Learn2Match: Matching Under Uncertainty

• Learn2Match is a research framework that learns latent features like worker skills and preferences to enable matching under uncertainty.
• It has evolved from threshold-based worker-task assignments to dynamic two-sided matching using techniques like reinforcement learning and Hungarian min–max estimation.
• Key studies demonstrate actionable insights, with methods achieving up to 85% of oracle performance and addressing challenges such as noisy feedback and evolving agent profiles.

Searching arXiv for "Learn2Match" and closely related papers to ground the article in current literature. Learn2Match is a polyvalent label in the arXiv literature rather than a single universally standardized algorithm name. In the 2017 paper “Learning to Match” (Ekman et al., 2017), it is used informally for a framework that repeatedly assigns workers to tasks when worker skills are unknown and must be inferred from noisy success or failure feedback. In the 2026 paper “Learn to Match: Two-Sided Matching with Temporally Extended Feedback” (Zong et al., 4 Jun 2026), Learn2Match denotes a formal framework and a multi-agent reinforcement-learning benchmark for dynamic two-sided matching with costly interviews, noisy post-match observations, evolving latent profiles, and endogenous dissolution. Related work extends the broader “learn-to-match” theme to retention-optimized two-sided ranking (Kishimoto et al., 17 Feb 2026), deep learning for mechanism design in two-sided matching (Ravindranath et al., 2021), and earlier learning-based matching under uncertainty or capacity constraints (Johari et al., 2016, Ahuja et al., 2016). The term should therefore be interpreted contextually: sometimes it names a specific benchmark, sometimes an informal shorthand for learning-to-match under uncertainty, and sometimes a broader research program on combining inference with matching optimization.

1. Terminological scope and historical usage

The official title of the 2017 paper is “Learning to Match” (Ekman et al., 2017). In that work, the problem is “matching under uncertainty”: a platform must assign workers to tasks repeatedly over time, while workers’ true skills are unknown and have to be learned from noisy performance feedback. The paper states that when there is only a single skill or job type, the problem is essentially a bandit problem, whereas the multi-skill, multi-task setting requires a new algorithmic treatment (Ekman et al., 2017).

The same label appears more explicitly in 2026 in “Learn to Match: Two-Sided Matching with Temporally Extended Feedback” (Zong et al., 4 Jun 2026). There, Learn2Match is both a formal framework and a benchmark environment. The benchmark models two-sided matching markets as a partially observable Markov game with evolving latent profiles, costly pre-match screening, noisy post-match observations, and continuation or dissolution decisions. Policies are evaluated using regret, social welfare, and information-friction loss (Zong et al., 4 Jun 2026).

A distinct but related usage appears in “Beyond Match Maximization and Fairness: Retention-Optimized Two-Sided Matching” (Kishimoto et al., 17 Feb 2026). That paper places itself squarely within the broader Learn2Match theme by replacing hand-crafted objectives such as total matches or fairness with a learned retention objective, operationalized through personalized retention curves and a dynamic learning-to-rank algorithm called Matching for Retention, or MRet (Kishimoto et al., 17 Feb 2026). This suggests that Learn2Match has evolved from a narrow worker-task assignment problem into a broader label for methods that learn the objective-relevant structure of a matching market and then optimize against it.

The string “l2Match” in “l2Match: Optimization Techniques on Subgraph Matching Algorithm using Label Pair, Neighboring Label Index, and Jump-Redo method” (Cheng et al., 2023) is terminologically similar but conceptually separate. That paper explicitly states that l2Match is not a learning-based algorithm; it is a deterministic, combinatorial subgraph isomorphism method (Cheng et al., 2023). In the matching-markets literature, by contrast, Learn2Match denotes problems in which uncertainty, inference, or learned utilities are central.

2. The worker–task formulation of “Learning to Match”

In “Learning to Match” (Ekman et al., 2017), the platform has a set of workers $\Workers$ and a set of tasks $\Tasks$. Each worker $\worker$ has an unknown skill vector

$\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$

with $\skill_{\worker,m} \in [0,1]$, and each task $\task$ has a known requirement vector

$\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$

with $\req_{\task,m} \in [0,1]$ (Ekman et al., 2017).

Time proceeds in discrete iterations $k = 1,2,\dots$. At each iteration, the algorithm chooses an assignment $a_k \in A$, consisting of worker-task pairs, and receives total reward

$\Tasks$0

where

$\Tasks$1

The objective is to maximize cumulative reward $\Tasks$2 (Ekman et al., 2017).

The feedback model is threshold-based and binary. For a worker-task pair, the observed outcome is Bernoulli with success probability $\Tasks$3 if the worker meets the requirement and Bernoulli with success probability $\Tasks$4 otherwise. The experiments denote the noise parameter by $\Tasks$5 and set $\Tasks$6 in subsequent experiments (Ekman et al., 2017). The paper considers both unrestricted matching, where a worker can be assigned to multiple tasks in the same iteration, and bipartite matching, where each worker is assigned to at most one task in each block (Ekman et al., 2017).

The paper emphasizes the difference between the single-skill and multi-skill cases. With one skill, the problem is equivalent to choosing an arm in a multi-armed bandit, so UCB, $\Tasks$7-greedy, and bounded $\Tasks$8-first apply directly. With multiple skills and multiple tasks, the platform faces high-dimensional skill learning, combinatorial matching, contextual structure induced by task requirements, and exploration–exploitation across both worker and task dimensions (Ekman et al., 2017).

3. Hungarian min–max estimation and empirical performance

The principal algorithmic contribution of “Learning to Match” (Ekman et al., 2017) is Hungarian min–max estimation, abbreviated HME. For each worker $\Tasks$9 and skill $\worker$0, the algorithm maintains a lower bound $\worker$1 and an upper bound $\worker$2, with point estimate

$\worker$3

Under deterministic feedback, a success raises the lower bound to the task requirement, while a failure lowers the upper bound to the task requirement (Ekman et al., 2017). The paper reports that, in the noisy Bernoulli setting, this min–max estimator gives better and faster skill estimates than a naive average-rating estimator (Ekman et al., 2017).

Once skill estimates $\worker$4 are available, the method constructs a cost matrix $\worker$5 over worker-task pairs and applies the Hungarian algorithm to compute an optimal bipartite matching under the estimated costs. If true skills were known, the same procedure would produce an oracle assignment; with estimated skills, it becomes a learning-based assignment rule (Ekman et al., 2017).

The paper compares HME against $\worker$6-greedy, a naive UCB extension, bounded $\worker$7-first, random assignment, and an oracle. In synthetic experiments with 3 skills, 10 workers, task counts ranging from 10 to 300 in the learning curves, and 25 repetitions per setting, HME is the best-performing method among those tested (Ekman et al., 2017). The headline result is that HME rapidly reaches and stabilizes around 85–86% of oracle performance, including around 85% at 50 tasks and about 85.7% at 300 tasks (Ekman et al., 2017). The abstract summarizes this as “results approaching 85% of oracle’s performance” (Ekman et al., 2017).

The same study is explicit about its limitations. It is primarily experimental and exploratory, provides no formal regret bounds or convergence theorems, assumes stationary worker skills and task requirement distributions, and relies entirely on synthetic data (Ekman et al., 2017). The authors invite the community to develop real-world benchmarks (Ekman et al., 2017). A plausible implication is that the original Learn2Match formulation established a practically motivated empirical template—estimate latent worker quality from task outcomes, then solve an assignment problem—but left much of the theory and benchmarking program open.

4. Dynamic and two-sided generalizations

A major line of generalization treats learning and matching as a dynamic two-sided problem rather than a static or repeated one-sided assignment. “Dynamic Matching and Allocation of Tasks” (Ahuja et al., 2016) studies a market with $\worker$8 workers and $\worker$9 clients, unknown productivities $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$0, unknown task qualities $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$1, and privately chosen effort levels. Because effort is unobserved, the environment exhibits moral hazard. The mechanism the paper derives includes a finite assessment phase in which each worker is matched to each task exactly once, a reporting phase in which workers rank tasks truthfully according to learned utilities, and an operational phase in which Gale–Shapley is run with workers as proposers. The induced MTBB strategy—maximum effort in assessment, truthful reporting, and bang-bang effort in the operational phase—is shown to be a weakly dominant strategy, and the resulting equilibrium is long-run coalitionally stable (Ahuja et al., 2016). Under additional assumptions of task homogeneity and a bimodal quality distribution, the mechanism achieves optimal long-run revenue (Ahuja et al., 2016).

“Matching while Learning” (Johari et al., 2016) develops a complementary constrained-bandit perspective. There is a finite set of worker types $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$2, a finite set of job types $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$3, Bernoulli payoff matrix $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$4, job arrival rates $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$5, and worker type distribution $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$6. Worker types are unknown at the individual level but known in aggregate; each worker remains in the system for $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$7 periods, and the platform seeks to maximize the steady-state rate of accumulation of payoff (Johari et al., 2016). The central result is that the optimal policy in the limit $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$8 has a two-phase structure: exploration followed by exploitation, both driven by shadow prices $\skill_{\worker} = (\skill_{\worker,1}, \dots, \skill_{\worker,M}),$9 obtained from the full-information linear program. These shadow prices define adjusted payoffs $\skill_{\worker,m} \in [0,1]$0, identify economically relevant confusions among types, and determine the optimal exploration policy. The DEEM policy achieves the asymptotically optimal regret rate $\skill_{\worker,m} \in [0,1]$1 when $\skill_{\worker,m} \in [0,1]$2 (Johari et al., 2016). This suggests that a principled Learn2Match system can often be decomposed into a global coordination layer based on dual prices and a local learning layer based on per-agent inference.

The 2026 Learn2Match benchmark pushes this dynamic view further by formulating matching as a partially observable Markov game

$\skill_{\worker,m} \in [0,1]$3

with latent worker profiles $\skill_{\worker,m} \in [0,1]$4, latent firm profiles $\skill_{\worker,m} \in [0,1]$5, belief states $\skill_{\worker,m} \in [0,1]$6, current matching $\skill_{\worker,m} \in [0,1]$7, and history $\skill_{\worker,m} \in [0,1]$8 (Zong et al., 4 Jun 2026). In that environment, both sides may propose interviews, only workers propose matches, either side may dissolve a match, and information is revealed gradually through interviews and tenure-dependent post-match signals (Zong et al., 4 Jun 2026). The model explicitly captures temporally extended feedback, rather than immediate one-shot feedback.

5. Learned objectives: preferences, retention, and mechanism design

Another branch of Learn2Match research changes not only how matching is optimized but also what objective is learned. “Online Dating Recommendations: Matching Markets and Learning Preferences” (Tu et al., 2014) is an early example. It models a recommendation $\skill_{\worker,m} \in [0,1]$9 with decision variable $\task$0, initiation probability $\task$1, reply probability $\task$2, and objective

$\task$3

subject to capacity constraints on both senders and receivers (Tu et al., 2014). Preferences are learned using an LDA-style latent type model over user profile tuples, yielding

$\task$4

On data from Baihe.com, the paper reports that two-sided preference ranking yields substantial gains relative to one-sided ranking, including median relative gain in success rate for male suitors of about 46.84% and for female suitors of about 16.5%, with the abstract summarizing this as improvement “by as much as 45%” (Tu et al., 2014). The work is one-sided in optimization despite the dating context, but it established the pattern of learning latent preferences and then embedding them into a constrained matching objective.

“Beyond Match Maximization and Fairness: Retention-Optimized Two-Sided Matching” (Kishimoto et al., 17 Feb 2026) replaces immediate matches or fairness with user retention as the primary objective. It defines personalized retention curves $\task$5, where $\task$6 is cumulative matches, and chooses a ranking $\task$7 to maximize total retention gain for both the receiver and recommended users (Kishimoto et al., 17 Feb 2026). Under a concavity assumption on $\task$8, the paper derives a lower-bounding score function

$\task$9

that yields a sorting-based dynamic learning-to-rank policy, MRet, with $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$0 complexity per request (Kishimoto et al., 17 Feb 2026). Synthetic and real experiments on online dating data show that MRet achieves higher retention than match-maximization or fairness baselines (Kishimoto et al., 17 Feb 2026). This suggests a broader interpretation of Learn2Match: learning the utility function that matters for the platform, not merely learning preferences or skills.

A different learned-objective perspective appears in “Deep Learning for Two-Sided Matching” (Ravindranath et al., 2021). That paper does not learn preferences; it learns a randomized matching mechanism $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$1 that maps ordinal preference profiles to feasible randomized matchings. It introduces differentiable surrogates for ex ante stability and ordinal strategy-proofness, denoted $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$2 and $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$3, and trains neural mechanisms by minimizing

$\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$4

The learned mechanisms trace a stability–strategy-proofness frontier that is strictly better than convex combinations of deferred acceptance, top trading cycles, and randomized serial dictatorship in $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$5 markets (Ravindranath et al., 2021). In this literature, Learn2Match thus includes the possibility of learning the mechanism itself rather than only latent agent attributes.

6. Benchmarks, metrics, and methodological distinctions

A recurring distinction across Learn2Match work is between matching under immediate feedback and matching under temporally extended feedback. In the original worker-task setting, feedback is binary and contemporaneous; the main evaluation metric is percent of oracle cumulative reward, with HME approaching 85% of oracle performance on synthetic data (Ekman et al., 2017). In “Matching while Learning” (Johari et al., 2016), performance is evaluated by steady-state payoff rate and asymptotic regret relative to a full-information benchmark. In the dynamic task-allocation model, revenue optimality and long-run coalitional stability are central (Ahuja et al., 2016).

The 2026 Learn2Match benchmark adds a richer metric suite. Worker-side regret and firm-side regret are defined relative to the worker-optimal stable matching under true latent preferences, cumulative social welfare sums both sides’ rewards across periods, and information-friction loss measures the welfare gap between the full-information stable matching $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$6 and the stable matching induced by current beliefs $\Reqs_{\task} = (\req_{\task,1}, \dots, \req_{\task,M}),$7 (Zong et al., 4 Jun 2026). This metric is specific to environments where incomplete revelation distorts the stable structure itself. In experiments, independent PPO achieves higher cumulative social welfare and lower cumulative regret than the bandit-style CA-ETC baseline under temporally extended feedback, but CA-ETC achieves lower friction loss (Zong et al., 4 Jun 2026). The paper interprets this as evidence that MARL exploits long-horizon dynamics effectively but does not yet provide the coordinated exploration structure of matching-bandit methods (Zong et al., 4 Jun 2026).

The literature also distinguishes learning-based matching from optimization without learning. The subgraph-isomorphism system l2Match introduces a Label Pair Index, Neighboring Label Index, BCPRefine, and Jump-Redo for small-query subgraph matching, and reports query-time improvements over several baselines (Cheng et al., 2023). Yet it explicitly states that there is no machine learning, no training phase, and no learned models (Cheng et al., 2023). This is not a Learn2Match method in the matching-markets sense, despite the orthographic similarity. A common misconception is therefore to treat every “match” algorithm bearing “l2Match” or “Match” in the name as part of the same research line; the data do not support that conflation.

A second misconception is that fairness objectives are interchangeable with broader platform objectives. The retention paper argues directly that fairness equalizes some fairness functional, not the probability that users come back, and shows empirically that fairness-based exposure control does not optimize retention (Kishimoto et al., 17 Feb 2026). A plausible implication is that modern Learn2Match work increasingly treats the choice of objective as an estimable object rather than a fixed normative primitive.

7. Open problems and research directions

The original “Learning to Match” paper explicitly positions itself as preliminary work and calls for real-world benchmarks (Ekman et al., 2017). That request remains salient across the later literature. “Matching while Learning” provides asymptotic optimality for a stylized capacity-constrained setting with known payoff matrix and worker-type prior, but its finite-type, stationary assumptions are strong (Johari et al., 2016). “Dynamic Matching and Allocation of Tasks” obtains incentive compatibility and optimality under task homogeneity and a bimodal quality distribution, again indicating a structured but restrictive regime (Ahuja et al., 2016). The 2026 Learn2Match benchmark broadens realism by allowing noisy interviews, gradual revelation, and dissolution, but it is benchmark-oriented and does not yet deliver a next-generation algorithm that combines RL adaptability, bandit discipline, and stable-matching structure (Zong et al., 4 Jun 2026).

Several concrete future directions recur across papers. Better feedback models, including graded feedback and richer temporal structure, are identified in the original worker-task framework (Ekman et al., 2017). Jointly learning match probabilities and long-term utilities such as retention is proposed in the retention-optimized two-sided matching work (Kishimoto et al., 17 Feb 2026). Online adaptation of learned utility functions through contextual bandits or RL is another explicit extension (Kishimoto et al., 17 Feb 2026). The deep mechanism-design paper highlights scalability of adversarial misreport search and the need for architectures with permutation structure or graph inductive biases (Ravindranath et al., 2021). The 2026 benchmark specifically calls for methods that are “adaptive like RL agents, statistically disciplined like bandit algorithms, and structurally aware like stable-matching mechanisms” (Zong et al., 4 Jun 2026).

Taken together, these works indicate that Learn2Match is best understood as a family of methods for matching under uncertainty in which some economically relevant object—worker skill, user preference, retention response, latent profile, or even the mechanism itself—is learned and then coupled to a matching or ranking procedure. The field has progressed from threshold-feedback worker assignment (Ekman et al., 2017) through shadow-price coordination and dynamic assessment (Johari et al., 2016, Ahuja et al., 2016) to learned long-term objectives (Kishimoto et al., 17 Feb 2026) and temporally extended multi-agent benchmarks (Zong et al., 4 Jun 2026). This suggests an increasingly unified agenda: learning what matters for matching, and designing algorithms that exploit that learned structure without sacrificing the combinatorial, strategic, and temporal structure of the market.