Two well-supported results about how networks sustain cooperation point in opposite directions. Simulations of self-interested agents show that cooperation survives when cooperators sever ties with defectors and rewire toward other cooperators, leaving defectors isolated at the edge of the network. Yet in a human experiment, a learned network planner reached its best cooperation rates by doing the reverse: it placed defectors inside small cooperative neighborhoods so that they would come round. This proposal argues that both results follow from one variable. That variable is how much each agent's choices are driven by social pressure from the agents it is currently connected to, as opposed to its own payoff. Psychologists call this pressure the subjective norm, meaning the perceived expectation of the people around you.
The mechanism is simple. If social pressure is felt only from one's current neighbors, then cutting a tie changes the pressure the defector feels as well as its payoff. Ostracism takes a defector out of the cooperative neighborhood that could have converted it. It also tends to group defectors together on the periphery, where they reinforce each other's defection. We predict a crossover. Below some critical weight on social pressure, ostracism wins. Above it, "conciliatory" rewiring wins, meaning that defectors are introduced into cooperative neighborhoods rather than cut off. We propose to locate this crossover in simulation, explain it, and test whether it predicts which rewiring policy works for populations of language-model agents whose sensitivity to social pressure we measure directly.
Core Claim
The research file sets up a direct conflict. Cooperation Prevails When Individuals Adjust Their Social Ties shows that rewiring away from defectors, fast relative to strategy change, lowers the benefit-to-cost ratio cooperation needs from above 30 to 2. To Promote Full Cooperation in Social Dilemmas, Agents Need to Unlearn Loyalty shows that learning agents reach full cooperation once they adopt "Out-for-Tat", a policy of keeping cooperative partners and cutting defecting ones. Both papers motivate the ostracism side. On the other side, Scaffolding cooperation in human groups with deep reinforcement learning reports that its graph-neural-network planner beat both static networks and a rule-based clustering heuristic by placing defectors among cooperators. In all three, agents update on payoffs alone, or the humans' internal motives were not modeled. None has a term for social pressure that depends on neighborhood.
Our claim is that the sign of the best rewiring rule is set by $\alpha$, the weight an agent puts on local normative pressure relative to payoff. We expect a critical $\alpha^*$ that falls as rewiring speeds up, because faster rewiring forms defector enclaves faster. The second, more testable part of the claim is that this crossover does not appear if the same pressure is computed over the whole population instead of the current neighborhood. Without that topological dependence, the effect is not about networks at all.
Model
We keep the standard co-evolutionary setup of Cooperation Prevails When Individuals Adjust Their Social Ties: a Prisoner's Dilemma on a graph, a strategy update by pairwise comparison, and a separate structural update that runs $W$ times faster. The only change is to the strategy update. Agent $i$ copies neighbor $j$'s strategy with probability
$$ p(i \leftarrow j) = \frac{1}{1+\exp\left(-\beta\left[(1-\alpha)(\pi_j - \pi_i) + \alpha\,\kappa\,\big(f_i(s_j) - f_i(s_i)\big)\right]\right)}. $$
Here $\pi$ is the accumulated payoff. $f_i(s)$ is the fraction of $i$'s current neighbors playing strategy $s$. $\kappa$ puts that fraction on the payoff scale. $\beta$ is selection intensity. Setting $\alpha=0$ recovers the original model.
Because $f_i$ is computed over current edges, every rewiring decision changes the social pressure felt by both endpoints. That coupling is the only new ingredient. Strictly, this term is a descriptive norm (what neighbors do), standing in for the theory-of-planned-behavior subjective norm (what neighbors expect). We also run an injunctive variant in which pressure comes from neighbors' explicit approval signals, which can differ from what those neighbors actually play.
We compare four structural rules that differ only in how they treat a defecting neighbor. Ostracism follows Santos et al.: cut the tie and rewire to a friend of a friend. Out-for-Tat is learned with Q-learning as in Leung et al. Conciliatory introduction is a decentralized analogue of the planner's behavior: a cooperator keeps the tie and adds an edge from the defector to one of its own cooperative neighbors. Learned linking uses the Sarsa co-learning agents of Co-Learning of Strategy and Structure Achieves Full Cooperation in Complex Networks with Dynamical Linking, with the action set extended to keep, cut, or introduce.
Research Program
The first stage maps the phase diagram over $(\alpha, W, b/c)$. We use $N=1000$ agents on random regular graphs as in Leung et al., plus the $z=30$ networks of Santos et al. For each point we record four quantities: final cooperation, the minimum $b/c$ at which cooperation survives, time to fixation, and the number and size of defector enclaves. A defector enclave is a connected set of defectors whose members each have a defecting majority among their neighbors. The mechanism predicts that, at high $\alpha$, ostracism leaves a stable population of such enclaves and cooperation plateaus below full. Conciliatory introduction should not.
The second stage asks what an optimal decentralized policy looks like. We train the extended Sarsa agents across $\alpha$ and test whether "introduce" emerges as the learned response to defectors exactly where the phase diagram says it should. We also train a centralized planner in the style of the scaffolding paper. It is not a competitor here. It serves as an upper bound that shows how much the decentralized rules leave unrealized at each $\alpha$.
The third stage moves to language-model agents. We extend the simulator of Evolution of Cooperation in LLM Societies : A Multi-Lingual Examination in two ways. The hard-coded imitate-the-best rule is replaced by the model's own decision given its neighborhood's history. Rewiring is enabled. First, we measure each model's effective $\alpha$ with counterfactual probes: we vary the cooperation fraction among neighbors while holding the agent's payoff comparison fixed, and fit the equation above to the choices. We do this separately per language, because that paper reports systematic drift toward defection in Hindi. Mapping the fitted $\alpha$ onto the stage-one diagram yields a prediction of which rewiring rule should win for that population, and stage three tests that prediction. We drop the paper's cache mirroring and sample every interaction independently, since mirroring can introduce correlations between players.
Baselines
The naive reference point is a static graph with payoff imitation. In LLM societies, Evolution of Cooperation in LLM Societies : A Multi-Lingual Examination reports cooperation of 0.6–0.7 on English scale-free networks after 40 generations, with collapse on star and complete graphs. In humans, the scaffolding paper reports 42.8% cooperation on static networks. Every dynamic rule must clear the static baseline on the same graphs and payoffs.
The strongest decentralized competitor is ostracism. We run it in two forms: the Santos et al. rule, which lowers the surviving $b/c$ from above 30 to 2 at $z=30$, and Leung et al.'s learned Out-for-Tat, which reaches above 90% Out-for-Tat adoption at high $W$. Learned co-linking via Sarsa reports 100% cooperation at a timescale ratio of 1, against below 20% under random matching. We reimplement all of these in our own code, on identical graphs, seeds, and iteration budgets. A dynamic rule beats ostracism if, at matched $W$ and $b/c$, it gives higher final cooperation or survives at a lower $b/c$ across at least a contiguous band of $\alpha$, with non-overlapping 95% intervals over 50 seeds. At $\alpha=0$ our reimplementation must reproduce the published numbers first. If it does not, the comparison is not valid.
The strongest structural intervention overall is the centralized planner. The scaffolding paper reports 77.7% cooperation against 61.2% for rule-based cooperative clustering. We give the planner the same environment interactions as the Sarsa agents' combined training budget and treat it as a ceiling, not as something to beat.
The ablation that isolates the key ingredient is the global-pressure variant, in which $f_i$ is replaced by the population-wide fraction. If the crossover survives this replacement, our claim that topology matters is false. A second ablation replaces normative pressure with the intrinsic prosocial weight of Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI, matched so that cooperation is equally attractive on a complete graph.
Why not just…?
Why not just use conformist imitation on static graphs, which evolutionary game theory already studies? A static graph removes exactly the coupling we care about: on a fixed graph, cutting or introducing ties cannot change the pressure anyone feels. We run static conformity as a condition anyway. If it reproduces the ordering of rules we see under rewiring, the dynamic-graph story is unnecessary, and we will say so.
Why not just let a centralized planner pick ties, as the scaffolding paper did? The planner works, but it explains nothing about why rehabilitation beat clustering, and in decentralized agent systems there is often no global observer. Training the planner across $\alpha$ is in fact our tool for answering the why-question. The decentralized introduce rule is what agents could actually run themselves.
Why not just add a prosocial preference, which the mechanism-design paper argues is necessary? A fixed regard for others' payoff pushes toward cooperation everywhere. Local pressure pushes toward defection inside defector enclaves. That difference is why we expect ostracism to fail under norms but not under prosociality. The matched-prosociality ablation tests this directly. If the crossover also appears under prosociality, our mechanism is wrong.
Why not just give everyone a global norm or a vote, as in the well-mixed LLM norm studies? The Role of Social Learning and Collective Norm Formation in Fostering Cooperation in LLM Multi-Agent Systems shows that global voting sustains a commons, with survival dropping to 16.22 rounds without it. Cultural Evolution of Cooperation among LLM Agents uses global reputation in the same way. A global norm cannot form enclaves, so it should erase the crossover. This is a real threat to the relevance of the work: if global pressure performs as well, topology adds nothing. The global-pressure ablation decides it.
Decisive Checks
The proposal succeeds only if three things hold. First, the $\alpha=0$ reimplementation reproduces the Santos and Leung numbers, and a crossover $\alpha^*$ then appears, below which ostracism wins and above which introduction wins. Second, the crossover vanishes under global pressure and under matched prosociality, and enclave counts track the loss of cooperation under ostracism. Third, for language-model populations, the probe-measured $\alpha$ correctly predicts the winning rewiring rule in more model-and-language combinations than chance. Here the English and Hindi populations of the multilingual simulator serve as a natural contrast, provided that their measured $\alpha$ values differ.
A secondary check repeats stage one under Snowdrift and Stag Hunt payoffs. Facilitating cooperation in human-agent hybrid populations through autonomous agents found that defectors can raise human cooperation in Snowdrift, so the crossover may shift or reverse sign there, and we report whichever happens.
If ostracism dominates at every $\alpha$, the conflict between the two literatures has some other cause. Likely candidates are human-specific features absent from the model, such as the planner's ability to choose which cooperators a defector meets.
References
- Cooperation Prevails When Individuals Adjust Their Social Ties
- To Promote Full Cooperation in Social Dilemmas, Agents Need to Unlearn Loyalty
- Scaffolding cooperation in human groups with deep reinforcement learning
- Co-Learning of Strategy and Structure Achieves Full Cooperation in Complex Networks with Dynamical Linking
- Evolution of Cooperation in LLM Societies : A Multi-Lingual Examination
- Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI
- The Role of Social Learning and Collective Norm Formation in Fostering Cooperation in LLM Multi-Agent Systems
- Cultural Evolution of Cooperation among LLM Agents
- Facilitating Cooperation in Human-Agent Hybrid Populations through Autonomous Agents