Multiverse of Human Games

• The multiverse of human games is a research framework defining the infinite space of human-designed games, parallel canons, and multiple strategic equilibria.
• It formalizes game interactions using goal-conditioned POMDPs, orchestrated reality pipelines, and equilibrium analyses to benchmark human-model performance.
• Empirical findings, including AI GameStore evaluations and role-play metrics, reveal significant human–model gaps and challenges in memory, planning, and canonical state persistence.

The Multiverse of Human Games denotes, in current research usage, a family of closely related but technically distinct ideas about plurality in games: the open-ended space of games “designed by humans for humans”; the distribution of rule systems, aesthetics, physics, and objectives across many games; the multiplicity of equilibria and dynamical regimes in multiplayer strategic interaction; and the existence of multiple persistent worlds, shards, timelines, or canon-specific versions in interactive simulation and role-play (Ying et al., 19 Feb 2026, Zhang et al., 11 May 2026, Gokhale et al., 2010, Huang et al., 14 Jun 2026, Ngokpol et al., 16 Oct 2025). Across these usages, the common theme is that game analysis is no longer confined to a single fixed environment. Instead, it concerns families of environments, worldlines, or strategic landscapes whose structure must be formalized, compared, simulated, and evaluated under human-relevant constraints.

1. Meanings of “multiverse” across game research

One line of work defines a human game as a game “designed by humans for humans,” and the Multiverse of Human Games as the entire open-ended, in-principle infinite space of all games that humans could conceivably create and enjoy, together with the human cultural distribution over that space—how likely such games are to be invented, propagated, and played (Ying et al., 19 Feb 2026). This definition intentionally excludes arbitrary computable environments that no human would design or enjoy. In that sense, the multiverse is a task space and a distribution.

A second line of work uses multiverse to describe the proliferation of possible outcomes in evolutionary game dynamics as group size and strategy count increase. In Gokhale and Traulsen’s formulation, the term refers to the multiplicity of possible equilibria and dynamical regimes as multiplayer and multi-strategy effects replace pairwise interactions; it is explicitly not a metaphysical multiverse (Gokhale et al., 2010). Here, the multiverse is an equilibrium landscape.

A third line of work treats the multiverse as a collection of parallel canons, worlds, shards, or timelines. “Beyond One World” studies superheroes across 30 iconic heroes and 90 canon-specific versions, with multiverse awareness defined by version-specific factual and moral fidelity (Ngokpol et al., 16 Oct 2025). “Orchestrated Reality” extends this to persistent, content-addressed game worlds that can branch, migrate entities, and support multiple concurrent worlds or timelines (Huang et al., 14 Jun 2026). Here, the multiverse is a provenance-governed space of narrative states.

A fourth line of work in generalist game playing frames games themselves as a vast multiverse because each game is a universe with its own rules, aesthetics, physics, and objectives. Human intelligence is characterized there by “omni-reality adaptability,” namely rapid transfer from one physical reality to many synthetic rule systems (Zhang et al., 11 May 2026).

Context	Meaning of “multiverse”	Representative formal object
Human-game evaluation	Open-ended space of all conceivable human games	$\mathcal{H}$ with associated distribution $P_{\mathcal{H}}$
Evolutionary games	Multiplicity of equilibria and dynamical regimes	Isolated internal equilibria in multiplayer games
Role-play and simulation	Parallel canons, worlds, shards, or timelines	Version tags, canonical state, content hashes
Generalist game playing	Distribution of heterogeneous game universes	Goal-conditioned POMDP instances

This plurality of meanings is not merely terminological variation. It indicates that “multiverse” has become a reusable abstraction for breadth, branching, and multiplicity in game-centered research.

2. Formal models: from game distributions to canonical world state

A central formalization treats interaction with a game as a goal-conditioned POMDP,

$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$

where $G$ are natural-language goals or tasks, $S$ states, $A$ actions, $T$ transitions, $R$ feedback conditioned on $g\in G$, $\Omega$ observations, $P_{\mathcal{H}}$0 observation model, and $P_{\mathcal{H}}$1 discount. The agent seeks a policy $P_{\mathcal{H}}$2 that maximizes

$P_{\mathcal{H}}$3

This formulation is used to describe the shift from fixed-rule, fully observable settings to partially observable, multimodal, long-horizon, real-time environments with open-ended goal spaces (Zhang et al., 11 May 2026).

For human-game evaluation, an inferred formalization denotes the set of all conceivable human games by $P_{\mathcal{H}}$4 and an associated distribution by $P_{\mathcal{H}}$5, with a budget-conditional evaluation function

$P_{\mathcal{H}}$6

where $P_{\mathcal{H}}$7 fixes the resource budget and $P_{\mathcal{H}}$8 outputs a normalized, human-relative score. This suggests that the multiverse is operationalized not only as a set of games but also as a distribution over culturally generated tasks under matched resource constraints (Ying et al., 19 Feb 2026).

“Orchestrated Reality” gives a more specific world-model formalization for an LLM-driven game world for a human player as a Parameterized-Action POMDP,

$P_{\mathcal{H}}$9

Here, each $$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$0 is a tree of typed JSON documents under a fixed schema, fully serialisable and addressable by content hash $$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$1. Actions decompose as

$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$2

with a discrete intent kind and structured JSON parameters. The player does not observe $$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$3 directly, but instead receives a narrative projection

$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$4

and the transition kernel

$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$5

is implemented by a Plan–Diff–Validate–Apply (PDVA) pipeline that proposes, validates, and atomically commits JSON deltas (Huang et al., 14 Jun 2026).

In that architecture, the world is the canonical object owned by a singleton orchestration agent analogous to a tabletop-RPG Game Master. The narrative voice is an observation of canonical state rather than a stateful authority. State persistence is achieved through a schema-validated JSON tree on disk, atomic content-hashed commits, and event journaling. This yields a mutation boundary at which prose may suggest outcomes, but only validated JSON deltas advance state (Huang et al., 14 Jun 2026).

A plausible implication is that the multiverse perspective requires more than broader task coverage. It also requires explicit contracts over state, action, observation, provenance, and branching semantics.

3. Human-game evaluation and the AI GameStore

The most explicit operationalization of the Multiverse of Human Games as a benchmark appears in AI GameStore (Ying et al., 19 Feb 2026). The platform uses an LLM+human-in-the-loop pipeline to source, generate, refine, and annotate games aligned with human play, while standardizing interfaces for reproducible evaluation. The first suite contains 100 adapted games spanning Action, Casual, Puzzle, Board, and other categories.

Its sourcing stage scraped 7,500 candidates across categories, countries, and platforms from Apple App Store and Steam, then filtered by popularity ($$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$6 reviews) and high rating ($$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$7). An LLM judge, Gemini 2.5 Flash, scored suitability using criteria including playable in minutes, implementable in p5.js, quantifiable scoring, and minimal reliance on niche knowledge; the top 100 were retained for adaptation. In refinement, Claude sonnet-4.5 produced p5.js games, automated debugging iteratively fixed bugs, and human players supplied natural-language feedback. The reported average was 4.7 refinement steps and roughly 30 minutes end-to-end per game (Ying et al., 19 Feb 2026).

Each game was labeled on a 0–5 scale for Visual Processing (VP), Spatial-Temporal Coordination (ST), Memory (ME), Planning (PL), World Model Learning (WM), Physical Reasoning (PH), and Social Reasoning (SO). This makes the benchmark diagnostic rather than purely aggregate. The paper emphasizes that many games jointly require multiple capabilities, enabling tests of integration such as VP+PL+WM or ME+PL (Ying et al., 19 Feb 2026).

The evaluation protocol matched humans and models on gameplay duration: 120 seconds per game. Humans interacted naturally through a unified web interface. Models interacted through a harness that pauses the game each second, provides game description, screenshots, an editable scratchpad, and past actions, then receives five action lists for the next one-second window partitioned into five 0.2-second segments. Scoring used the author-provided normalization

$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$8

with the geometric mean across games as the principal aggregate. Human median is set to 100 per game (Ying et al., 19 Feb 2026).

The empirical setup included N=106 Prolific participants, each playing 10 games, and seven frontier VLMs: GPT-5.2, GPT-5-mini, Gemini-2.5-Pro, Gemini-2.5-Flash, Claude-Opus-4.5, Qwen-3-VL-32B, and Llama-4-Maverick. Each model was run three times per game with default temperature and “auto” thinking budgets; results were averaged across runs (Ying et al., 19 Feb 2026).

The headline result is a large human–model gap. Geometric-mean normalized scores are under 10% of human baseline for the best models, and models require on average 12–18$$\mathcal{M}=\langle G,S,A,T,R,\Omega,O,\gamma\rangle,$$9 more wall-clock time than humans for the same 120 seconds of in-game interaction. Performance is bimodal: on roughly two thirds of games, models make some progress, often 10–30% of human median; on the remaining 30–40%, they fail to achieve meaningful progress (<1% of human median). Capability-conditioned analyses show pronounced deficits in Memory, Planning, and World Model Learning, and performance declines sharply as games demand more distinct capabilities simultaneously. Restricting to games with ST $G$0 does not materially improve the top models’ aggregate performance, suggesting that slow reaction or harness cadence is not the primary failure mode (Ying et al., 19 Feb 2026).

These findings situate the Multiverse of Human Games as a living, open-ended meta-benchmark rather than a static leaderboard. The platform’s anti-saturation measures—only 10 games public, 90 kept as a private test set, and continual introduction of new variants—are part of the benchmark definition itself (Ying et al., 19 Feb 2026).

4. Multiversal role-play and persistent playable worlds

Multiverse-aware role-play focuses on character identity across parallel canons. Beyond One World benchmarks character-grounded role-play for 30 iconic heroes and 90 canon-specific versions, with three narrative phases per character: Childhood, Pre-Hero, and Hero. The benchmark contains two tasks: Canon Events, with 1,346 items total, and Moral Dilemmas, with 1,080 dilemmas across four archetypes (Ngokpol et al., 16 Oct 2025).

The scoring framework explicitly separates internal deliberation from outward decisions. Responses are segmented into <thinking> and <acting> spans, then scored on a 0–5 scale by Sonnet 3.7 using version-specific attributes. The benchmark introduces Think–Act Matching (TAM), computed from all-mpnet-base-v2 embeddings as the maximum cosine similarity between thinking and acting spans:

$G$1

Dataset-level TAM is the mean of instance-level scores (Ngokpol et al., 16 Oct 2025).

The reported findings are threefold. First, chain-of-thought prompting improves narrative coherence in weaker models but can reduce canonical accuracy in stronger ones. Second, cross-version generalization within a character remains a major obstacle. Third, models often excel at either thinking or acting, but rarely both. For example, sonnet3.7 acts highest with $G$2 but thinks moderately with $G$3, whereas gemini2-flash thinks highest with $G$4 and has moderate acting with $G$5 (Ngokpol et al., 16 Oct 2025).

The benchmark’s practical recommendations are structurally similar to those found in persistent world simulation: modular canon memory keyed by character, universe, and phase; explicit version tags and provenance; periodic Canon Event probes; and reason–action audits using TAM. This suggests a shared requirement across multiverse role-play and persistent worlds: canonical state must be explicit, versioned, and resistant to leakage across worldlines (Ngokpol et al., 16 Oct 2025).

That same requirement is formalized at the world-simulation level in Orchestrated Reality. The framework treats the game world as the canonical object and assigns ownership to a single orchestration agent—the world-agent—analogous to a tabletop RPG Game Master. It addresses three failures of many LLM-driven experiences—statelessness, unvalidated writes, and monolithic agency—by modeling the world as a persistent, schema-validated JSON tree on disk, advancing it through atomic, content-hashed commits, and separating concerns across sub-agents coordinated by a singleton GM (Huang et al., 14 Jun 2026).

The paper provides a worked single-turn example, a JSON-state sketch, and a catalogue of 15 illustrative incidents drawn from a real deployment. It also describes scaling from a single world to a multiverse of human games through branching timelines and versioning, world seeding and rule parameterization, cross-world portability of typed JSON entities, migration intents across world boundaries, multi-NPC concurrent agency under deny-first permissions, and RL interfaces that expose $G$6 as an environment for automated play, regression testing, and content QA (Huang et al., 14 Jun 2026).

The limitations are equally central. Open problems include deterministic record/replay under LLM stochasticity, per-actor cost discipline as NPC concurrency grows, narrative drift for under-specified details, incomplete rule coverage $G$7, and the fact that prose observations $G$8 are not themselves validated (Huang et al., 14 Jun 2026). In both role-play and orchestrated worlds, multiversality depends on canonical memory and provenance control, but factual fidelity remains an active systems problem.

5. Strategic and behavioral multiverses in human interaction

In multiplayer evolutionary game theory, the multiverse is a landscape of equilibria. For two-player games with any number of strategies there can be at most one isolated internal equilibrium. For any number of players $G$9 with any number of strategies $S$0, the maximum number of isolated internal equilibria is

$S$1

This upper bound is obtained by treating internal equilibria as solutions to $S$2 polynomial equations of degree at most $S$3 and applying a Bézout-type counting argument (Gokhale et al., 2010).

For two strategies, the replicator equation is

$S$4

and in $S$5-player games the payoffs are degree-$S$6 polynomials in $S$7:

$S$8

Hence there can be up to $S$9 interior fixed points. With $A$0, intersections of higher-degree algebraic curves or surfaces on the simplex generate multiple isolated equilibria with diverse stability patterns (Gokhale et al., 2010).

The paper maps this formal proliferation to human-relevant scenarios. In public goods games, linear returns yield at most one interior equilibrium, but threshold public goods can yield up to two interior equilibria. In multiplayer stag hunts, increasing thresholds and synergies can generate multiple interior equilibria and fragile payoff dominance. The broader implication is that pairwise-derived results, including the one-third law, do not generalize cleanly to multiplayer multi-strategy settings (Gokhale et al., 2010).

A complementary behavioral notion of multiverse appears in the dynamics of human cooperation in repeated Prisoner’s Dilemma. There, distinct long-run behavioral regimes emerge from the closed loop between simple human reinforcement learning and an opponent’s strategy. Using the stage-game payoffs

$A$1

and learning rates

$A$2

the paper studies human play against strong extortion

$A$3

and strong generous

$A$4

zero-determinant opponents (Spanknebel et al., 2015).

In the context-conditioned model, dynamics unfold in the two-dimensional space $A$5 with unstable fixed points and sharply partitioned basins of attraction. Against the strong generous ZD, the most important stable manifold empirically is the one with $A$6, where mutual cooperation is absorbing. Against the strong extortion ZD, the dominant basin intersecting the initialization range leads to permanent mutual defection through the stable manifold $A$7 (Spanknebel et al., 2015).

The empirical benchmark matched human data from Hilbe et al. (2014). Observed initial human cooperation was 30–40%; against generous opponents, cooperation rose to ~70–80% with mean payoff ≈ €0.27; against extortion, it declined to ~30–40% with mean payoff ≈ €0.15. Simulations with N = 10201 agents reproduced the time courses and end-of-game distributional shapes, and individual-level prediction using the first 10 rounds gave a linear regression with slope 0.69 and intercept 0.26 for predicted versus observed cooperativity (Spanknebel et al., 2015).

Here, the multiverse is a parameterized behavioral regime space. Extortionate and generous opponents carve different basins; learning parameters and initial conditions determine which basin is reached. This is not a separate use of the term so much as a dynamical specialization of it.

6. Formal universes, limits, and research directions

Absolute combinatorial game theory introduces yet another rigorous sense in which human games inhabit a multiverse. Under misère play, there are infinitely many absolute universes, obtained by recursive parental expansions from dicot and dead-ending universes. Under normal-play, there are exactly two absolute universes: the full space of short normal-play games and the universe of all-small games (Larsson et al., 2023).

A universe $A$8 is required to contain all atoms, be closed under disjunctive sum and conjugation, and be hereditarily closed. The key local comparison result is the Basic Order Theorem, under which $A$9 in an absolute universe iff both Proviso and Maintenance hold. The minimal parental universe is the dicot universe $T$0. The dead-ending universe $T$1 is parental and contains placement rulesets, with $T$2 (Larsson et al., 2023).

This framework has direct implications for human-playable rulesets. Nim belongs to $T$3 under normal-play and to $T$4 under misère. Kayles, Dawson’s Kayles, and Treblecross are dead-ending. Domineering is dead-ending under misère and typically not dicot under normal-play. Hackenbush and Toads-and-Frogs are neither dead-ending nor dicot, and therefore live in the full universes (Larsson et al., 2023). The multiverse here is a family of comparison universes in which equivalence and strategy depend on the closure conditions one assumes.

Across the broader literature, several open problems recur. In orchestrated world simulation, planned evaluations include a within-subjects player study using miniPXI and PAX, promotion of multi-NPC concurrent agency to a parameterized-action Markov game, and deployment as an RL environment (Huang et al., 14 Jun 2026). In AI GameStore, next steps include richer social interaction, long-horizon games with narrative and large state, automated reliable level generation, deeper capability profiles, strict real-time constraints, and continued anti-cheating safeguards through private test sets and replay logging (Ying et al., 19 Feb 2026). In generalist game playing, the field is organized around four pillars—Dataset, Model, Harness, and Benchmark—and constrained by five trade-offs: Scale vs Fidelity vs Diversity, Breadth vs Depth, Reasoning vs Reactivity, Modular Workflow vs Model-as-Whole, and Code Engine vs World Model (Zhang et al., 11 May 2026).

That same work proposes a five-level roadmap from single-game task mastery to cross-task transfer, cross-genre generalization, lifelong adaptation, and finally the Creator (Demiurge) stage, in which the agent generates and evolves worlds, rules, tasks, rewards, and co-evolves populations (Zhang et al., 11 May 2026). A plausible implication is that the Multiverse of Human Games is becoming both a descriptive object and a systems objective: a way to characterize the breadth of human-designed games, the branching of persistent worlds, the multiplicity of strategic equilibria, and the technical path toward agents that can act, learn, and remain consistent across all of them.