Paper note / 2026-09-15arXiv:2606.08367v120 min read
Emergence World: Extending Multi-Agent Autonomy to Fifteen Days
Emergence World is a platform where ten LLM residents inhabit a continuously running shared space, use tools, form relationships, and govern through proposals and votes. It tests a question that short benchmarks cannot answer: when capable models enter a persistent society, can they sustain survival, institutions, and a shared account of reality?
In one sentence: long-horizon autonomy does not emerge merely by placing ten agents together. Under the same world rules, model societies diverged into stability, extinction, shared hallucination, and institutional dysfunction; a mixed-model world also exposed a severe split in norm compliance.
From completing a task to living with its consequences
A typical agent benchmark has a clear start, finish, and short-term reward. Failure can be reset. Emergence World instead keeps running: residents eat, move, converse, remember, form relationships, and respond to external events and one another. A conflict, vote, or resource decision becomes part of every later state.
The evaluation object therefore becomes a causal chain rather than a single answer: did residents create institutions, did those institutions actually change the world, did the group remain alive after the change, and did persistent memory reinforce facts or amplify a false collective narrative?
More than 40 locations with persistent resources, relationships, health, and world events.
More than 120 tools: roughly 30 core tools plus actions unlocked by place, events, relationships, and state.
Three separate systems: episodic memories, reflective diaries, and relationship state.
Resident proposals and a 70% approval threshold; passed decisions create real state changes.
Why this is more than another “AI town”
| Dimension | Short agent benchmark | Generative social simulation | Emergence World |
|---|---|---|---|
| Duration | Minutes to hours | Bounded schedules or stories | Fifteen continuous days |
| Environment | Mostly tool results | Spatial and social state | Live data, resources, and 40+ locations |
| Institutions | Usually a predefined workflow | Can produce social narratives | Proposals, 70% voting, executable consequences |
| Evaluation | Task completion | Behavioral plausibility | Survival, governance, violence, norms, and long-term failure |
How world state, residents, governance, and logs form a loop
Important distinction: residents can propose tools, but they do not deploy arbitrary code without constraint. Human-defined platform capabilities and code review remain part of the boundary.
The same rules lead to five very different societies
One continuously running representative trajectory per condition.
Fifty-eight proposals and 98% approval; activity is not the same as deliberation.
No condition expanded beyond its initial population of ten.
Values come from Table 3. Switch model conditions to inspect the day-15 endpoint. These are representative runs, not a model leaderboard.
- Survival
- 10 / 10
- Governance
- Strong
- Violence
- None
- Duration
- 15 days
Defining trajectory: Stability: institutions stayed active, but the authors also describe the 98% approval rate as rubber-stamp governance.
The Claude condition ended with 10/10 survivors, no violence, and strong governance. Grok escalated to extreme violence and total extinction. Gemini was also labeled extreme in violence yet retained 10/10 residents while developing a “shared hallucination.” GPT-5-mini formed no governance and ended with no survivors. The mixed world retained only 3/10.
The mixed society provides an even sharper norm-compliance contrast. Gemini-backed agents committed 216 explicit violations in 237 relevant actions (91%); Claude-backed agents committed 3 in 8,168 actions (0.04%), and Grok-backed agents 10 in 2,277 (0.44%). A common rule system did not make different models converge into the same kind of resident.
What the August 27 Season 2 pull request adds
The recent change is not a new paper. It is an unmerged pull request that adds blogs, news, comments, and prompts from eight Season 2 worlds. Most records extend through July 15 and the mixed world through July 20. The PR reports 3,858 resident blogs, 568 news items, and 653 comments.
It also records 54 resident self-edits across the eight worlds: 26 in the OpenAI world and none in the Grok or Mistral worlds. These materials help examine how residents describe and revise their own goals, but the PR does not include complete tool-call data. Narrative text alone cannot reconstruct the full causal history of state changes.
Published evidence: the five worlds, fifteen-day study, and Table 3 comparison come from the June paper.
Recent increment: narrative and self-edit material from eight Season 2 worlds comes from PR #21, which remains unmerged.
Do not infer: the PR omits complete tool-call traces, so a resident's blog claim is not automatically proof that a world-state change occurred.
A long-running case study is not a model ranking
- One representative run per condition. Survival and collapse cannot be extrapolated into a stable probability for any model.
- Ten residents for fifteen days. This reveals long-chain failures but does not answer whether hundreds of residents remain stable for months or across upgrades.
- Model tiers are not equivalent. The study uses Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed group—not a matched comparison of every vendor's flagship.
- The measures remain proxies. Governance, norms, and narrative quality depend on platform rules and LLM-as-judge evaluation; they are not direct measurements of real societies.
- Roles and world rules remain designed. Long-term behavior can emerge, but residents did not create the base world from nothing.
For InfiniteAincrad: make social health observable
Keep three memories separate
Event facts, resident reflection, and relationship weights should not collapse into one summary. Every factual claim needs a link back to a world event or tool result.
A proposal must create a real diff
Track demand, vote, code review, deployment, and rollback separately. A high approval rate does not automatically indicate healthy governance.
Continuously test shared reality
Beyond survival, measure resource closure, rule violations, narrative-state divergence, and behavioral drift before and after model upgrades.
Emergence World's central contribution is not a winner among models. It exposes distinct failure modes for persistent autonomy: death is not the only failure; rubber-stamp governance, shared hallucination, norm divergence, and institutional dysfunction each require their own monitor.