Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating

Daria Leshchikova; Valentina V. Kuskova; Dmitry Zaytsev; Valerii Klimov

cs.AIAug 18, 2026

Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating

Authors: Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, Valerii Klimov

Organizations: Fleamily, Inc. Delaware, USA · Lucy Family Institute for Data & Society University of Notre Dame Notre Dame, IN, USA

Abstract

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.

Figures & tables

Item	Role	Stem (abridged)	Response options (code)
Y1	P	First emotional reaction to the agent concept	very interesting (2); neutral (1); didn’t seem trustworthy (0); not interesting (0)
Y2	P	Attitude to configuring the agent’s activity and tone	useful – saves me time (2); too complicated – can’t be bothered (1); risky – might write something inappropriate (0)
Y3	P	“How would you feel if the agent talked with a potential partner while you are away?”	I want to try it (2); I might try it (1); I don’t want an agent responding for me (0)
Y4	C	“How would you feel if someone’s AI agent responded to you?”	would engage in conversation with the agent (2); might engage (1); weird – only want a real person (0); would react negatively (0)
Y5	C	Your agent chats with another agent to find common interests before you join	super idea (3); funny – would watch their conversation (2); doubtful – why are we needed then? (1); strange and frightening (0)
Y6	C	Humans and agents converse together in a group chat	super idea (2); funny – would participate (1); doubtful / strange (0)

Table 1. Instrument B: agent-receptivity items. English form shown; higher codes indicate greater receptivity. P = respondent as principal (send dimension), C = respondent as counterpart (receive dimension).

Female	35.4%
Age 18–24	21.1%
Age 25–34	40.8%
Age 35–44	29.3%
Age 45+	8.8%
Tenure: $<$ 1 month	24.6%
Tenure: Several months	44.5%

Table 2. Sample descriptives, agent-survey complete cases (N = 2,499).

Model	$-\log L$	Params	BIC
1D	13,707.4	31	27,657.2
2D	13,650.1	39	27,605.4

Table 3. Model comparison. The two-dimensional model is preferred.

Item	Dim.	$a$	$b_{1}$	$b_{2}$	$b_{3}$	$b_{4}$
Y1	S	2.97	$-0.23$	$0.36$	—	—
Y2	S	3.44	$-0.33$	$-0.12$	—	—
Y3	S	4.20	$-0.38$	$0.38$	—	—
Y4	R	2.82	$0.32$	$1.39$	—	—
Y5	R	3.28	$-1.64$	$-0.36$	$0.99$	—
Y6	R	3.36	$-0.09$	$1.17$	—	—

Table 4. Item parameters of the 2D graded response model: discrimination

a

and ordered thresholds

b_{k}

(the latent-trait location at which the probability of responding in category

k

or higher reaches 50%). Item content and response categories are given in Table 1 ; dashes mark items with fewer categories; S = send, R = receive dimension.

Figure 1. Item response curves for deploying one’s own agent (Y3) and engaging a counterpart’s agent (Y4). The engagement curve is displaced

\approx

0.7 SD to the right.

Covariate	$\beta_{\mathrm{send}}$	$\beta_{\mathrm{recv}}$
Female	$-0.38$ [-0.57, -0.23]	$-0.30$ [-0.47, -0.18]
Age band	$+0.07$ [0.00, 0.15]	$-0.01$ [-0.07, 0.06]
Platform tenure	$-0.14$ [-0.24, 0.00]	$-0.11$ [-0.19, 0.04]
Match volume	$+0.09$ [-0.04, 0.17]	$+0.08$ [-0.02, 0.14]
Match $\to$ conversation rate	$-0.17$ [-0.27, -0.11]	$-0.16$ [-0.26, -0.09]
Negative affect (stalled matches)	$+0.14$ [0.07, 0.26]	$+0.14$ [0.06, 0.23]

Table 5. Latent regression of send and receive receptivity on centered covariates (trait SD = 1). Brackets: 95% nonparametric bootstrap CIs (200 replicates over respondents), recentered on the full-precision estimates.

Score	AUC	Top-quartile lift
Covariates only (logistic)	0.60	1.5 $\times$
Raw $\mathrm{Y5{+}Y6}$ sum score	0.85	3.0 $\times$
Logistic on six non-target items	0.89	3.2 $\times$
Latent (EAP, leave-Y4-out)	0.88	3.1 $\times$

Table 6. Routing-score baselines on identical folds (full engagement, out of fold). The logistic baselines are trained on the target labels; the latent score never uses Y4.

	Strict		Soft
Scenario	Vol.	Qual.	Vol.	Qual.
Baseline (unconstrained)	0.044	0.116	0.128	0.256
Male $\to$ female contacts	0.035	0.084	0.114	0.213
Female $\to$ male contacts	0.042	0.134	0.122	0.279
Reciprocity gate	0.020	0.116	0.046	0.256
Routing: top 10% receivers	0.025	0.648	0.041	0.809

Table 7. Random-pairing stated-preference counterfactuals. Vol. = viable dyad rate; Qual. = engagement per agent contact.

Figure 2. Receptivity-aware routing: engagement per agent contact (soft operationalization) as a function of receiver coverage

q

. The dotted line marks the unrouted baseline.

Human-AI Interaction

Human-AI interaction studies how people and AI systems communicate, collaborate, make decisions, and shape one another over time.

See all 330 papers

From Preference to Reciprocity: Decentralized Matching with Empirically Grounded LLM-agent Based Modeling
Sep 28, 2026Wangxuan Fan, Xiaoyu Nie, Zhoutian Shi +5Agent-Based Modeling Large Language Model-Based Social Simulation
The Chinese University of Hong Kong, Shenzhen
The User Asks, Platforms Compete: How Agentic Recommendation Markets Take Shape
Jul 28, 2026Deyao Hong, Kehan Zheng, Qian Li +3Mechanism Design Recommender Systems
Tsinghua University · Tencent Inc.
Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems
Jul 5, 2026Xinyu Lin, Yashar Deldjoo, Sunhao Dai +7LLM Agent Evaluation Recommender Systems
National University of Singapore · Polytechnic University of Bari · Renmin University of China +1

Sep 28, 2026J/K move · Enter open · S save

From Preference to Reciprocity: Decentralized Matching with Empirically Grounded LLM-agent Based Modeling

Wangxuan Fan, Xiaoyu Nie, Zhoutian Shi, Xiangcheng Meng+4

The Chinese University of Hong Kong, Shenzhen

Agent-Based Modeling Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

Bipartite matching is a fundamental problem in game theory and market design. Classical approaches such as Gale--Shapley assume complete preferences and centralized computation, whereas many real-world matching processes are decentralized, asynchronous, and shaped by sequential interaction under limited information. We propose a dynamic bipartite matching framework that combines large language model (LLM) agents with contextual bandits. In a simulated Chinese marriage market, economically grounded LLM agents evaluate locally encountered candidates, while agent-specific Logistic-UCB models learn reciprocal acceptance from realized proposal outcomes. The mechanism therefore separates two decisions---whom do I like? and who is likely to like me back?---without requiring ex ante market-wide preference rankings. We first validate LLM-induced mate preferences against the empirical conditional-logit reference across multiple LLM backbones. In the

50\times50

matching experiment, Bandit-UCB achieves the highest mean mutual welfare (56.01 versus 54.87 for Gale--Shapley), a smaller gender rank gap than the classical baselines, and the fewest blocking pairs among the LLM-ABM policies. Learned acceptance models show economically interpretable gender-differentiated associations, while counterfactual setups reveal no systematic unilateral advantage from prior search knowledge. Overall, these results support the advantages of decentralized matching with LLM-based behavioral modeling and online learning under incomplete information for economic simulation and computational social science research.

Abstract

Human-AI Interaction

Explore similar work

From Preference to Reciprocity: Decentralized Matching with Empirically Grounded LLM-agent Based Modeling

The User Asks, Platforms Compete: How Agentic Recommendation Markets Take Shape

Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems