RESEARCH · CHI PLAY COMPANION ’26 · ACM

Orchestrated Reality

From Character Role-Play to Orchestrated Reality — LLM-driven game world simulation as a Parameterized-Action POMDP
Yuhang Huang, Chenmiao Li (The University of Tokyo) · Chaowei Fang (Independent Researcher)
CHI PLAY Companion ’26 · Companion Proceedings of the Annual Symposium on Computer-Human Interaction in Play
📄 ACM · DOI arXiv preprint ⭐ Star WorldLines 🎮 Play Stoneford
ABSTRACT
Bridging tightly-authored narrative with deeply-simulated worlds has been prohibitively expensive. LLM-driven worlds open a new path — but today's deployed systems let the narrative voice assert state in free prose, with no validated representation of the world. We treat this as an architectural choice, not a limitation of language models: Orchestrated Reality makes the world a canonical object owned by a singleton orchestration agent, analogous to a tabletop-RPG Game Master. We formalise an LLM-driven game world for a human player as a Parameterized-Action POMDP — state is a tree of canonical JSON entities, actions decompose as a = (k, xk), the player observes only a narrative projection o = O(s), and the transition kernel F is the LLM-driven Plan–Diff–Validate–Apply pipeline committing schema-validated, content-hashed JSON deltas — with a worked single-turn example and a catalogue of 15 incidents from a real deployment.
SYSTEM Ⅰ — THE TRANSITION KERNEL

The PDVA pipeline

The narrative never owns the world. Every turn, the orchestrator runs one transaction over canonical JSON state — story is rendered from state, not asserted into it.

soul α soul β soul γ propose Plan Diff Validate Apply ✗ invalid → bounced, never committed world state ├─ places · actors ├─ inventory · wounds └─ promises · ledger canonical JSON · content-hashed
CONCEPT · ONE WRITER HOLDS THE PEN — EVERY ACTION PASSES VALIDATION BEFORE IT BECOMES WORLD STATE
01 · PLAN

Plan

The LLM interprets the player's intent (k, xk) against the current world and proposes what should change.

→
02 · DIFF

Diff

The proposal is expressed as explicit JSON deltas against the canonical entity tree — never free prose.

→
03 · VALIDATE

Validate

Deltas are schema-validated and content-hashed; anything inconsistent with the world's rules is rejected.

→
04 · APPLY

Apply

Valid deltas commit atomically. The new state is narrated back as o = O(s).

SYSTEM Ⅱ — THE AGENT SOCIETY

The World-Agent architecture

The world is a society of independent agents connected by one harness — physical consistency in the world, cognitive consistency in every NPC. Don't code the agent; use the agent to orchestrate the world.

WORLD-AGENT

The singleton Game Master. Owns the canonical world — state, routing, narration, archive. The only agent allowed to commit deltas through PDVA.

PLACE-AGENTS

Locations with behaviour. Towns, dungeons, combat and story scenes — routed to by the world-agent as the player moves through the world.

SOUL-AGENTS

One per character, each its own mind. Private memory, secrets, and an agenda — NPCs perceive, think, and act on their own, not as lines of dialogue data.

PLAYER

You, through an avatar terminal. You see only the narrative projection o = O(s) — the world persists whether you look or not.

s ∈ 𝒮

Canonical state: a tree of schema-typed JSON entities — who is where, what has just happened, what is currently true.

a = (k, xk)

Parameterized actions: a discrete intent kind plus structured JSON parameters.

o = O(s)

Partial observability: the player reads a narrative projection of state, never the state itself.

F = PDVA

The transition kernel is the Plan–Diff–Validate–Apply pipeline: schema-validated, content-hashed deltas.

RESEARCH SERIES

The papers behind WorldLines

PAPER Ⅰ · CHI PLAY COMPANION ’26 · THIS PAGE

From Character Role-Play to Orchestrated Reality: LLM-Driven Game World Simulation as a Parameterized-Action POMDP

The formal model — Parameterized-Action POMDP, canonical JSON state, and the PDVA transition kernel.

DOI: 10.1145/3800965.3834285 → arXiv → Reference implementation →
PAPER Ⅱ · IPSJ EC2026 · KYOTO

Multi-Agent LLM Simulation for Living Game Worlds: Coordinating World, Soul, and Player Agents over a Pub/Sub Message Bus

The society layer — world, soul, and player agents as peers on a pub/sub bus; souls run concurrently, the world agent is the single writer. Presented in Kyoto (Sep 2026) with the playable demo 時間陣 (Jikanjin).

Paper page →
PAPER Ⅲ · UNDER ANONYMOUS REVIEW · 2026

RP-Worlds: An Architecture for Unbounded yet Consistent World Simulation via PA-POMDP

The growth paper — play expands the canonical world through validated transitions; it grows without drifting. Preprint after review.

Paper page →

Overview of all three: worldlines.gg/research →

CITATION
@inproceedings{huang2026orchestrated,
  title     = {From Character Role-Play to Orchestrated Reality: LLM-Driven
               Game World Simulation as a Parameterized-Action POMDP},
  author    = {Huang, Yuhang and Li, Chenmiao and Fang, Chaowei},
  booktitle = {Companion Proceedings of the Annual Symposium on
               Computer-Human Interaction in Play (CHI PLAY Companion '26)},
  publisher = {Association for Computing Machinery},
  year      = {2026},
  doi       = {10.1145/3800965.3834285},
  url       = {https://doi.org/10.1145/3800965.3834285}
}

The reference implementation is open.

WorldLines — the orchestrator-agents engine for living worlds. File-backed, event-sourced, playable today.

⭐ Star WorldLines on GitHub