Skip to content

Related work

Research context

A resident accepts an invitation. What evidence would show that two residents actually met? The answer depends on who controls the world, what the residents can observe and what the experiment counts.

Who came to the party?

In Smallville, Isabella starts with the idea of a Valentine’s Day party. Over two simulated days, invitations spread through a town of 25 residents. Park and colleagues checked residents’ claims to have heard the news against dialogue stored in their memories. They counted attendance separately. Five of the twelve invitees appeared at the café.

Park et al., UIST 2023 · End-to-end measurements and attendance, §7

The server keeps track of positions and objects and sends agents what lies within visual range. Each agent retrieves memories by relevance, recency and importance, develops reflections and adjusts its plans. The authors also asked human evaluators to judge interview answers with different parts of this architecture disabled. Those judgments test believability. The attendance count tests whether the meeting took place.

Memory, reflection and planning, §4 · Sandbox state, §5 · Controlled evaluation, §6

In Village Zero’s tavern we can ask one more question. Did Nessa get a turn while Mara’s invitation was still open? In R06/A, event 101 opens the invitation. At tick 55, her menu lists three rest options. She can join, decline or rest alone. She chooses to join, and event 105 refers back to the invitation. The same kind of menu record lets us count turns in which joining was possible and another action was chosen.

R06/A decision menus and linked events · Long-Life opportunities and outcomes

Who settles a market order?

In Concordia a Game Master mediates between a player's attempted action and the world. The founding paper describes a character proposing an action in ordinary language. The GM can assess its physical plausibility or call a digital service to carry it out. Player components can combine associative memory retrieval with language-model calls, while the GM determines what each player observes.

Vezhnevets et al., founding paper, v2 · Concordia project

The marketplace implementation shows how much authority can sit in code. A configurable GM component supplies observations, specifies the next action format and resolves orders. In its order book, a matching bid and ask still require stock and sufficient cash. The code moves goods and money between parties and logs the quantity filled, the price and whether the order failed, filled partially or filled completely.

Marketplace GM composition · Order records · Matching and settlement

Scenario designers can therefore choose which consequences need interpretation and which follow explicit rules. In the Concordia Contest, researchers used different partners and mixed-motive scenarios to test whether cooperative strategies generalized. Village Zero’s Long-Life study asks what changes when the same eight starting histories receive different decision schedules. Its engine defines the operations available and validates their effects. Giving a new kind of contract enforceable consequences would require adding a corresponding rule.

Smith et al., Concordia Contest evaluation, NeurIPS 2025 · Village Zero decision and validation rules · Eight matched histories

When does a conversation begin?

AI Town is a starter kit for a multiplayer world that people can join. A character can accept a conversation invitation while still too far away to talk. The game tracks the walk and marks the character as participating once close enough. Dialogue prompts can draw on earlier encounters. The memory layer summarizes conversations and ranks a player’s memories using similarity, recency and importance.

AI Town project · Conversation membership · Player-scoped memory retrieval

AI Town batches simulation ticks into a step. Its saveWorld mutation commits game-state updates transactionally in Convex, and submitted inputs have recorded results or errors. This lets an engineer check that a world update was applied consistently. To measure missed opportunities in Village Zero, we also need the options the particular resident received before the update. A committed database write and a record of the choices available at decision time establish different facts.

Input results and game steps · World-saving mutation · Convex transaction semantics · Village Zero decision record

What changes when the unit is a population?

OASIS

A recommender selects the posts an agent sees. An hourly activity profile governs when it can respond. Agents remember encountered posts and choose actions such as reposting or doing nothing. The server updates the platform state. In a propagation experiment the authors compared simulated trajectories with the first 150 minutes of observed activity. This makes exposure through a feed part of the explanation for diffusion.

Yang et al., OASIS · Recommendation, memory and activation, §2 · Propagation experiment, Appendix F.2.2

AgentSociety

AgentSociety links urban movement, social contacts and an economy. A resident’s memory includes a profile, changing state and streams for events and perceptions. In its basic-income study, the authors compared simulations with and without a monthly transfer, following consumption and survey responses. Those measures describe the simulated population. Applying them to people requires separate empirical evidence.

Piao et al., AgentSociety · Resident memory, §3.6 · Income intervention and measurements, §7.4

What does a minute measure?

The Smallville authors report several days of elapsed time and thousands of dollars in token credits for two simulated days. OASIS reports 26 elapsed minutes for a particular 300-agent, 30-step run on one A100. AgentSociety measured wall time per round under several parallel group configurations. AI Town supports local or hosted inference, and Concordia’s model calls depend on the agents and GM components a scenario selects.

Smallville costs, §8.2 · OASIS runtime, Appendix F.2.2 · AgentSociety timings, Table 4 · AI Town inference options · Concordia component design

Across Village Zero’s eight Long-Life pairs, the post-fork branches used the same local model. A used 1,152 decision calls and 148.9 minutes of model evaluation. B used 2,943 calls and 387.6 minutes. These times sum the provider’s prompt processing and generation durations. Comparing efficiency across projects would require matching models, decision opportunities, simulated horizons and timing boundaries. The available studies do not supply that matched comparison.

Calls and durations by phase · Measured hardware and timing definitions · Long-Life outcomes

Sources checked . Implementation links use Concordia commit b3de736 and AI Town commit c9a37aa.