Is Prosociality a Preference or a Learning Signal in Threshold Dilemmas?

Many shared-safety problems only pay off once enough people chip in. A disaster is averted only if total contributions cross a cutoff, and a regional hospital network gets a safety benefit only if enough sites adopt an expensive platform. Reinforcement-learning agents are bad at these "threshold" games when their wealth is unequal. In a collective-risk game, players pay into a pot, and if the pot misses a target $T$ everyone loses their remaining wealth with probability $r$. In that game, independent learners fall short of the game's own equilibrium for every risk level $r \ge 0.3$. The reason is that poor agents rarely change the outcome, so they get almost no learning feedback about it (Cooperation between Independent Reinforcement Learners under Wealth Inequality and Collective Risks). A separate literature gives agents a "social value orientation": a weight, written as an angle, that mixes the agent's own reward with its peers' rewards. That literature reports that prosocial agents close cooperation gaps that contracts cannot (Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI).

These findings are easy to combine carelessly. One could add prosocial weights to the unequal-wealth threshold game, report that cooperation goes up, and stop. That result would be ambiguous. Prosocial weighting could help because it changes what agents want, or only because it gives poor agents a denser learning signal. A selfish credit-assignment fix could do the second job just as well. This note proposes an experiment that separates the two explanations. It also tests one design that only makes sense if both explanations are partly true: prosociality that switches on only when an agent is likely to be pivotal, meaning near enough to the threshold that its own contribution could decide the outcome.

Core Claim

We claim that the effect of social preferences in unequal threshold dilemmas splits cleanly by risk regime.

At high collective risk, selfish equilibria already reach the threshold. The class-based equilibrium has rich players contributing about 100% of the time and poor players about 40%. The failure there is in learning. A selfish learner given a counterfactual "difference reward" should therefore match prosocial learners. A difference reward pays an agent the change in its own outcome caused by its contribution.

At low risk, selfish equilibria abandon the threshold. Only a change in preferences, such as a prosocial weight or an external mechanism, can recover it.

The second claim is that a gated prosocial weight, which scales with the agent's estimated chance of being pivotal, can match the threshold success of a uniformly prosocial agent at the same average prosocial "budget". It should lose less to exploiters, because it stops paying when contributions are futile (far below the threshold) or redundant (far above it).

We also name the failure mode we expect this design to create. A competitive agent can hold back deliberately, push the group to the brink, and make the gated agent pivotal. We call this brinkmanship, and testing for it is part of the proposal.

Research Program

We reimplement the collective-risk game from the unequal-wealth study and its follow-up on diverse individual risk (Cooperation and Learning Dynamics under Wealth Inequality and Diversity in Individual Risk). We keep their log-utility, endowment classes, threshold $T$, risk $r$, and Roth-Erev learners. Roth-Erev learners update action propensities in proportion to the payoffs they receive. Keeping these choices means our selfish baseline reproduces their numbers before anything else changes. We add a policy-gradient learner as a second learning rule, so conclusions do not depend on one algorithm.

Agents learn on a remapped reward, $\hat r_i = r_i\cos\theta_i + \bar r_{-i}\sin\theta_i$. Here $r_i$ is agent $i$'s material payoff, $\bar r_{-i}$ is the mean payoff of the others, and $\theta_i$ is its social value angle. This follows Heterogeneous Social Value Orientation Leads to Meaningful Diversity in Sequential Social Dilemmas, which used the remapping for policy diversity in gridworlds with no threshold cliffs. The gated variant sets

$$\theta_i(s) = \theta_{\max}\, g\big(\hat p_i(s)\big),$$

where $\hat p_i(s)$ is agent $i$'s running estimate of the probability that its contribution flips the threshold outcome, computed from recent rounds' contribution totals. The function $g$ maps that probability to $[0,1]$ and is monotone. To compare gated and static agents fairly, we choose $\theta_{\max}$ so that the episode-averaged $\sin\theta_i$ equals that of a static angle. Any gain from gating is then a gain from where prosociality is spent, not how much of it there is.

The cost-safety literature supplies one structural axis and nothing more. Clinical economics separates sunk capital costs from per-use consumables, and consumables can eat most of a fixed fee: 92.7% for laparoscopic versus 58.9% for open gastrectomy (Assessment of the Cost Performance of Laparoscopy-Assisted Gastrectomy). Capital barriers keep safer platforms in well-funded centers (Robotic Retrograde Intrarenal Surgery: A Journey from “Back to the Future”). We therefore let a fraction $\kappa$ of each agent's contribution be paid up front, irreversibly, before the threshold is resolved. The remainder is paid per round.

This axis matters because the main mechanical defense against the "sucker's payoff" behaves differently across $\kappa$. Refunding contributions when the threshold is missed protects consumables but cannot refund capital that is already sunk. We use the clinical numbers only to place $\kappa$ in a plausible range. We do not claim to model hospitals.

As a transfer check, we port the threshold and $\kappa$ structure into GovSimContract inside Flecart/prosocial-agents. There we give large language model (LLM) agents static or pivotality-gated prosocial instructions. That paper fits frontier models' default prosocial weight at about zero. The reported equality paradox in Evaluating Cooperation in LLM Social Groups through Elected Leadership, where prosocial leaders raised welfare but also raised the Gini index, gives a specific hypothesis to test: prosocial agents end up carrying the cost of avoiding the threshold disaster.

Baselines

The naive reference is selfish Roth-Erev learners on the unmodified game. The cited paper reports cooperation rates of 60–80% for rich agents and 40–50% for poor agents under learning, against roughly 100% and 40% at the class-based Nash equilibrium. It also reports that learned performance falls below the equilibrium for $r \ge 0.3$ (Cooperation between Independent Reinforcement Learners under Wealth Inequality and Collective Risks). We first reproduce these numbers. If we cannot, the comparisons that follow are not trustworthy. We also report the equilibrium itself as an analytic upper reference for selfish play.

The strongest learning baseline is a selfish learner trained on a difference reward. This targets exactly the poor-agent credit problem that paper diagnosed. We know of no reported result on Lacuna for difference rewards in collective-risk threshold games, so this baseline is new to the comparison and is the main threat to our story.

The strongest preference baseline is a static prosocial weight. We use the $\lambda$-cooperation utility of Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI, which reports a 60–68 percentage point efficiency gap left by contracts in its matrix games and +119.6 units of gain from prosociality in GovSimContract. We also use static social value angles at prosocial (45°) and altruistic settings.

The strongest mechanism baseline is a refund-if-missed rule combined with a minimum-participation quorum, in which a venture aborts before any costs are sunk unless enough agents commit (The Anh Han et al.). We found no reported result on Lacuna for quorums in unequal-wealth collective-risk games with RL agents.

Every method gets the same environments, seeds (at least 30 per cell), episode budget, and hyperparameter search budget. Hyperparameters are tuned on a held-out set of $(T, r)$ values. The key ablation is gated versus static prosociality at matched average $\sin\theta$. A second ablation feeds the pivotality estimate to a selfish agent as an extra input, without changing its utility. This shows whether gated prosociality works through information or through preference.

"Beating" a baseline has three meanings, one per claim. At high risk, the prosocial agents must fail to beat the difference-reward baseline for the learning-signal claim to hold. At low risk, they must beat it by a margin outside the seed confidence interval for the preference claim to hold. For the gating claim, gated agents must reach the static agents' threshold success with lower losses against competitive partners, again outside seed confidence intervals.

Why not just…?

Why not just give everyone a fixed prosocial weight? That baseline is in the design. The literature already shows it closes gaps, and it is also exposed to the sucker's payoff that both Mechanism Design Is Not Enough and Heterogeneous Social Value Orientation flag. Our gated-versus-static comparison at matched budget asks whether placing prosociality only at pivotal states keeps its benefits while reducing exploitation. If static agents lose no more than gated ones against competitive partners, gating is unnecessary, and we will report that.

Why not just fix credit assignment with a difference reward and skip social preferences entirely? This is a real threat, and we expect it to win at high risk. Ivanov et al.'s diagnosis is a weak learning signal, and a difference reward attacks that directly. Our claim is that it cannot work at low risk, where the selfish equilibrium abandons the threshold, because a correct selfish signal points toward abandonment. The risk sweep settles this. If difference rewards match prosocial agents at every $r$, our preference claim is false and the paper becomes a negative result about social value orientation.

Why not just refund missed contributions or require a quorum? A refund rule removes the sucker's payoff for per-round costs. A quorum taken before commitment is the right defense when costs are sunk, and it may dominate everything we propose. The $\kappa$ sweep is designed to show where each mechanism breaks. We expect refunds to weaken as $\kappa$ grows, and quorums to stall when poor agents cannot afford the capital share. Mechanisms also need a party to enforce them. The equality-paradox study reports that elected policies worked only as soft guidance, with no enforcement (Evaluating Cooperation in LLM Social Groups through Elected Leadership). If quorums dominate across all $\kappa$ in our runs, we will say so.

Proposed Experiments

The first decisive check is the risk sweep over $r \in [0, 1]$ under unequal wealth, plotting threshold success and total welfare against $r$ for the selfish, difference-reward, static-prosocial, and gated learners. The prediction is a crossover. Difference-reward and prosocial curves should coincide above the risk level where the selfish equilibrium begins to reach the threshold, and separate below it. Whether that crossover exists, and where, answers the title question.

The second check uses mixed populations with 25–50% competitive agents (negative angle). We measure the prosocial agents' material losses, the Gini index of final wealth, and how often brinkmanship appears. We define brinkmanship as competitive agents withholding until the gated agents' pivotality estimate peaks.

The third check sweeps the up-front cost fraction $\kappa$ against the refund and quorum baselines.

As a sanity check for the reward remapping, we confirm the two-player Stag Hunt result from Melting Pot 2.0. Stag Hunt is a threshold of two, and the cited work reports a normalized 0.876 for the best response trained on diverse partners versus 0.850 for the selfish baseline (Heterogeneous Social Value Orientation Leads to Meaningful Diversity in Sequential Social Dilemmas).

Finally, the GovSimContract transfer runs eight-agent groups under static versus gated prosocial instructions. It measures survival, welfare, and Gini, and checks whether the equality paradox tracks which agents paid the safety costs.