Registration open

SocioVerse Challenge 2026

AI4SS Challenge for Human–AI Collaboration and Social Governance

Awards on Nov 8 · LASS 2026 @ CIKM 2026 · Rome

Register Learn more
Live now · open registration

Social evolution,
visible and testable.

SocioVerse2 is a longitudinal social-simulation framework where researchers and AI agents build studies together. People (P) aligned with real populations live through a changing environment (E) grounded in real-world signals, and their behavior (B) is recorded at every step. Intervene to branch a counterfactual, revise any part of the study, and reproduce every version.

Start a study See it in action ↓ Technical report ↗
Overview film · 90 s

SocioVerse2 intro video

16 ready-made studies in the catalog
11 classic ABMs, rule vs LLM vs hybrid
7 conversation-driven workflow skills
1 simple loop behind every run: P × E → B
Case studies

Already running on SocioVerse2

Five studies from the catalog — urban policy, social media, consumer research and the classics — each with a browsable live-dashboard export.

Urban policy · real census data Path C · zero-edit adapter

Chicago segregation under a transit intervention

Starting from real 2010 census tracts, a CTA subway station opens on the South Side at step 2 with a policy broadcast. LLM residents decide whether to move; dissimilarity and isolation indices are tracked step by step against a no-intervention baseline.

D_black_white Isolation_black census tracts policy sandbox
Live demo →
Opinion dynamics · hybrid engine Path B · built on core

Hybrid social-media opinion dynamics

A hybrid design: a handful of LLM-driven core users post and react in a Twitter-like feed while thousands of bounded-confidence ABM users assimilate. Trigger news lands, and population bias and diversity shift turn by turn.

bias diversity LLM × ABM trigger news
Live demo →
Consumer confidence · macro forecasting Path C · zero-edit legacy adapter

US consumer-confidence backtest on a fixed panel

A fixed, demographically stratified US panel answers the University-of-Michigan consumer-sentiment questions month by month, grounded in real macro news and FRED indicators, with the aggregate scored against the index published afterwards — 1.66 index points of mean absolute error across 11 months.

fixed stratified panel monthly news + indicators UMich questions vs UMCSENT
Live demo →
Auto market · real registration data Path B · built on core

Brand-share dynamics in the German car market

1,500 buyer personas anchored to Germany's real demographic joint distribution choose brands month by month across 2024-01 → 2026-05 — EV-subsidy removal, price moves and BYD's entry included — reproducing the KBA-registered share trajectories of 12 named brands (Tesla roughly halved, BYD up from near zero).

1,500 personas 12 brands · KBA-grounded EV-subsidy shock
Live demo →
Why SocioVerse2

Built for the questions that need time

Most LLM-agent frameworks stop at cross-sectional snapshots. SocioVerse2 tracks the same people through a changing world — and hands you the panel data to prove it.

⏱

Longitudinal by design

Persistent agent IDs turn every run into a panel dataset: one row for each person, at each step. Ask "who changed, when, and why" — not just "what did the crowd look like".

⧉

People × Environment → Behavior

Describe who your people are and the world they live in; SocioVerse2 plays out what they do. Each step, the world updates, everyone observes and decides, and every outcome is recorded.

✦

Intervene, branch, compare

An intervention is one entry in the environment that fires at the step you pick. Branch at that step: earlier steps are replayed, not re-run, and both branches follow the same people.

⑂

A research loop you control

Each version snapshots the whole workspace: change anything at any stage, roll back anytime. Start from a catalog study (Path A), from scratch (Path B), or your own simulator (Path C).

◉

Watch the run, live

A local dashboard shows stage status, streaming step-by-step metrics, and a per-agent inspector — drill into any agent's state, decision and generated output mid-run.

◍

Population and event services

Two standard data services supply real people to sample from and real-world signals (news, indicators, prices) for the environment, each item stamped with its source and version.

Workflow

Five commands from question to report

Every stage writes a validated artifact; every handoff is schema-checked. Re-run any stage and the dashboard flags what's now stale downstream. Two more skills round it out: /sv-build-model for from-scratch (Path B) studies, and /sv-iterate for changing a study after a run: a new version, or a branch with an intervention.

/sv-init route the question, scaffold or fork a study
→
/sv-build-environment compose E: layers, events, schedules
→
/sv-build-population materialize P: personas with persistent IDs
→
/sv-run play out the society step by step, stream the panel live
→
/sv-report metrics, figures and a written report
Contact us

Tell us what you want to simulate

Questions, collaboration ideas, or a quota top-up — tell us who you are and what you'd like to simulate, here or at contact@socioverse.fudan-disc.com. We reply by email.

Registration is instant; an invite code grants a higher starting quota.

Register / Sign in →

We reply by email, usually within a couple of days.