Source-linked AI summary
Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating
Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, Valerii Klimov
TL;DR
Agent-mediated dating features require acceptance from both senders and receivers, but existing measures do not distinguish these roles or locate receptivity thresholds. The paper measures both within person and finds a robust delegation asymmetry: users are substantially more willing to deploy agents than engage with others’ agents, leaving only 4–13% of random directed dyads combining deployment and engagement.
Problem
Existing attitude measures cannot determine whether both sides of an agent-mediated feature’s market exist or locate sender and receiver receptivity thresholds.
Method
The paper jointly models within-person send and receive receptivity with graded-response latent factors, then uses estimated propensities in validation and market counterfactuals.
Results
4–13% of random directed dyads combine agent deployment with receiver engagement, reflecting a robust 0.7 SD delegation asymmetry favoring deployment.
Takeaways & Limitations
Agentic matching should treat receive-side consent and receptivity-aware routing as design primitives rather than optimize only for sender deployment.
Takeaways & Limitations
Estimates are stated preferences from concept mockups on one self-selected platform, so their relationship to behavior under deployed agents is uncertain.
Abstract
from arXiv · showhide
Autonomous LLM agents that converse on a user's behalf are an emerging design pattern in matching platforms, yet their viability depends on a condition rarely examined: users must accept not only delegating conversation to an agent, but also receiving agent-mediated communication from others. We study this condition using two large-scale surveys of active users of a major dating platform (N=2,894 on generative profile features; N=2,617 on autonomous conversational agents, fielded in two languages). We develop a latent-variable measurement model of agent receptivity based on graded response models with latent regression, and show via model comparison that willingness to send and willingness to receive agent communication are distinct constructs: highly correlated (rho=0.92) but separable (Delta BIC=52), with partial measurement invariance across languages. The model quantifies a systematic delegation asymmetry: deploying one's own agent requires far lower receptivity (threshold -0.38) than engaging a counterpart's agent (+0.32; full engagement +1.39), and mean deployment propensity exceeds engagement propensity roughly threefold. Under a random-pairing counterfactual derived from stated receptivity, only 4-13% of directed dyads combine agent deployment with receiver engagement, with a pronounced gender-directional imbalance. Design counterfactuals quantify the levers: a reciprocity requirement cuts interaction volume by half or more by excluding nearly two-thirds of would-be deployment, while routing agent contacts on receive receptivity triples per-contact engagement, a lift that survives out-of-sample validation with the target item held out (AUC 0.88, 3.1x quartile lift under respondent-level cross-validation). We discuss implications for agentic recommender design, including disclosure, opt-in mechanics, and receptivity-aware matchmaking.
1 Introduction
This study jointly measures users’ willingness to delegate communication and engage with agent-mediated messages, treating agentic dating as a two-sided market condition. It finds distinct but correlated receptivity dimensions, a substantial delegation asymmetry, and limited baseline viability that can be improved through receptivity-aware design.
- Measurement design: Two surveys of active users (N = 2,894 and N = 2,617) support a latent-variable model measuring send and receive receptivity jointly within person.The graded response model with latent regression uses seven attitudinal items and was fielded in two languages.
- Core findings: ρ = 0.92 and ΔBIC = 51.8 show send and receive receptivity are highly correlated but decisively separable, with partial measurement invariance across languages.The constructs represent willingness to delegate one’s own communication and willingness to engage with agent-mediated communication from others.
- Core findings: 0.71 SD separates deployment from engagement: deploying one’s own agent requires θ = −0.38, whereas engaging a counterpart’s agent requires θ = +0.32.The displacement survives every tested recoding.
- Market consequences: 4–13% of directed dyads combine agent deployment with receiver engagement under random pairing, with a gender-directional imbalance.The estimate comes from a counterfactual based on the fitted stated-preference model.
- Design levers: −65% of deployers are excluded by a reciprocity norm, while receptivity-aware routing produces a 3× per-contact engagement gain that survives leave-target-item-out cross-validation.Receive-side consent emerges as a first-class design primitive and routing signal.
2 Related Work
Prior work frames online dating as a reciprocal, two-sided matching market shaped by asymmetric attention and scarce replies, while AI-mediated communication exposes distinct sender and receiver concerns. Research on LLM agents develops the supply side of agent-mediated interaction; this study measures demand jointly and applies latent-variable psychometrics to identify receiving willingness as a constraint.
- Online dating as a matching market: Online dating is a two-sided matching market where users sort on observable attributes, direct attention upward, and leave most first messages unanswered.Reply behavior is described as scarce, skewed, and predictable from user and dyad features.
- Reciprocal recommendation: Reciprocal-recommendation research models mutual interest, directional preferences, behavioral matching, latent factors, and exposure trade-offs in two-sided platforms.This literature includes RECON, large-platform behavioral models, latent-factor formulations, and fairness-aware reciprocal recommendation.
- AI-mediated communication: AI-mediated communication research documents receiver-side trust penalties and sender–receiver asymmetries, including authenticity concerns among senders and digital betrayal among receivers.These findings motivate measuring send- and receive-willingness jointly within individuals on one latent scale at population size.
- LLM agents and agentic recommenders: LLM-agent research develops autonomous simulacra, recommender-tool planners, user-side shields, and dating-specific multi-agent matching prototypes, but primarily builds the supply side.The paper positions its contribution as measuring the demand side, with receiving demand as the binding constraint.
- Measuring attitudes toward AI: Technology-adoption research and AI-attitude instruments establish heterogeneous responses to algorithmic judgment, while item-response models support ordinal measurement, latent regression, and cross-group invariance testing.The framework uses graded response models estimated by marginal maximum likelihood and interprets claims through thresholds and cross-group comparisons.
3 The Receptivity Audit
The receptivity audit tests whether both sides of an agent-mediated communication market exist by measuring deployment and receiving roles within the same respondents. It combines a multidimensional measurement model with validated propensity estimates and dyadic counterfactuals to connect receptivity to market design levers.
- Instrument: The audit measures receptivity in both deployment and receiving scenarios within each respondent, separating role effects from population effects.This within-person coverage is the instrument’s non-negotiable design element.
- Measurement model: The graded response model estimates separate send and receive factors, tests dimensionality, and provides common-scale thresholds, invariance assessments, and uncertain per-user scores.Latent traits also regress on user covariates, while model comparison determines whether the dimensions are separable.
- Propensity estimation: Model-implied endorsement propensities translate factor scores into deployment and engagement probabilities under transparent strict and soft bounds for intermediate responses.These propensities are treated as scenario bounds rather than calibrated behavior and are predictively validated before policy use.
- Market counterfactuals: Dyadic simulation converts individual propensities into interaction volume, engagement per contact, directional imbalances, and counterfactual effects of eligibility, routing, disclosure, and consent rules.Because the levers use the model’s own parameters, each counterfactual is computed in closed form rather than estimated through a new experiment.
- Scope: The audit’s contribution is the receptivity construct and its bridge to market primitives, using established machinery and extending beyond dating to other delegated-communication settings.Examples include recruiting outreach, sales prospecting, agent-mediated customer contact, and professional networking.
4 Data and Instruments
The study used two voluntary, self-administered surveys of active users of a large dating platform: one on autonomous agents in Russian and English, and one on generative features in Russian. The instruments elicited graded receptivity responses, measured covariates, and released coded, de-identified records with documented preprocessing.
- Survey instruments: 2,617 respondents completed the autonomous-agents instrument in Russian or English, while 2,894 completed the generative-features instrument in Russian; the samples were not linked.The agents survey ran from November 12 to December 15, 2025, and the generative-features survey from March 22 to April 1, 2026.
- Survey instruments: Three generative-AI concepts were rated on a three-level interest scale: an AI profile summary, conversation tips, and an AI couple description.A minority supplied 473 free-text comments across items instead of ratings; these responses were excluded from corresponding item denominators and reviewed only for context.
- Survey instruments: The autonomous-agent instrument presented annotated mockups of a configurable agent messaging matches on users’ behalf, followed by seven attitudinal items rated immediately after each scenario.Users could configure activity level—initiates versus replies only—and tone.
- Measurement inputs: Six ordinal covariates captured gender, age band, platform tenure, perceived match volume, match-to-conversation conversion, and negative affect about stalled matches.Two questions on conversation-decay attributions and coping strategy were not used in the measurement model.
- Measurement inputs: Responses were coded so higher categories indicated greater receptivity, with selected rejection options collapsed and Y7’s 1–10 rating binned into five ordered levels.Sensitivity analyses instead treated the collapsed categories as distinct.
- Data release: 2,617 Instrument B records and corresponding Instrument A records were deposited with verbatim and analysis-coded responses, covariates, language flags, and documented de-identification.Released data removed contact information and free text, coarsened timestamps to ISO week, and enforced a minimum released demographic cell size of k=5.
5 A Measurement Model of Agent Receptivity
The section models agent receptivity as a latent construct measured by seven ordinal items in a graded response model with latent regression. A correlated two-dimensional specification separates send and receive factors, with estimation and identification defined through marginal maximum likelihood and standardized traits.
- Model specification: Seven ordinal items are modeled with a graded response model that estimates positive item discriminations and ordered response thresholds.The model treats observed responses as ordinal indicators of latent receptivity rather than directly summing raw scores.
- Model specification: The two-dimensional specification assigns items to a send factor (Y1, Y2, Y3, Y7) or a receive factor (Y4, Y5, Y6).The latent traits are modeled in a correlated space, allowing the dimensions to remain distinct while related.
- Latent regression: Latent regression incorporates centered covariates including gender, age, tenure, match volume, conversion, negative affect, and language.Centering makes the latent regression a pure slope structure while thresholds absorb location.
- Estimation and identification: The analysis estimates parameters by marginal maximum likelihood with 132-node Gauss–Hermite quadrature and gradient-based quasi-Newton optimization under automatic differentiation.The correlated latent space is integrated numerically during estimation.
- Estimation and identification: Identification fixes both trait variances to one, centers covariates, and selects dimensionality and latent correlation using BIC and likelihood-ratio tests.Per-user factor scores are computed as expected a posteriori values under the fitted model.
6 Results: The Structure of Receptivity
The results show that willingness to send and receive agent communication are strongly related but distinct, with deployment substantially easier to endorse than engaging another user’s agent. This asymmetry persists across robustness checks, while receive-side responses show limited cross-language noninvariance and reveal a sizable asymmetric-delegator class.
- Two-dimensional receptivity: ΔBIC = 51.8 favors a two-dimensional model, with latent correlation ρ = 0.92: send and receive receptivity are strongly related but not interchangeable.All seven items discriminate strongly, with a between 1.7 and 4.2; deployment is the most informative send-side item.
- Delegation asymmetry: −0.38 is the deployment threshold versus +0.32 for engaging a counterpart’s agent and +1.39 for full engagement, a 0.71 SD displacement.The displacement remains stable across alternative category orderings, ranging from 0.70 to 0.72.
- Delegation asymmetry: 40.7% of respondents rated deployment above engagement, versus 2.1% in the opposite direction, a 19:1 imbalance.Joint rejection was the modal paired response at 36.1%, while only 11.0% fully endorsed both sending and receiving.
- Measurement invariance: Two receive-side items violate full cross-language invariance, while the remaining five are invariant and freeing Y4/Y5 leaves the gap at 0.69 and ρ = 0.92.All send-side items function equivalently across language forms; the asymmetry-defining items Y3 and Y4 show no gender DIF.
- Latent classes: ≈26% are asymmetric delegators: 98% want to try or might try their own agent, but only 10% would fully engage an agent contacting them.This class enjoys agent-to-agent conversation, establishing a discrete population pattern behind the continuous asymmetry.
7 Stated-Preference Market Counterfactuals
Random-pairing stated-preference counterfactuals produce low engagement, directional imbalance, and a tradeoff between interaction volume and engagement quality. Reciprocity gates reduce volume, while receive-receptivity routing substantially improves engagement per contact at lower coverage.
- Baseline: 4.4%–12.8% of directed dyads yield engaged agent interactions, while engagement per contact is 11.6%/25.6% under strict/soft operationalizations.Most modeled agent contacts therefore do not produce endorsed engagement.
- Directional imbalance: 8.4% versus 13.4% engagement per contact occurs for male-deployed agents contacting women versus female-deployed agents contacting men.The heavier prospective male-to-female traffic faces the colder reception.
- Reciprocity gate: 65% of likely deployers are excluded and strict interaction volume falls from 0.044 to 0.020 when soft engagement propensity must reach 0.5 for deployment.The reciprocity gate removes traffic from users willing to send agents but not receive them.
- Receptivity-aware routing: 39.4% strict engagement per contact results from routing to the top receptivity quartile, versus 11.6% baseline, while coverage falls to 25% of receivers.Under soft operationalization, engagement rises from 25.6% to 64.0%, a 3.4× quality gain; the policy uses receive-side receptivity.
- Caveats: Random dyad formation, static propensities, and omitted matching structure bound the simulation, while receive-side scarcity, directional imbalance, and volume–quality tradeoffs remain qualitative conclusions.The authors identify an equilibrium treatment as future work.
8 Design Implications for Agentic Recommenders
Agentic recommender design should treat receive-side consent and receptivity as core matching primitives, while recognizing reciprocity as a quantified platform-values choice. Rollouts should progress from less intrusive features to autonomous delegation with opt-in on both sides, and the audit framework can extend to other delegated-communication markets.
- Receive-side consent: Three quarters of agent contacts reach users who did not want them, making receive-side consent a mechanism for market quality rather than merely a compliance feature.The proposed receiving-consent surface would specify whether, and from whom, agent-mediated contact is acceptable.
- Receptivity-aware routing: AUC 0.88 shows that receive receptivity predicts held-out engagement even when the engagement item is excluded from scoring.Routing agent traffic to the top receptivity quartile more than triples per-contact engagement.
- Reciprocity: Halving agent-interaction volume, a symmetric-willingness rule excludes two thirds of would-be deployers, quantifying reciprocity’s platform-level price.The excluded group is almost exactly the asymmetric-delegator segment, so adopting the rule is framed as a values decision rather than a technical necessity.
- Beyond dating: The audit transfers beyond dating wherever delegation meets a human receiver, including recruiting, where agent-authored candidate outreach is already deployed but candidate-side receptivity is unmeasured.The proposed generalization rests on mechanisms described as non-dating-specific and requires re-fielding the instrument to test it.
- Rollout sequencing: A staged rollout should introduce profile-level generative features first, conversation assistance with strong user control second, and autonomous delegation last with opt-in on both sides.The sequence follows increasing relational intrusion; deploying autonomous delegation first is characterized as demand-inefficient and trust-corrosive.
9 Limitations
The study measures stated preferences from concept mockups rather than behavior with deployed agents, and its samples are self-selected from one platform. Cross-language conclusions are constrained because the English subsample is small and language is confounded with population composition.
- Measurement and sampling limitations: Stated preferences from concept mockups may diverge from adoption behavior under deployed agents, with the direction of divergence for receiving agents unknown.The technology-acceptance literature suggests stated and revealed adoption correlate but diverge.
- Measurement and sampling limitations: Both samples are self-selected respondents from a single dating platform, limiting generalizability beyond that platform and population.
- Measurement and sampling limitations: The English subsample contains n = 232 self-selected respondents, confounding language comparisons with population composition.Accordingly, crosslingual claims are limited.
10 Conclusion
Agent-mediated matching is being built sender-first, but its success depends on receiver receptivity. The study finds distinct sending and receiving constructs, a robust asymmetry, and only 4–13% of directed dyads combining deployment with engagement unless receive-side consent becomes a design primitive.
- Conclusion: Sending and receiving agent communication are distinct constructs, separated by a robust 0.7 SD asymmetry.The asymmetry is measured within person.
- Conclusion: A quarter of users want to send agent traffic they would not fully accept.
- Conclusion: 4–13% of directed dyads combine agent deployment with engagement under a random-pairing counterfactual.The passage identifies receive-side consent as necessary for broader viability.
Ethical Considerations
The study uses de-identified, aggregate survey data with safeguards against re-identification and reports ethical risks surrounding receptivity scoring, gender differences, disclosure, and agent-mediated intimacy. It argues for mandatory agent-authorship disclosure while treating receptivity measurement as informing, rather than resolving, broader debates about delegated romantic communication.
- Data handling and review: Both surveys were voluntary secondary analyses of de-identified platform data, reviewed as not human subjects research and reported only in aggregate.The public deposit removes identifiers and free text, coarsens timestamps, and excludes small demographic cells and volunteered candidate contact information.
- Dual use of receptivity scoring: Receptivity scoring can protect users by routing unwanted agent traffic away, or exploit emotionally vulnerable users by maximizing AI exposure.The most receptive users are also the most frustrated with their current experience, sharpening the dual-use concern.
- Gender differences: Aggregate gender differences should not gate features or justify individual differential treatment; their design relevance is directional rather than individual.The reported pattern is that the heaviest prospective agent traffic faces the least receptive audience.
- Deception and disclosure: Undisclosed agent communication turns receive-side receptivity estimates into deception measurements, supporting mandatory disclosure of agent authorship.Users who would decline agent contact cannot decline what they cannot detect, and prior work reports trust penalties even for suspected AI authorship.
- Agent-mediated intimacy: Agent-mediated romantic communication raises unresolved questions about whose words ground relationships and how users transition to unassisted interaction.Users articulated these concerns unprompted; the study presents measurement and market analysis to inform the debate rather than settle it.