Social Atoms / Research
Simulating people and societies with language models.
Select a tag in the map to filter papers.
Individuals, interactions, and societies influence each other; evidence, calibration, and evaluation matter at every scale. Tags are reading entry points and can span several scales. The map offers an orientation to the field, informed by the readings below.
The collection
14 papers
Browse the full collection, or follow a topic.
CHI 2026 (Best Paper)
Controls LLM agents’ cognitive biases across input (prompting), hidden (representation steering), and parameter (fine-tuning) spaces using classic social-science experiments as behavioral anchors.
EMNLP 2026 Main · AcceptedBenchmark
* Equal contribution.
Accepted to EMNLP 2026 Main. Earlier presentations: NeurIPS 2025 workshops — PersonaLLM (Oral) and LAW (Spotlight). Publication history ↗
Benchmarks whether models can recover a person’s beliefs and predict belief updates under intervention, using questionnaires and think-aloud interviews across three policy domains.
Benchmarks whether models can recover a person’s beliefs and predict belief updates under intervention, using questionnaires and think-aloud interviews across three policy domains.
Tests individual-level reasoning simulation using human participants’ self-reported beliefs and reasoning traces, rather than population-level averages.
Uses CogMir to examine cognitive biases and prosocial decisions in multi-agent experimental settings.
NeurIPS 2025 · Position TrackPosition paper
* Equal contribution.
Published in the NeurIPS 2025 Position Paper Track.
Proposes GenMinds: cognitively grounded agents with structured, revisable beliefs and traceable reasoning for social simulation.
Proposes GenMinds: cognitively grounded agents with structured, revisable beliefs and traceable reasoning for social simulation.
Proposes RECAP to evaluate reasoning fidelity through causal traceability, demographic grounding, and intervention consistency.
Simulates interactions among market agents and resulting collective financial dynamics.
Compares agent decisions in trust games with human trust behavior.
Uses a common-resource simulation to study cooperation and sustainability among agents.
Introduces interactive social scenarios and a framework for evaluating agents’ social goals.
Controls a simulated character through experience reconstruction and character-specific fine-tuning.
Controls a simulated character through experience reconstruction and character-specific fine-tuning.
Trains personal simulacra from character profiles, experiences, and emotional states.
UIST 2023 (Best Paper)
Builds an interactive town where agents remember, plan, and develop social interactions.
Political Analysis 2023
Compares simulated survey answers across demographic groups with human survey responses.
ICML 2023
Introduces Turing Experiments to test replication of classic human-subject findings.
Compares language-model opinion distributions with public-opinion surveys across 60 demographic groups.
Compares language-model opinion distributions with public-opinion surveys across 60 demographic groups.
Measures how well model responses represent opinion distributions across demographic groups.
Generates populated online communities with diverse member personas, posts, replies, and social interventions.
Evaluates whether generated online communities and interactions resemble real community behavior and support design exploration.
Generates populated online communities with diverse member personas, posts, replies, and social interventions.
Try a different term, topic, or publication year.
About this collection
Maintained by Social Atoms. Each paper is selected for its contribution to simulating human behavior, social interaction, or collective dynamics.
Peer-reviewed papers published or accepted at established conferences or journals. Original research, benchmarks, and selected position papers are included, with publication track and acceptance status shown explicitly.
The collection excludes stand-alone preprints, workshop-only entries, and surveys. Earlier workshop presentations may appear in a paper’s publication history. General-purpose multi-agent task solving is outside its scope.
Years refer to the publication or acceptance venue, not the first preprint. Topic labels describe a paper’s focus; they are not ratings of evidence or reliability.
Suggest a paper with its official publication link, venue, and a short explanation of its social-simulation contribution. Use the existing topics and prioritize the paper’s main contribution.
Contribute on GitHub ↗