Multi-Agent Algorithmic Recourse

• Multi-agent algorithmic recourse is a framework that redefines individualized recourse by incorporating strategic interdependencies, ethical constraints, and quota limitations.
• It employs methodologies such as agent-based simulations, structural causal models, and reinforcement learning to analyze temporal dynamics and reliability.
• The approach has practical implications for enhancing fairness, reducing wasted effort, and optimizing system-level welfare in competitive, retrained environments.

Multi-agent algorithmic recourse studies recourse when a recommendation issued to one decision subject is not independent of the rest of the system. In the standard single-agent formulation, an individual with features $x$ or $x^F$ seeks a minimally costly change that flips an unfavorable model output to a favorable one. The multi-agent extension begins when this independence assumption fails: outcomes may be strategically interdependent, the favorable outcome may be quota-limited and rank-based, multiple rejected individuals may act on recourse simultaneously, providers may be heterogeneous and capacity-constrained, and model retraining may shift the very standard that recourse was meant to satisfy. Across recent work, the topic has been formalized through ethical constraints on structural counterfactuals, agent-based simulations with moving thresholds, population-coupled reinforcement learning, capacitated bipartite matching, and repeated deployment–response–retraining loops under scarcity (O'Brien et al., 2021, Fonseca et al., 2023, Ceccon et al., 26 Sep 2025, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).

1. From individualized recourse to coupled environments

The classical recourse problem is written as a minimum-cost modification subject to a favorable outcome constraint. One formulation asks for a minimally costly $x'$ such that $f(x')=1$, and another writes the same idea as a minimum-cost action $\delta$ such that $h(x^F+\delta)$ exceeds a target threshold $t$. In that setting, recourse is effectively interpreted as “what is the cheapest change I can make to get a better outcome?” (O'Brien et al., 2021, Fonseca et al., 2023).

The multi-agent critique begins with the assumptions embedded in that formulation. One paper states that standard recourse typically assumes that only one agent matters at a time, that an agent’s utility depends only on their own action and resulting prediction, that the recommendation does not materially alter other agents’ outcomes, and that the ethical objective is fully captured by improving the principal agent’s prediction at minimum cost. A later temporal paper adds a distinct assumption: the environment is effectively static, so a recommendation issued at time $t$ is presumed redeemable at time $t+\delta$ without meaningful threshold movement, competition, or population turnover (O'Brien et al., 2021, Fonseca et al., 2023).

Relaxing those assumptions produces several non-equivalent multi-agent settings. In one line of work, strategic dependence is explicit: one person’s action can change another person’s payoff, as in the prisoner’s dilemma or route recommendations in traffic networks. In another, interaction is mediated through a top-$k$ ranking rule: success depends on beating a moving threshold induced by other applicants. In a third, recourse is many-to-many, with multiple seekers, multiple provider-specific classifiers, and provider capacities. In a fourth, competition and retraining generate endogenous distribution shift over repeated rounds. This suggests that multi-agent algorithmic recourse is best understood as a family of recourse problems with externalities rather than a single canonical formalism (O'Brien et al., 2021, Fonseca et al., 2023, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).

2. Ethical criteria, structural counterfactuals, and strategic dependence

The most explicit normative formulation defines an “ethically ideal” recommendation as one that improves the prediction of the principal agent, does not worsen the prediction for any other agent, and increases the sum of predictions over all agents. These desiderata are encoded through constraints inspired by Pareto efficiency and social welfare efficiency. For $x^F$0 agents, Pareto non-harm is expressed as

$x^F$1

while social welfare improvement is

$x^F$2

The resulting framework augments the usual minimum-cost recourse objective with group-level constraints, yielding single-agent efficient, social-welfare-efficient, Pareto-efficient, and combined formulations (O'Brien et al., 2021).

That framework is causal rather than purely observational. Recourse actions are modeled as interventions in a structural causal model $x^F$3, with counterfactual state computed under no hidden confounders and invertible $x^F$4 by

$x^F$5

The point of this SCM layer is not only to ensure structural validity, but also to propagate the effect of one agent’s action to all affected agents. The paper is explicit that assuming access to such a causal model is a major limitation, because learning causal graphs from data remains difficult (O'Brien et al., 2021).

The constructive counterexample is the prisoner’s dilemma. With actions Betray $x^F$6 and Silent $x^F$7, and payoffs $x^F$8, $x^F$9, $x'$0, and $x'$1, single-agent recourse for agent 1 at $x'$2 recommends $x'$3, moving to $x'$4 and improving $x'$5 from $x'$6 to $x'$7. Yet agent 2’s outcome falls from $x'$8 to $x'$9, and total payoff falls from $f(x')=1$0 to $f(x')=1$1. From $f(x')=1$2, the same action improves agent 1 and reduces agent 2, but aggregate welfare rises. The example therefore exhibits a real tension among principal-agent improvement, Pareto non-harm, and aggregate welfare (O'Brien et al., 2021).

A recurring misconception is that this literature already solves recourse through equilibrium analysis. The game-theoretic work is explicitly “game theory inspired,” but it does not define Nash equilibrium, subgame perfect equilibrium, correlated equilibrium, or best-response dynamics. Its contribution is the importation of efficiency criteria into recourse optimization, not an equilibrium solution concept (O'Brien et al., 2021).

3. Competition over time, moving thresholds, and recourse reliability

A second line of work argues that time is intrinsic to recourse itself. The standard one-shot formulation presumes that if an individual is told what to change after an unfavorable decision, then acting on that recommendation later should remain sufficient. The temporal critique rejects this premise in capacity-limited settings. Instead of a binary classifier boundary, the environment is score-based and rank-based: a black-box model outputs $f(x')=1$3, agents are ranked, and only the top $f(x')=1$4 receive the favorable outcome. If the current top-$f(x')=1$5 threshold is the $f(x')=1$6-th highest score $f(x')=1$7, then multi-agent recourse seeks a minimally costly $f(x')=1$8 such that $f(x')=1$9. The threshold is therefore endogenous, since it depends on who else acts, who else enters, and how many slots exist (Fonseca et al., 2023).

The agent-based simulation framework makes the temporal coupling explicit. At time $\delta$0, the population is $\delta$1, consisting of previously unsuccessful agents plus new entrants. Agents are scored, the top $\delta$2 are selected, winners exit, losers receive recourse $\delta$3, and losers transition to the next round according to an action model $\delta$4 that captures adaptation and effort. The framework distinguishes binary versus continuous adaptation and constant versus flexible effort. It also introduces a system-level reliability metric. Let

$\delta$5

the agents who have reached or exceeded the previous threshold. If $\delta$6 denotes the actual winners at time $\delta$7, then

$\delta$8

This is the proportion of “expectant” recourse-takers who are actually rewarded. A related threshold-stability condition is

$\delta$9

When this equality fails, the threshold may move, and recourse based on $h(x^F+\delta)$0 becomes unreliable (Fonseca et al., 2023).

The central empirical claim is that only a small set of specific parameterizations yields reliable recourse over time. In the binary adaptation with constant effort regime, $h(x^F+\delta)$1 becomes close to zero when $h(x^F+\delta)$2 is around $h(x^F+\delta)$3 or higher and the number of new agents is around $h(x^F+\delta)$4 or higher. In continuous adaptation regimes, threshold and reliability become more volatile. The discussion adds an important asymmetry: if the threshold increases, then agents who did what was recommended may fail anyway; if the threshold decreases, recourse may remain “reliable” but the effort may have been excessive. Reliability failure is therefore not limited to broken promises; it also includes wasted effort (Fonseca et al., 2023).

A related but more prescriptive framework introduces durable recourse over a predefined time horizon $h(x^F+\delta)$5. Here the environment is again quota-limited: at each step $h(x^F+\delta)$6 candidates are accepted, $h(x^F+\delta)$7 new candidates enter, some rejected candidates leave, and others implement recommendations stochastically and reapply after delays. The problem is formalized as a POMDP, and the learned policy is hierarchical: a predictor policy $h(x^F+\delta)$8 selects a target score $h(x^F+\delta)$9, and a recommender policy $t$0 outputs individualized counterfactuals aiming at that score. Durability is evaluated through horizon-aware reliability and feasibility,

$t$1

The method is not a decentralized multi-agent game; it is a centralized RL controller over a population of adaptive candidates. Experimentally, it achieves a better Pareto frontier between reliability and feasibility than standard baselines and hybrids. For $t$2 at $t$3, the reported feasibility values are $t$4 and $t$5 for the learned method at $t$6 and $t$7, compared with $t$8 for Ustun, $t$9 for Wachter, and $t$0 for DiCE (Ceccon et al., 26 Sep 2025).

4. Many-to-many recourse, provider heterogeneity, and welfare gaps

A different generalization treats recourse as a system with multiple recourse seekers and multiple recourse providers. Seekers are $t$1, providers are $t$2, and each provider $t$3 has a binary classifier $t$4 and capacity $t$5. For each seeker–provider pair, the recourse cost is

$t$6

These costs are transformed into weights

$t$7

so lower recourse cost corresponds to higher welfare weight (Khotanlou et al., 14 Aug 2025).

The key distinction is between individual welfare and social welfare. If capacities are ignored, seeker $t$8 would choose the best edge $t$9, yielding

$t+\delta$0

Under capacities, assignments are binary variables $t+\delta$1, and social welfare is the optimum of the capacitated weighted bipartite matching problem

$t+\delta$2

subject to

$t+\delta$3

The welfare gap is

$t+\delta$4

This quantity measures the price of treating recourse as independent when provider capacities make many individually optimal recommendations jointly infeasible (Khotanlou et al., 14 Aug 2025).

The framework proceeds in three layers. Layer 1 solves the basic capacitated matching problem for fixed capacities. Layer 2 redistributes capacity while holding total capacity $t+\delta$5 fixed, thereby minimizing the welfare gap. Layer 3 introduces adjustment penalties $t+\delta$6, where $t+\delta$7, to trade off welfare gain against the cost of changing capacities. The second layer is accompanied by a direct capacity-allocation rule: compute each seeker’s preferred provider $t+\delta$8, keep the top-$t+\delta$9 seekers ranked by $k$0, and assign capacity by

$k$1

The reported complexity is $k$2 to compute all $k$3 and $k$4 for sorting (Khotanlou et al., 14 Aug 2025).

Empirically, the framework is evaluated on Two-Moon, COMPAS, and Credit using four provider models—logistic regression, multilayer perceptron, decision tree, and random forest—and pairwise recourse costs computed by MACE. For $k$5 costs, arbitrary initial capacities produce social welfare equal to $k$6, $k$7, and $k$8 of the individual-welfare benchmark on Two-Moon, Credit, and COMPAS, respectively. Optimal redistribution reaches $k$9 of individual welfare on all three datasets. Penalized redistribution remains near-optimal at $x^F$00, $x^F$01, and $x^F$02. The appendix reports the same qualitative pattern for $x^F$03, with $x^F$04, $x^F$05, and $x^F$06 (Khotanlou et al., 14 Aug 2025).

5. Repeated retraining, escalating boundaries, and inequitable recourse

Another branch of the literature studies multi-agent recourse under endogenous model evolution. The interaction is round-based. At round $x^F$07, a dataset $x^F$08 is sampled, a subset of rejected users strategically responds to the deployed model $x^F$09 via recourse, the modified dataset $x^F$10 is scored, a constrained labeling rule $x^F$11 accepts at most $x^F$12 points, and the next model $x^F$13 is retrained on $x^F$14. The recourse step is written either as the hard-constrained problem

$x^F$15

or in relaxed form as

$x^F$16

Competition is introduced through the Top-$x^F$17 rule,

$x^F$18

so acceptance is a function of the entire score vector rather than only the individual’s modified score (Yang et al., 12 Mar 2025).

The main analytical claim is Theorem 1, an increasing decision boundary condition. In the setting of logistic regression, linear cost, and cross-entropy update, increasing the decision boundary decreases the loss if and only if the pseudo-label of the $x^F$19-th sample in $x^F$20 is $x^F$21. Intuitively, if recourse creates too many pseudo-positive points relative to capacity, the retrained model is pressured to become stricter. The paper uses this argument to explain the first of three recurring phenomena: escalating decision boundaries (Yang et al., 12 Mar 2025).

The second phenomenon is non-robust prediction or model collapse. The reported metrics include Test-Acceptance Rate (TAR), Short-Term Accuracy (STA), Model Shift (MS), Fail to Recourse (FTR), and Ratio of Effort (RoE). Under Top-$x^F$22 plus standard retraining, TAR falls close to zero after a few rounds across the synthetic, UCI DefaultCredit, and Credit datasets, indicating that the evolved model has become much stricter relative to the original distribution. STA degrades over rounds, and the accepted and rejected distributions become increasingly mixed. The third phenomenon is inequitable recourse actions: newcomers face higher cost and worse reliability than incumbents, with newcomers’ cost reported as about $x^F$23 that of existing users in logistic regression on synthetic data (Yang et al., 12 Mar 2025).

Two mitigation strategies are proposed. Fair-top-$x^F$24 replaces deterministic score-only selection with a KDE-biased weight

$x^F$25

so that accepted points are selected according to a diversity-aware distribution rather than a single crowded metric. Dynamic Continual Learning modifies a Synaptic Intelligence regularizer by making its strength depend on the Jensen-Shannon divergence between positive and negative data distributions in the previous round,

$x^F$26

The reported combination of Fair-top-$x^F$27 and DCL performs best: TAR remains much higher than in Top-$x^F$28 baselines, STA stays near $x^F$29 in long-term rounds, and model shift remains controlled. Fair-top-$x^F$30 alone keeps TAR around $x^F$31 and substantially reduces RoE, but FTR is around $x^F$32–$x^F$33, which the paper attributes partly to a mismatch between the recourse generator and the KDE-based selection rule (Yang et al., 12 Mar 2025).

6. Conceptual status, limitations, and open problems

A notable feature of the literature is that “multi-agent” does not imply a single methodological commitment. The game-theoretic efficiency paper borrows Pareto and social-welfare criteria without solving for Nash equilibrium or related concepts. The temporal simulation framework is population-level and competition-mediated, not a formal strategic game; it explicitly positions itself relative to game-theoretic multi-agent recourse papers, especially O’Brien and Kim. The durable recourse paper formulates a centralized POMDP and learns a recourse policy with Soft Actor-Critic, but applicants are not utility-maximizing best responders. The welfare-gap paper assumes a centralized planner, passive providers in the objective, and static one-shot capacities. The model-evolution paper captures strategic dependence through ranking and retraining rather than through explicit equilibrium analysis (O'Brien et al., 2021, Fonseca et al., 2023, Ceccon et al., 26 Sep 2025, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).

The limitations are correspondingly heterogeneous. Ethical-constraint formulations rely on access to an accurate SCM and acknowledge that causal structure learning is difficult. Simulation-based temporal work relies on synthetic populations and simplified adaptation functions, while noting that human behavior is complex and not always rational. RL-based durable recourse is entirely simulation-based, assumes a fixed scoring model, and omits causal structure. Capacitated matching presumes that pairwise recourse costs $x^F$34 are known for every seeker–provider pair and that capacity redesign is centrally implementable. Repeated-retraining models rely on pseudo-labeling under scarcity, simplified user response, and only partial formal theory (O'Brien et al., 2021, Fonseca et al., 2023, Ceccon et al., 26 Sep 2025, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).

The open questions are therefore not merely algorithmic but conceptual. One direction is to evaluate these formulations on additional real-world datasets and domains such as GPS route recommendation, where one driver’s recourse can worsen congestion for others. Another is to expose uncertainty directly, for example by communicating recourse as a probability of success rather than as a deterministic promise. A further direction is to model applicants as utility-bearing agents with heterogeneous costs and preferences, and to analyze equilibrium, congestion, or mean-field effects. Other open directions include strategic providers, decentralized coordination, fairness-aware welfare objectives, dynamic systems in which providers retrain and seekers reapply, and joint treatment of causal actionability with competition and temporal validity (O'Brien et al., 2021, Fonseca et al., 2023, Ceccon et al., 26 Sep 2025, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).

Taken together, the literature redefines algorithmic recourse as a system property rather than a purely local explanation. A recommendation that is individually actionable at issuance may be ethically harmful, collectively infeasible, temporally unreliable, or invalidated by competition and model evolution. The central implication is that recourse in multi-agent settings must be evaluated not only by whether a single individual can cross a model boundary, but by how recommendations reshape the environment that all agents subsequently face (O'Brien et al., 2021, Fonseca et al., 2023, Ceccon et al., 26 Sep 2025, Khotanlou et al., 14 Aug 2025, Yang et al., 12 Mar 2025).