acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Bipartite Matching, decentralized and Dynamic Matching, LLM Social Simulation, LLM-agent Based Modeling.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_A7jnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper splits marriage matching into “whom do I like?” and “who might accept me?” An LLM scores the former, while each agent learns the latter with its own Logistic-UCB model. The authors compare matching outcomes in a synthetic 50×50 market and provide local learning guarantees with a limited scope.

Critical strengths
  1. Each agent learns acceptance probabilities only from feedback on its own proposals. This fits settings where agents can meet only some of the available candidates. (pp. 5–6)
  2. The experiments include a Utility-only baseline that does not learn acceptance probabilities, under the same market rules. The paper also clearly states that local guarantees do not imply global stability or optimal welfare. (p. 7, Table 1; Appendix L)
Critical weaknesses
  1. The behavioral validation is not sufficiently independent: the prompts and validation reference come from the same study, and the no-CoT baseline also removes the gender-specific preference instructions. This makes it hard to tell whether the gains come from the prompt content or the reasoning steps. Independent behavioral data are also missing. (pp. 3–4, 7)
  2. The gains from reciprocity learning are small: compared with Utility-only, mean welfare rises by just 0.09 and the number of blocking pairs falls by 1.30, while the gender rank gap grows. The paper does not report confidence intervals for the paired differences, and the main experiments all use the same 100 agents. (p. 7, Table 1; p. 19)
Questions for authors
  1. Do the gains remain when the preference instructions stay the same and only the reasoning steps are switched on or off? Is there a behavioral reference that was not used to design the prompts and that supports the current validation conclusions?
  2. What are the confidence intervals for the paired differences between UCB and Utility-only? Do the welfare and stability gains hold across different populations?
Reviewer_ptk2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

The paper studies decentralized matching markets, where an agent (man or woman) can propose and the proposee can reject or accept based on their preferences. The paper uses LLMs to assess a man's (woman's) preference over a woman (man) in a few features, and then uses a probablistic model to compute an overall score. If the proposer has higher score than the benchmark (the current match), then the proposee would accept it; otherwise, the proposee would reject it. The paper studies a contextual-UCB algorithm, where each proposer computes a score as utility improvement times UCB estimate of acceptance probability, and proposers use this score to decide whom they want to propose. The paper implements their algorithms on a simulated Chinese marriage market dataset, and compares their algorithms with a few baselines.

Critical strengths

This paper seems to follow the framework of Hosseini et al, 2026, and extends the framework in the following dimensions: (1) the paper studies a decentralized market, where each agent can only see a few potential proposees. (2) the paper studies a contextual setting and proposes a UCB algorithm, and provides its theoretical guarantee in Theorem 1, though the regret bound is not surprising.

Critical weaknesses
  1. The biggest weakness that mutes my enthusiasm is poor writing and lacked motivation. On the one hand, it is unclear why one wants to model human preferences using LLM, and the current introduction (Line 63-78) does not motivate the choice of modeling clearly. As a result, it's unclear why one would like to use these methods in real matching markets. On the other hand, the current writeup is not easy to follow. For example, the authors introduce reward model error ηn\eta_n with three lines of texts in informal Theorem 1. I believe the authors could improve its clarity by introducing the full model in the main text. In Table 1, The authors do not introduce the definition of metrics reported, including Rank Gap and mutual welfare.

  2. Compared to Random eligible, UCB algorithm does not have statistical improvement in mutual welfare (56.01 vs. 55.24), and it has worse random gap and proposals. The only improvement is less blocking pairs. These results weaken the necessity of the UCB algorithm.

Questions for authors
  1. Can you explain the motivation of using LLM in matching markets?

  2. Random eligible method seems to be a naive approach, in which a proposer uniformly proposes an acceptable proposee. However, in Table 1, it has statistically undistinguishable multual welfare as UCB. This seems to be very counter-intuitive to bandit learning theory. Can you explain why?

Reviewer_XhNsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary
  • The paper looks at a decentralized, asynchronous two-sided matching model through a synthetic Chinese marriage market.
  • agents possess socioeconomic profiles, and gender-specific prompts guide an LLM to assign desirability scores to encountered candidates -each proposer separately fits a logistic model to its own proposal outcomes, estimating whether a candidate will reciprocate
  • proposals maximize the product of the proposer’s utility improvement and an optimistic acceptance estimate.
  • The paper then empirically evaluates on synthetic data.
Critical strengths
  • clear model of decentralized matching under limited information. The model allows the reader to assess what reciprocal learning adds w.r.t. how much an agent likes a candidate vs how likely that candidate is to accept, and it compares with a utility-only version
Critical weaknesses
  • the experiments does not yet show that reciprocal learning gives a meaningful improvement over Utility-only. Bandit-UCB raises mean mutual welfare only from 55.92 to 56.01 and lowers the number of blocking pairs from 17.58 to 16.28, while worsening the gender rank gap from 1.84 to 2.11. Since reciprocal learning is a central part of the paper, this make it difficult to judge whether it really provides an improvement, thereby hurting the significance of the paper
  • the main welfare measure averages scores only over realized pairs and leaves out unmatched agents. A method could therefore obtain a higher value by matching a smaller or more selected group. Also, the paper treats the LLM's scores as directly comparable across agents, even though the validation mainly checks rankings, which I find to be inconsistent.
Questions for authors

see weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.