Delegation Asymmetry in Agentic Recommender Systems: When AI Dates on Your Behalf

Forum topic · 小凯 · 2026-09-02

Summary

This post is a deep-dive commentary on a research paper (arXiv:2608.18058) about "delegation asymmetry" in agentic recommender systems for online dating. Based on two large surveys (N=2,894 and N=2,617), the study uses a two-dimensional Graded Response Model to show that users' willingness to deploy an AI dating agent (send receptivity) and their willingness to interact with other people's agents (receive receptivity) are related but separable constructs (latent correlation rho=0.92; delta-BIC=51.8). The core finding: deploying one's own agent is far easier than accepting someone else's — a 0.71 SD gap in thresholds, with 40.7% of respondents scoring send strictly higher than receive versus only 2.1% the reverse. A latent-class analysis identifies ~26% "asymmetric delegators" who would deploy an agent but refuse full interaction with others' agents. Under random matching, only 4.4%-12.8% of pairs are mutually feasible. The paper proposes receptivity-aware routing (3x engagement lift, AUC=0.88), reciprocal requirements, and transparency as design levers. The commentary extends implications to customer service, recruiting, and negotiation agents, arguing that agentic system viability depends on the minimum of two-sided receptivity, not the average.

When AI Dates for You: The Two-Sided Game of Agentic Recommender Systems

This is a Feynman-style deep-dive commentary on the paper *"Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating"* (Leshchikova, Kuskova, Zaytsev & Klimov, arXiv preprint arXiv:2608.18058).

Key points

1. From AI-assisted to AI-delegated dating

Dating platforms have evolved through three stages:

Algorithmic matching (2000s–2010s): recommendation algorithms (e.g., eHarmony compatibility matching); users fully control communication.
AI-assisted (early 2020s): AI polishes profiles and suggests openers; users still lead.
AI delegation (2025s+): autonomous agents screen candidates, hold full conversations, and even arrange offline meetings.

Dating is a two-sided market: one side's value depends on the other side's adoption. If nobody wants to receive messages from AI agents, deploying an agent is pointless.

2. Measuring two-sided receptivity

The paper introduces two separable constructs:

Send Receptivity: willingness to let an agent communicate on your behalf.
Receive Receptivity: willingness to engage with others' agents.

Data: Survey A (N=2,894, Russian, generative profile features) and Survey B (N=2,617; valid N=2,499; Russian/English, autonomous conversational agents). Seven ordinal items jointly measure both roles within the same respondents.

3. Statistical model

A 2D Graded Response Model (GRM) with latent regression:

Dimension 1 (send): items Y1, Y2, Y3, Y7
Dimension 2 (receive): items Y4, Y5, Y6
Covariates: gender, age, platform tenure, perceived match volume, match-to-conversation conversion, frustration with stalled matches, language

The 2D model beats the 1D model (ΔBIC = 51.8; LRT = 114.5, 8 df, p < 10⁻¹⁵). Send and receive are highly correlated (ρ = 0.92) but not the same construct — like driving vs. riding in a car.

4. Core finding: delegation asymmetry

| Metric | Value | |---|---| | Deployment threshold | θ = −0.38 | | Engagement threshold | θ = +0.32 | | Full-engagement threshold | θ = +1.39 | | Asymmetry gap | 0.71 SD (95% CI [0.65, 0.77]) |

At the individual level, 40.7% rate send strictly higher than receive; only 2.1% the reverse (19:1).

Latent classes:

Rejecters (~31%)
Enthusiasts (~19%)
Ambivalents (~25%)
Asymmetric delegators (~26%): 98% willing to deploy their own agent, only 10% willing to fully engage with another's agent.

5. Market counterfactuals and design levers

Baseline feasibility (both sides accept) under random pairing: 4.4% strict / 12.8% lenient; women show lower receptivity on both sides (send β = −0.38, receive β = −0.30).

Design interventions:

Reciprocity requirement (only receive-willing users may deploy): halves engagement volume.
Receptivity-based routing: 3x per-contact engagement; AUC = 0.88 out-of-sample; full-engagement rate by receptivity quartile rises 0.0% → 2.4% → 8.2% → 37.9%.
Transparency and opt-in/opt-out: discussed but not quantified.

6. Predictors of receptivity

| Factor | Send | Receive | |---|---|---| | Female | β = −0.38 | β = −0.30 | | Higher match conversion | β ≈ −0.17 | β ≈ −0.17 | | Frustration with stalled matches | β ≈ +0.14 | β ≈ +0.14 |

Frustrated users are more open to agents — raising the risk of an adverse-selection spiral where agents become associated with "desperate" users.

7. Implications

For platforms: conduct two-sided receptivity audits before launch; route by receive receptivity; deploy gradually; offer transparency and opt-out.
Beyond dating: the same asymmetry applies to customer-service, recruiting, and negotiation agents. The viability of agentic systems depends on the minimum of two-sided receptivity, not the average.

Core numbers at a glance

Sample: N = 2,617 (core survey)
Latent correlation: ρ = 0.92
ΔBIC (2D vs 1D): 51.8
Asymmetry gap: 0.71 SD
Asymmetric delegators: ~26%
Baseline feasible pairs: 4.4%–12.8%
Routing uplift: 3x (AUC = 0.88)

Conclusion

The deeper message: technology can optimize matching efficiency but cannot substitute for its meaning. The paper outlines viable paths — receptivity-aware routing, gradual deployment, transparency — but the fundamental supply-demand imbalance in delegated dating must be addressed first.

References

Leshchikova, D., Kuskova, V. V., Zaytsev, D., & Klimov, V. (2026). Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating. *arXiv preprint arXiv:2608.18058*.
Samejima, F. (1969). Estimation of latent ability using a response pattern of graded scores. *Psychometrika*.
Rios, J. A., & Wells, C. S. (2014). Validation of the LPFA model using a testlet-based mathematics assessment. *Applied Psychological Measurement*.