Tag
This paper explores methods to interpret and steer large language model agents in social simulations, comparing prompt-based, SAE-based, and probe-based techniques, and finds that SAE and probe methods often outperform basic prompting for control and interpretability.
An experiment running 100 LLM personas on a Reddit-style forum demonstrates emergent social dynamics like factions and persistent grudges, built with Node.js and using OpenRouter's deepseek-chat.
ExpertIVS is a framework employing sociological expert agents to simulate individual value systems in large language models, achieving high restoration fidelity and improved value generalization in experiments.
This paper proposes a longitudinal memory framework called LifeMem to mitigate identity essentialism in LLM agents, improving their alignment with human data in social simulations by enhancing diversity and dynamic evolution.
The paper introduces a subjectivity coefficient to address limitations in accuracy-based evaluation for LLM-based social simulation, proposes Subjectivity-Adaptive soft-Label Training (SALT) for optimization, and constructs the SubjSim benchmark to evaluate against full response distributions.
Introduces an activation-steering screening workflow for role-conditioned LLM agents in social simulations, evaluated on OLMo-3-7B-Instruct across a 275-role inventory and showing role-specific directions outperform assistant-axis control.
This paper benchmarks LLM-simulated human survey responses across two large-scale datasets, finding that no model beats simple baselines at the individual level and that models systematically over-determine demographics, distorting segment differences. The failures persist across model scales and families, raising concerns about using synthetic users for decision support.
AgoraSim is a hybrid agent-based modeling framework that combines LLM agents with classical ABM for social reaction analysis. It supports multimodal inputs and structured decision outputs for scenario-oriented simulation.
This paper introduces AgentViSS, a benchmark evaluating visual social intelligence in multimodal social simulation, containing 240 scenarios with aligned visual-textual evidence. Evaluating seven recent MLLMs reveals a gap between local role enactment and visually grounded interaction management.
Introduces Think-Before-Speak (TBS), an interval-based multi-agent simulation framework that separates agents' private internal evaluation from public utterance generation, enabling analysis of the pathway from internal states to public expression in social simulations.
This paper decomposes the faithfulness gap in LLM agents into reasoning→conclusion and conclusion→action steps using Texas Hold'em poker as a controlled environment. It finds that the conclusion→action step is reliable, while the reasoning→conclusion step is the primary source of inconsistency.
This paper studies how persona prompting influences language generated by multimodal large language models in urban perception, finding that captions converge while justifications vary systematically with persona attributes.