Why Slower Learners Make Better Cooperators

In a population of purely selfish learners, cooperation should collapse. That’s the prediction of classical game theory: when individuals act only to maximize their own payoff, the only stable outcome in a social dilemma like the iterated prisoner’s dilemma is mutual defection. Yet in this new analysis, Wolfe and Plotkin (2026) show that cooperation not only emerges—it thrives—when learners differ in how fast they adapt. Not because they’re altruistic. Not because they plan ahead. But simply because some learn slowly.

The most striking result? A population with identical learning rates converges to universal defection, where average cooperation drops to near zero. But introduce variation—some agents learning quickly, others slowly—and the same selfish, myopic update rule leads the population to generous tit-for-tat (GTFT), a strategy that sustains cooperation at nearly optimal levels. The difference isn’t in the rules of learning. It’s in the rhythm.

This isn’t a simulation of robots playing nice. It’s a mathematical model of how individuals—humans, algorithms, organisms—adjust their behavior in response to others in a repeated social dilemma. And it suggests that diversity in learning speed, far from being noise, is a hidden engine of cooperation.

The Science

The researchers studied the iterated prisoner’s dilemma (IPD), the canonical model of cooperation under conflict. In each round, two players choose to cooperate—paying a cost $c$ to deliver a benefit $b>c$ to the other—or defect, paying nothing and giving nothing. A single round favors defection. But when the game repeats, players can condition their actions on past behavior, opening the door to reciprocity.

Instead of modeling evolution by imitation (as in most prior work), Wolfe and Plotkin modeled individual learning via gradient ascent. Each agent adjusts its strategy in the direction that most rapidly increases its payoff against the current population. The strategy space is two-dimensional: a reactive memory-one strategy defined by two continuous traits:

• Reciprocity ($p$): the probability of cooperating after the co-player cooperated.
• Forgiveness ($q$): the probability of cooperating after the co-player defected.

This space includes well-known strategies: always cooperate $(1,1)$, always defect $(0,0)$, tit-for-tat $(1,0)$, and generous tit-for-tat $(1,1-\delta )$, which forgives defection with some probability.

Each agent updates its $(p,q)$ strategy using the gradient of its payoff against all others in the population. The update is scaled by a learning rate ${\eta }_i$, which can be uniform across agents or vary from one to another. The model includes self-play and assumes a well-mixed population of $N=2500$ agents, with strategies bounded away from the edges by $\epsilon =0.001$ to ensure stable long-run payoffs.

The key innovation is treating learning rates as a fixed, individual trait—some agents are fast learners, others slow—and asking how this heterogeneity shapes collective outcomes. The dynamics are simulated over thousands of discrete learning steps, and also analyzed in the infinite-population limit using a continuous density formulation.

What They Found

The divergence between uniform and variable learning rates is stark. In simulations, every population—regardless of learning rate structure—passes through a phase of collapsing forgiveness: all agents rapidly reduce their $q$ (forgiveness) to the minimum $\epsilon$. This creates a transient “unforgiving state” where strategies are pinned to the line $q=\epsilon$, and only reciprocity $p$ can evolve.

What happens next depends entirely on the distribution of learning rates.

In a population with uniform learning rates (${\eta }_i=0.2$ for all), fast convergence to low forgiveness drags reciprocity down with it. By the time all agents reach $q=\epsilon$, their $p$ values have also dropped, leaving the population clustered near $(\epsilon ,\epsilon )$—effectively always defect (ALLD). From there, no escape: the gradient pushes further toward defection, and the population locks in.

But when learning rates vary—drawn from $\text{Uniform}[0.01,0.39]$—a different story unfolds. Fast learners still plunge to low forgiveness and low reciprocity. But slow learners, moving in smaller steps, preserve high $p$ values even as their $q$ drops. When the population hits the unforgiving line, there’s a residual tail of high-reciprocity agents near tit-for-tat $(1,\epsilon )$.

This tail changes everything.

Once on the $q=\epsilon$ line, the gradient begins to favor increases in $p$. Agents with high $p$ start pulling the population upward. Then, when $p$ nears 1, the gradient rotates: now it favors increases in $q$. The population begins to climb the right edge of the strategy space, moving up the line $p=1-\epsilon$ toward GTFT.

Figure 3 from the paper captures this four-stage trajectory: (1) collapse of forgiveness, (2) transient unforgiving state with stratified reciprocity, (3) move toward GTFT, and (4) stable cooperation. The color-coding by learning rate shows how slow learners (blue) anchor high reciprocity, enabling the transition.

The final state is generous tit-for-tat: $(p,q)\approx (1,1-c/b)$. With $c/b=0.5$, this means $q\approx 0.5$, so agents forgive defection half the time. This strategy is stable because it resists invasion: defectors get punished, but cooperation is restored after a mistake. And it produces nearly optimal collective payoffs—far above the zero-sum outcome of universal defection.

Average Payoff Over Time: Uniform vs. Variable Learning Rates

Label	Value
Uniform Learning	0.02
Variable Learning	0.42

Source: Simulations from Wolfe & Plotkin (2026). Payoff stabilizes near 0.45 with variable rates, near 0.02 with uniform.

The payoff difference is dramatic. In uniform-learning populations, average payoff per round collapses to near zero. In variable-learning populations, it stabilizes around 0.45—close to the theoretical maximum of $b-c=0.5$.

Even more striking is the recovery from perturbation. The researchers tested what happens when a stable GTFT population is slightly disturbed—say, by a random shock that increases forgiveness in some agents. Under uniform learning, the population cannot recover; it drifts away from cooperation. But under variable learning, slow learners act as anchors, pulling the population back to GTFT.

Figure S7 shows this explicitly: after a tiny perturbation, the uniform-learning population (dashed line) diverges, while variable-learning populations (colored lines) return to high cooperation. This resilience is provably impossible under uniform learning.

Why This Changes Things

For decades, the puzzle of cooperation has centered on mechanisms: direct reciprocity, reputation, punishment, group selection. The assumption has been that cooperation requires special cognitive or institutional scaffolding—foresight, memory, communication, or social norms.

This paper flips the script. It shows that cooperation can emerge from purely selfish, myopic agents—no foresight, no communication, no punishment—provided only that they learn at different speeds.

The implication is profound: cooperation may not depend on how smart we are, but on how differently we learn.

This reframes our understanding of social dynamics. In schools, workplaces, online communities, and international relations, we often assume that faster learning is better. But this model suggests that slow learners play a critical stabilizing role. They preserve cooperative norms during periods of upheaval, when fast learners might otherwise rush to defection.

Consider climate negotiations. Countries update their policies based on others’ actions. Fast-acting nations may defect if cooperation seems weak. But slower, more deliberative governments—those with institutional inertia—may maintain high reciprocity, anchoring the system until a cooperative equilibrium re-emerges.

Or consider AI systems trained to interact in repeated games. If all use the same learning rate, they may converge to mutual defection. But if learning rates are varied—by design or by accident—cooperation becomes possible, even likely.

The model also explains why cooperation persists in biological systems without centralized control. In microbial communities, for instance, individual cells may respond to neighbors at different speeds due to metabolic or genetic variation. This heterogeneity could sustain collective behaviors like biofilm formation or nutrient sharing, even under selfish incentives.

The threshold $c/b$—the cost-benefit ratio—also matters. When cooperation is cheap relative to its benefit, even modest heterogeneity can sustain it. But when $c/b$ approaches 1, cooperation requires greater variation in learning rates.

Probability of Cooperation by Benefit-Cost Ratio and Learning-Rate Heterogeneity

Label	Value
b/c = 1.5	20
b/c = 2.0	45
b/c = 3.0	75
b/c = 5.0	95

Source: Figure S9, Wolfe & Plotkin (2026). Cooperation probability rises with b/c and learning-rate variance.

shows how the probability of cooperation increases with both $b/c$ and learning-rate diversity.

This aligns with real-world patterns. High-stakes dilemmas (e.g., nuclear disarmament) require more structural support for cooperation—treaties, verification, phased commitments—than low-stakes ones (e.g., recycling). The model suggests that such structures may function, in part, by introducing temporal heterogeneity: some actors move fast, others slow, creating the conditions for stable reciprocity.

What's Next

The model has limits. It assumes well-mixed interactions, no noise in implementation, and a fixed set of learning rates. Real populations have network structure, mistakes, and evolving learning rules. Future work could test how robust the effect is under these conditions.

One open question is whether learning-rate heterogeneity can evolve. If fast learners gain short-term advantages, why wouldn’t selection eliminate slow ones? The answer may lie in the long-term stability slow learners provide. Populations with variable rates may be more resilient to shocks, giving them an evolutionary edge over time.

Another direction is policy design. Can we engineer learning-rate diversity in human or artificial systems to promote cooperation? In online platforms, for example, could we slow down certain users’ feedback loops to stabilize prosocial norms? In AI training, could we assign different learning rates to agents to prevent collapse into mutual defection?

The model also raises philosophical questions. If cooperation emerges not from virtue but from variation, what does that say about the nature of social order? Perhaps the secret isn’t in our morals, but in our mismatched tempos.

One caveat: the effect depends on the initial distribution of strategies. If all agents start with low reciprocity, even variable learning may not save cooperation. And in very small populations, stochastic effects can override the trend.

Still, the core insight stands: diversity in learning speed can transform a population of selfish agents into a cooperative one. Not by changing their goals. Not by adding new rules. But by varying the pace at which they adapt.

In a world obsessed with speed—faster algorithms, faster decisions, faster reactions—this is a quiet reminder that sometimes, the slow ones hold the system together.