Agents"}0 0 to=arxiv_search.search เงินไทยฟรี 天天中彩票被arguments={"query":"(Wang et al., 2023) GenSim Generating Robotic Simulation Tasks via LLMs"}0 0 to=arxiv_search.search գործում 娱乐开号arguments={"query":"(Chen et al., 2024) On the Evaluation of Generative Robotic Simulations GenSim"}0 0 to=arxiv_search.search 菲娱 дәриҗarguments={"query":"(Li et al., 26 Sep 2025) Generalized Multi-agent Social Simulation Framework GenSIM"}0 0 GenSIM is a general-purpose social simulation platform built around LLM agents. It is defined by three stated design goals: abstracting a set of general functions for constructing customized social scenarios, supporting one hundred thousand agents for large-scale population simulation, and incorporating error-correction mechanisms for more reliable long-term simulation. In the paper that introduces it, GenSIM is positioned as an initial step toward a general, large-scale, and correctable LLM-agent social simulation platform intended to advance computational social science and related simulation research (Tang et al., 2024).

1. Motivation and problem setting

GenSIM was proposed in response to two limitations identified in prior LLM-based social simulations. The first is scenario specificity at small scale: many prior systems focus on narrowly defined tasks or small populations, which constrains external validity, reproducibility, and realistic population modeling. The second is the lack of adaptability and error recovery in multi-round simulations: if unreasonable actions accumulate and propagate over time, simulation quality degrades. GenSIM addresses both by combining a general programming framework with explicit correction loops based on feedback and fine-tuning (Tang et al., 2024).

The platform’s stated motivation is methodological rather than merely engineering-oriented. A simulation framework that requires each scenario to be rebuilt from scratch impedes cross-domain reuse; similarly, a framework that only functions at very small population sizes forces researchers to subsample large populations and thereby increases variance. GenSIM therefore treats generality, scale, and correctability as co-equal requirements. This suggests a view of social simulation in which scenario construction, population size, and longitudinal reliability are not separate concerns but coupled design constraints.

The paper explicitly contrasts this design with prior work exemplified by small-scale or domain-specific systems, including “Generative Agents” with 25 agents and specialized platforms such as RecAgent and EconAgent. GenSIM’s claim is not simply to add more agents, but to expose reusable abstractions that reduce repeated implementation effort across domains while preserving the possibility of counterfactual interventions and long-horizon updates (Tang et al., 2024).

2. Core architecture and functional abstraction

GenSIM organizes simulation around three modules: single-agent construction, multi-agent interaction scheduling, and environment setup. These modules are intended to capture the essential operations shared by many social scenarios, so that a new simulation can be assembled primarily by configuration rather than by re-implementing core logic (Tang et al., 2024).

At the single-agent level, the platform defines agents through profile configuration, memory composition, and action prompting. Profiles include public attributes such as name, gender, and birthplace, together with private attributes such as income and health. Memory is decomposed into short-term memory, long-term memory, and reflection. Action generation is LLM-driven and conditions on the agent’s profile, memories, environment signals, and interaction history. This architecture gives the agent a persistent state rather than reducing it to a stateless prompt-response process (Tang et al., 2024).

At the multi-agent level, GenSIM offers two interaction-generation modes. In script mode, a single LLM call produces an entire multi-party interaction from a third-person, meta-agent perspective. In agent mode, multiple agents generate turns from a first-person perspective, each call accessing the full interaction history. Script mode is described as efficient for one-shot interaction generation, whereas agent mode preserves first-person turn structure at the cost of multiple LLM calls (Tang et al., 2024).

At the environment level, GenSIM represents exogenous systems and global controls. The environment can encode recommendation algorithms, market-matching rules, or other scenario-specific systems, and it also exposes utilities such as interview_agent(agent_id), search_agents(query), store_agent(agent), and global_intervene(policy_change). These functions make the environment an active simulation component rather than a passive container (Tang et al., 2024).

The paper’s conceptual workflow proceeds in five stages: initialize the scenario; create agents; execute a round through scheduled tasks or interactions; optionally run error correction by collecting feedback on action pairs; and continue with an updated backbone model. The abstraction is explicitly designed to prevent “reinventing the wheel,” so that scenario design becomes a matter of choosing profile structures, memory modules, prompting strategies, and environment functions rather than rebuilding the simulation stack itself (Tang et al., 2024).

3. Population scale, orchestration, and statistical stability

A central claim of GenSIM is support for up to 100,000 agents. The implementation relies on distributed parallel computing across GPUs and CPUs, with the reported evaluation performed on a server containing a 192-core CPU, eight A100-40G GPUs, and 440 GB memory. On that hardware, one simulation round with 100,000 agents took 15,492 seconds in the job market scenario and 3,024 seconds in the recommendation scenario. The paper further reports that adding GPUs decreases runtime per round, with monotonic speed improvements in the reported scaling experiment (Tang et al., 2024).

The scaling argument is not limited to throughput. GenSIM also links larger simulated populations to lower output variance. In a preliminary experiment on MovieLens-32M, the paper studies simulated rating distributions under sample sizes of 3.2K, 32K, 320K, and 3.2M user-item pairs, repeating each simulation 10 times. For rating domain $R = [0.5, 1.0, \ldots, 5.0]$ , if $p_i(r)$ denotes the empirical probability of rating $r$ in repetition $i$ , per-rating variance is defined as

$v(r) = \sigma\big(p_1(r), p_2(r), \dots, p_{10}(r)\big).$

The fluctuation metric is the sum of $v(r)$ over $r \in R$ , and the reported finding is that this fluctuation decreases as sample size grows (Tang et al., 2024).

The stated implication is that small-agent simulations can be unreliable because aggregate outcomes fluctuate substantially across runs, whereas large-scale simulations are more reproducible. This claim matters methodologically: GenSIM’s scaling is presented not only as a systems result but also as a response to statistical instability in small-population simulations.

4. Error correction, feedback loops, and long-horizon reliability

GenSIM formalizes each agent action as a pair $(q, a)$ , where $q$ is the action-driving prompt and $a$ is the produced action. Feedback is then introduced in two forms. A scalar score $p_i(r)$ 0 for $p_i(r)$ 1 yields triplets $p_i(r)$ 2 used for PPO fine-tuning, while a revised action $p_i(r)$ 3 yields pairs $p_i(r)$ 4 used for supervised fine-tuning (SFT). The platform accepts both LLM-based feedback, exemplified by GPT-4o, and human-in-the-loop feedback through interfaces for scoring or revision (Tang et al., 2024).

This mechanism is meant to address specific failure modes: unreasonable actions, drift from plausible human behavior, and cumulative error across rounds. LLM-based feedback is described as efficient and requiring no human effort, though potentially biased. Human-based feedback is described as more faithful to real-world judgments but labor-intensive. GenSIM therefore does not assume a single correction oracle; instead, it provides a shared abstraction for feedback collection and model updating (Tang et al., 2024).

In the job market scenario, the paper reports that both PPO and SFT improve performance relative to no correction in single-round settings, and that SFT tends to outperform PPO. The paper’s intuition is that revised actions $p_i(r)$ 5 provide more informative guidance than scalar scores $p_i(r)$ 6. In multi-round settings, the reported pattern is stronger: without correction, performance degrades over rounds, whereas with PPO or SFT, performance improves and continues to increase with successive rounds (Tang et al., 2024).

A plausible implication is that GenSIM treats simulation drift as a trainable systems problem rather than merely an evaluation artifact. By feeding action-level judgments back into the backbone model, the platform attempts to convert labeling effort into improved future rollouts, thereby making correction endogenous to the simulation loop.

5. Scenarios, interface, and operational workflow

GenSIM includes three default scenarios as templates: job market, recommender system, and group discussion. These are presented as configurable exemplars rather than fixed benchmarks, and they illustrate how the same abstractions can be adapted across labor-market negotiation, recommendation dynamics, and discourse settings (Tang et al., 2024).

Scenario	Main elements	Demonstrated use
Job market	skills, desired salary, market matching rules	interviews, negotiations, error correction
Recommender system	user preferences, watched/rated history, recommendation algorithm	scalability timing, cohort search, interventions
Group discussion	multi-agent discourse	script mode and agent mode

In the job market scenario, agent profiles include skills, desired salary, and role types; the environment contains market matching rules; and agent mode can be used for interviews and negotiations. In the recommender system scenario, agents maintain preferences and memories of watched or rated items, while the environment contains the recommendation algorithm itself. The platform also supports search_agents for cohort selection and global_intervene for counterfactual modification of recommendation parameters (Tang et al., 2024).

The paper describes a graphical interface with four major areas: Get Started, Configuration, Simulation view, and Labeling. Through this interface, researchers can choose a scenario, define profiles and memory types, set the number of agents and LLM parameters, inspect profiles and behaviors, search across agents, intervene globally, stop after several rounds for labeling, fine-tune the backbone, and then resume simulation (Tang et al., 2024).

The practical guidance given in the paper recommends starting from a default scenario template, scaling agent count when output variance is high, preferring agent mode for first-person fidelity and script mode for efficiency, and budgeting time for feedback cycles when long-term reliability matters. Code and documentation are provided through the project repository at https://github.com/TangJiakai/GenSim (Tang et al., 2024).

The name GenSim is not unique in the recent literature, and the social simulation platform introduced in (Tang et al., 2024) is distinct from several other systems with closely related names. “GenSim: Generating Robotic Simulation Tasks via LLMs” concerns LLM-generated robotic tasks and demonstrations rather than social simulation (Wang et al., 2023). “On the Evaluation of Generative Robotic Simulations” studies quality, diversity, and generalization metrics for generative robotic simulation pipelines and treats GenSim as one evaluated pipeline among others (Chen et al., 2024). “Generalized Multi-agent Social Simulation Framework” describes a later modular social simulation framework with a memory summarization mechanism, but it is a separate system with a different architecture and evaluation setup (Li et al., 26 Sep 2025).

The term is also distinct from “G-Sim: Generative Simulations with LLMs and Gradient-Free Calibration,” which focuses on hybrid simulator construction with likelihood-free and gradient-free calibration for domains such as healthcare and logistics (Holt et al., 10 Jun 2025); from “GENESIM,” a genetic algorithm for converting an ensemble of decision trees into a single interpretable tree (Vandewiele et al., 2016); and from several papers that use the Gensim Python library for document embedding or topic modelling rather than referring to an agent-based simulation platform (Patil et al., 2023, Igbomezie et al., 2024, Saleh et al., 1 May 2025). A later climate-modeling paper also uses GenSIM for a generative sea-ice surrogate model, again in an unrelated sense (Finn et al., 20 Aug 2025).

Within its own scope, the social-simulation GenSIM remains explicitly limited. The paper notes LLM feedback bias, the labor cost of human correction, and the computational intensity of scaling to 100,000 agents. It further describes the platform as an initial version and points to future work on more advanced acceleration strategies and adaptive self-correction (Tang et al., 2024).

The paper also acknowledges ethical concerns at a high level. Although detailed safeguards are not specified, the design is said to recognize the need to align agent behaviors with real human norms and to manage bias, with careful evaluation and feedback recommended to avoid amplifying harmful patterns. This suggests that GenSIM is best understood not as a finished normative model of society, but as an extensible research infrastructure for controlled, correctable, and large-scale LLM-agent simulation (Tang et al., 2024).