Rules that keep a shared resource from collapse, such as quotas, sanctions, or elected leaders, are now routinely designed and tested inside simulations populated by artificial agents. Those simulated agents quietly fix one behavioral fact: how noisy their choices are. We propose to measure that noise instead of assuming it. The model we use is quantal response, in which an agent picks better options more often than worse ones but still errs, with a single "rationality" dial $\lambda$ controlling how often. Measuring $\lambda$ matters because in commons problems noise is not a nuisance term. Early, clumsy learners preserve a resource by accident, and competent ones destroy it (A multi-agent reinforcement learning model of common-pool resource appropriation). A governance rule tuned against a simulated population with the wrong noise level can therefore succeed for the wrong reason, and then fail when it meets agents who are more or less careful than the simulator assumed.
The proposal has three parts. First, we fit a quantal response model, with a separate $\lambda$ for each agent, to a small amount of observed harvesting behavior. Second, we build a cheap simulator from that fitted population. Third, we search for governance rules in the simulator and test whether they transfer to real agent populations. The claim is specific. A fitted, heterogeneous rationality distribution should rank candidate governance rules in the same order the real population does, and should do so better than three alternatives: a fixed noise level, a black-box behavioral clone trained on the same data, or blind randomization over noise. Rule search against a fixed noise level should also converge prematurely onto rules that only work because simulated agents make mistakes.
Core Claim
Two facts in the file make this testable now rather than hypothetical. The first is that bounded rationality makes behavior informative. Mistakes reveal how large utility gaps are, so a structural choice model can be identified from far less data than a deterministic one. This is shown for Stackelberg followers (Wu et al.) and for Boltzmann-rational demonstrators in inverse reinforcement learning (Skalse et al.). The second is that the noise of large language model (LLM) agents can be set by hand. Decision Aggregation under Quantal Response reports that raising GPT-4o-mini's sampling temperature from 0 to 1 moves its fitted $\lambda$ from effectively infinite to 8.93. Together these give us an off-simulator "reality" whose rationality we can vary on purpose: LLM agent societies in the Governance of the Commons Simulation (GovSim). That lets us ask whether a calibrated noise model predicts transfer.
We model agent $i$'s choice of harvest $h$ at stock $s$ with the logit rule
$$P_i(h \mid s) = \frac{\exp(\lambda_i\, U(h; s, w_i))}{\sum_{h'} \exp(\lambda_i\, U(h'; s, w_i))}, \qquad U(h; s, w) = r(h, s) + w\, V(s - h),$$
Here $r$ is the immediate catch, $V$ is the continuation value of the remaining stock under the environment's regrowth rule, $w_i$ is the agent's farsightedness, and $\lambda_i$ is its rationality. We place a hierarchical prior over $(\lambda_i, w_i)$ so that populations can be mixed.
The farsightedness term is not decoration. Non-Myopic Negotiators See What's Best shows that long horizons alone can produce restraint, so a model with $\lambda$ but no $w$ would mistake farsighted restraint for noise.
The known limitations of quantal response constrain the design. The file lists i.i.d. errors, a single shared $\lambda$ across agents, and brittleness when an agent deliberately changes how rational it behaves (A Complete Analysis on the Risk of Using Quantal Response). We address the first two by fitting per-agent parameters. We address the third by including a deliberate greedy newcomer in every evaluation.
Research Program
Stage one: calibration. We run GovSim Fishery and Pasture (Cooperate or Collapse) with five-agent groups. The groups include mixed-model and mixed-temperature populations, at temperatures of 0, 0.3, 0.7, and 1.0. We log each monthly harvest decision. We then fit the hierarchical logit model to a deliberately small budget of logged months, and hold out other models and temperatures entirely.
Stage two: the simulator and a sharp prediction. The fitted population defines a simulator in which each rollout costs milliseconds instead of dozens of LLM calls. Before any governance is added, we use the simulator to predict how the chance of survival depends on the population's rationality. That predicted curve is then checked against real GovSim runs across temperatures.
The prediction could turn out non-monotone. Leibo et al. saw exactly that shape along a learning trajectory: sustainable while naive, collapsed once competent. If the simulator reproduces a shape that simpler models miss, that is direct evidence the calibration captures something real. A useful side result is theoretical. A Unified Approach to Reinforcement Learning, Quantal Response Equilibria, and Two-Player Zero-Sum Games connects entropy-regularized learners to quantal response equilibria. If that connection carries over, the naive-to-tragedy transition becomes a critical value of $\lambda$ that can be read off the fitted model, rather than a curiosity of one grid world.
Stage three: governance search and transfer. The governance mechanism is a small, interpretable rule family injected into GovSim as a binding planner. It has three components: a per-agent quota as a fraction of the sustainable harvest, a trigger for sanctions, and a sanction length that withholds an agent's share for $k$ months. The sanction component mirrors the temporary sanctions discovered by Deep reinforcement learning can promote sustainable human behaviour in a common-pool resource problem.
We search this family with a population-based optimizer inside each candidate simulator, then evaluate the resulting rules in real GovSim runs. The premature-convergence question is asked directly during this search. We track the diversity of the rule population and check which rules the search collapses onto. We expect that a simulator with a single fixed $\lambda$ will reward lax quotas whenever its agents' errors happen to under-harvest. We also expect that ranking candidates over short windows will reward rules whose collapse comes after the window closes. This is the same greediness that Multiple-Frequencies Population-Based Training documents for hyperparameters, but in a regime where the delayed cost is built into the problem.
Baselines
The naive reference point is ungoverned GovSim with no planner, which fixes the collapse rate each population shows on its own. For context, the strongest published figure for that setting is GPT-4o's 53.3% survival across scenarios (Cooperate or Collapse). We re-measure all numbers under our own populations rather than borrowing them.
The strongest existing governance method in this benchmark family is elected leadership. Evaluating Cooperation in LLM Social Groups through Elected Leadership reports that elected leaders raise social welfare by 55.4% and survival time by 128.6% over leaderless groups, and that the election itself adds 16.7% welfare over an appointed leader. It is strong because it adapts to whichever agents are present without any model of them. We run it on the identical populations, including the greedy newcomer, with the same LLM call budget. Our method counts as beating it only if calibrated-simulator rules match or exceed its survival time, reach higher welfare or a lower Gini coefficient (a standard 0-to-1 inequality measure), and hold up better against the newcomer. That paper notes that elected leadership raises inequality, which gives us a second axis on which to win.
We also compare against universalization prompting ("what if everyone acted this way?"), reported to extend average survival by four months (Cooperate or Collapse), run under the same budget.
The most important baseline is the Koster et al. pipeline itself. It uses a recurrent behavioral clone trained on the same logged months, with the same rule search run inside it. This is the closest existing sim-to-real design, and on commons governance it reports roughly 150% more surplus than proportional allocation and 65% of games completed versus 5% under equal allocation. Those numbers come from a different game with human participants, so they motivate the comparison rather than setting a target. We give the clone and the quantal response model exactly the same data, and give both simulators the same rule-search budget.
Two ablations isolate our key ingredient, the fitted heterogeneous $\lambda$. One replaces it with a single pooled $\lambda$. The other replaces it with uniform randomization over a wide $\lambda$ range, with no fitting at all.
Why not just…?
Why not just randomize the noise level widely instead of fitting it? This is the most serious threat. Training against the whole range of rationality levels might yield rules that are robust everywhere. We expect wide randomization to be safe but over-conservative. It must guard against implausibly greedy populations, so it will pick tight quotas and pay in welfare. The uniform-randomization ablation settles this directly. If it matches fitted calibration on both survival and welfare in held-out populations, then fitting adds nothing, and we will report the result as evidence that robustness, not calibration, is what matters.
Why not just clone behavior with a neural network, as Koster et al. did? Cloning is the proven route to real-world transfer (Deep reinforcement learning can promote sustainable human behaviour). Its weakness is that it has no knob for extrapolation. A clone trained at one temperature or model mix has no principled way to represent a more careful or more careless population, whereas the structural model changes one parameter. The held-out-temperature and held-out-model tests measure exactly this. We also sweep the amount of logged data, because the identifiability results suggest the structural model should win when data is scarce and may lose when data is plentiful.
Why not just search governance rules directly in GovSim with real LLM agents? Direct search removes the reality gap by construction. However, each twelve-month GovSim episode requires many LLM calls, and population-based search needs thousands of episodes. We include a compute-matched direct-search arm that spends the same LLM budget on real episodes instead of on calibration data. If it wins, a simulator is not worth building at this scale, and that is itself a useful finding.
Why not just add an entropy bonus or diversity pressure to the rule search? Generic diversity pressure keeps the rule population spread out, but it cannot tell the search which region of rule space is actually safe. If the simulator's agents are wrong about noise, a diverse search simply explores a wrong landscape more thoroughly. We run the fixed-$\lambda$ simulator with and without an explicit diversity operator. We predict that diversity alone does not repair transfer, while calibration does.
Proposed Experiments
The decisive check is a ranking test. We generate about fifty governance rules spanning the rule family, score each in every candidate simulator, and score each again in real held-out GovSim populations. We then report the rank correlation between simulated and real survival, and between simulated and real welfare. The fitted heterogeneous model should give the highest correlation, especially for populations whose temperature or model was never seen in calibration.
The second check is the collapse curve. We compare predicted and observed survival across temperatures 0 to 1, scored with the Brier score (the mean squared error of predicted survival probabilities) against real survival outcomes.
The third check is end-to-end governance. The best rule from each simulator is deployed in real GovSim Fishery and Pasture, with and without a greedy newcomer. We report survival rate, months survived, total harvest, and Gini coefficient against elected leadership, universalization, and the no-planner baseline.
Diversity diagnostics run alongside these checks. We track how spread out the rule population is during search, and compare ranking over three-month versus twelve-month windows, to test whether fixed-noise simulators drive the search onto lax rules early.
As a secondary substrate, Melting Pot's Commons Harvest (Melting Pot 2.0) lets us fit $\lambda$ to trained reinforcement learning agents at successive checkpoints. There we test whether the naive-to-tragedy transition occurs near the predicted critical value. We found no reported scalar result on Lacuna for baseline returns on Commons Harvest, so that comparison is internal.
We are explicit that LLM societies are a stand-in for reality and not people. A small human study in the Fishery game, testing whether human harvesters fall inside the fitted $\lambda$ range, is the natural next step once the method passes these checks.
References
- A multi-agent reinforcement learning model of common-pool resource appropriation
- Inverse Game Theory for Stackelberg Games: the Blessing of Bounded Rationality
- Misspecification in Inverse Reinforcement Learning
- Decision Aggregation under Quantal Response
- Non-Myopic Negotiators See What's Best
- A Complete Analysis on the Risk of Using Quantal Response: When Attacker Maliciously Changes Behavior under Uncertainty
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents
- A Unified Approach to Reinforcement Learning, Quantal Response Equilibria, and Two-Player Zero-Sum Games
- Deep reinforcement learning can promote sustainable human behaviour in a common-pool resource problem
- Multiple-Frequencies Population-Based Training
- Evaluating Cooperation in LLM Social Groups through Elected Leadership
- Melting Pot 2.0