Multi-Turn Ultimatum Game Dynamics

• Multi-Turn Ultimatum Game is an extension of classic bargaining that models sequential negotiation with repeated offers and adaptive strategy updates.
• Key mechanisms include reactive adjustment rules, spatial evolutionary imitation, and mobility that collectively drive fairness and compromise.
• Quantitative analyses reveal that varied adaptive strategies, such as Greedy, Moderate, and Conservative, converge to fair splits and reduced payoff inequality.

A multi-turn ultimatum game generalizes the classic ultimatum bargaining protocol to accommodate repeated, iterated, spatially-structured, or mechanism-driven negotiation scenarios, allowing for richer learning dynamics, evolutionary adaptation, and complex emergent behavior. In contrast to the one-shot game, where a single proposal divides a unit surplus and the responder accepts or rejects, multi-turn models introduce explicit sequential structure: reactive adjustments by the proposer, role alternation, stochastic continuation, or endogenous adaptation of strategies over time and across social networks. These extensions allow the investigation of fairness emergence, compromise rules, strategic imitation, and the impact of noise and spatial mobility.

1. Reactive Evolutionary Foundations

Two principal frameworks dominate the study of multi-turn ultimatum games: repeated bargaining with probabilistic-reactive update rules and evolutionary adaptation in spatially structured populations. In the reactive paradigm, agents iteratively adjust their proposals in response to accept/reject outcomes, often governed by a linear rule: $p_{t+1} = p_t + \alpha(\text{result}_t - \beta)$ where $p_t$ is the proposer’s offer at turn $t$, $\alpha$ a small learning rate, and $\beta$ a reference aspiration level. Acceptance is typically stochastic, for example by comparing $p_t$ to a uniformly drawn random variable $r \in [0,1]$ so that $\Pr(\text{accept}|p_t) = p_t$ (Silva et al., 2019).

Embedding such dynamics within a finite lattice (e.g., two-dimensional periodic square lattice) enables modeling of local interaction, evolutionary competition, and imitation (so-called "Darwinian copying") based on cumulative payoffs. The probability that an agent copies the strategy of a more successful neighbor is governed by the Fermi rule: $$W(i \leftarrow j) = \frac{1}{1+\exp[-(\pi_j-\pi_i)/K]}$$ where $\pi_i$ is cumulative payoff and $K$ a noise parameter (Silva et al., 2019).

2. Taxonomy of Adaptive Strategies

Three principal adjustment strategies for reactive updating have been identified (Silva et al., 2019), forming a basis for broad behavioral classification:

• Greedy (G): Decrease offer following any acceptance, otherwise increase.
• Moderate (M): Decrease if at least half of proposals in a round are accepted.
• Conservative (C): Only decrease if all proposals are accepted; otherwise increase.

The dynamical evolution of average offers $\langle p_t \rangle$ under these strategies obeys high-degree polynomial mean-field equations, with parameter dependence set by the mix of employed strategies. Pure and mixed portfolios admit analytic fixed points; for instance, pure Greedy equilibrates to $p_\infty \approx 0.16$, Moderate to $p_\infty \approx 0.39$, Conservative to $p_\infty \approx 0.84$, and a specific mixed portfolio achieves $p_\infty = 0.5$ (Silva et al., 2019).

These results are structurally consistent with other models employing single-parameter Pavlovian update rules (Silva et al., 2015): when offer-dependent acceptance $p_t = O_t$ is used, mean-field and networked populations converge to the fair split ($x^* = 1/2$) under most adaptive policies.

3. Network Structure and Social Learning

Population topology—specifically, network structure and local mobility—profoundly impacts equilibrium outcomes and fairness emergence. In spatially embedded models (e.g., square lattices or arbitrary graphs with coordination number $k$), imitation and role alternation lead to patchwise dominance of certain strategies. Mobility (diffusion) implements spatial remixing: random exchanges of agent positions with probability $\alpha$ per pair.

A critical value $\alpha_c$ partitions phase space: for $\alpha < \alpha_c \approx 0.26$, conservative strategies dominate, sustaining high average offers ($p_\infty \approx 0.72$); for $\alpha > \alpha_c$, moderates prevail, reducing average offers to $p_\infty \approx 0.5$ and diminishing payoff inequality as measured by the Gini index. At $\alpha \approx \alpha_c$, long-lived coexistence and clustering of C and M patches are observed (Silva et al., 2019).

Stochastic evolutionary imitation, even with occasional error (nonzero $K$), ensures that locally competitive but globally myopic strategies are not stable in high-mobility, well-mixed populations.

4. Stochasticity, Continuation, and Fairness

Variants such as the Not Quite Ultimatum Game (NQUG) (Ichinose et al., 2014) highlight how stochastic decision-making and voluntary continuation drive the evolutionary stability of fairness. Here, both proposers and responders sample actual offers and thresholds from truncated normal distributions centered at genotypic means, with noise parameter $\sigma$: $$p' \sim \text{TruncNorm}(p, \sigma), \quad q' \sim \text{TruncNorm}(q, \sigma)$$ If $p' \geq q'$, acceptance occurs; else, rejection triggers, with probability $r$, a new round with the same roles. Stochasticity (nonzero $\sigma$) penalizes low offers due to risk of rejection, while the possibility of continued bargaining (high $r$) allows responders to extract higher offers.

Evolutionary simulation reveals three regimes:

• For $\sigma \approx 0$ or $r$ small, rational (subgame-perfect) outcomes prevail (low $p, q$).
• For intermediate $\sigma$ and high $r$, fairness emerges ($p \approx q \in [0.4, 0.5]$).
• For very large $\sigma$ and $r \to 1$, "naysaying" occurs (responders demand more than they would themselves offer).

These dynamics are formalized by analytic expressions for equilibrium payoffs, acceptance probabilities, and responsiveness to mutation and noise (Ichinose et al., 2014).

5. Mechanism Design for Multi-Turn Compromise

The multimatum mechanism (Echenique et al., 29 Jan 2026) generalizes multi-turn negotiation to abstract compact policy spaces $X$, using set-valued offers (menus) instead of pointwise proposals. Rounds alternate: the proposer submits a closed nonempty set $A_1 \in \mathcal{F}(X)$; the responder either accepts by selecting $x \in A_1$ or counteroffers a set $A_2$ such that $\nu(A_2) \geq \nu(A_1)$ for a reference measure $\nu$.

This protocol fully implements compromise solutions—maximizing the minimum cardinalized rank of each player's preference $v_i(x) = \nu(L_i(x))$—in subgame-perfect Nash equilibrium. The mechanism is robust to infinite and high-dimensional outcome spaces and admits applications in public goods bargaining, other-regarding preference models, facility location, and lotteries (Echenique et al., 29 Jan 2026).

Unlike classic alternating-offer games, the multimatum protocol introduces endogenous discipline (via menu size constraints) and requires no discounting. Equilibrium is characterized using backward induction and ensures that the implemented compromise maximizes the minimum benefited cardinally across both agents.

6. Quantitative Results and Phase Behaviors

Key metrics in the study of multi-turn ultimatum games include:

• Strategy densities $(\rho_G, \rho_M, \rho_C)$
• Average offer $\langle p \rangle$
• Proposer/responder payoffs $\langle \pi_\text{pr/ac} \rangle$
• Inequality (Gini coefficient)

Evolutionary and Monte Carlo simulations (Silva et al., 2019, Ichinose et al., 2014, Silva et al., 2015) typify the following phenomena:

• Reactive-only dynamics reach strategy-specific equilibria unless diversified by mixed policies or imitation.
• Under spatial competition and mobility, fairness-promoting strategies invade at high diffusion.
• Mixtures of adaptive policies (e.g., equally weighted Greedy, Moderate, Conservative) yield robust $p^* = 0.5$ for any network degree $k$ (Silva et al., 2015).
• More conservative updating tends to raise mean offers and lower inequality; greedy updating depresses both and increases inequity.
• Mechanism-based negotiation, via the multimatum, implements fairness in both cardinal (lottery, location) and ordinal preference spaces (Echenique et al., 29 Jan 2026).

7. Theoretical and Practical Implications

Multi-turn ultimatum games demonstrate how coupling simple adaptive or stochastic learning with local imitation can yield robustly fair outcomes without requiring global statistics, explicit inequity aversion, or informational enhancements such as reputation. Fairness, understood as near 50--50 splits, emerges through local social learning and sufficiently frequent mixing; conversely, isolation or dominance by any one strategy (greedy or conservative) can cause persistent inequity.

The multimatum framework expands the scope of deliberative compromise to continuous, high-dimensional domains relevant to facility siting, political policy negotiations, and social choice without transfer payments, leveraging endogenous threat discipline to secure both equilibrium existence and implementability (Echenique et al., 29 Jan 2026).

A plausible implication is that real-world repeated bargaining environments—when equipped with sufficient social mobility and local information transfer—will stabilise fair, moderate negotiations, while static, unmixing environments risk entrenchment of excessive caution or exploitation, depending on the dominant reactive baseline.