Love First, Know Later: Persona-Based Romantic Compatibility Through LLM Text World Engines¶

🕒 Published (v1): 2025-12-04 02:07 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper proposes "Love First, Know Later," a compatibility assessment framework that replaces static profile matching with LLM-simulated romantic interactions. Compatibility is formalized as a reward modeling problem, where LLMs simultaneously act as persona-driven agents and as the environment simulating interaction dynamics. Theoretical convergence to optimal stable matching is proved under sparse-rewards and deterministic-decisions hypotheses.

Problem¶

Existing dating platforms compare static profiles, which cannot capture the emergent, interaction-driven nature of compatibility. Decades of relationship research show compatibility arises from responses to critical moments, not attribute similarity—a gap that profile-matching algorithms cannot bridge.

Method¶

An LLM text world engine operates in dual capacity: (1) as persona-driven agents π̂_i, π̂_j approximating human policies from natural-language persona narratives, and (2) as the environment that generates topics, models emotional state transitions, and evolves dialogue context. Compatibility is formalized as inverse RL: given observed matching outcomes, a linear reward function is learned over ratings extracted by a "love observer" LLM (participant self-ratings r1, r2 + external observer rating r3). Simulation is made tractable via two hypotheses grounded in relationship psychology—sparse rewards (only critical moments matter) and deterministic decisions (behavior is consistent at critical moments). Two modes are implemented: Speed Dating Mode (brief dyadic dialogues) and Critical Events Mode (pivotal scenarios like career conflicts). A separate world engine narrates environmental state in the divorce task, enforcing separation of agent policies from environmental dynamics.

Key Contributions¶

"Love First, Know Later" paradigm: simulate interactions first, then extract compatibility signals.
LLM text world engine architecture: dual-capacity LLM as both agent and environment.
Formalization of compatibility as reward modeling (inverse RL flavor) with a three-component observer rating system.
Theoretical convergence proof: as policy approximation error ε→0, predicted matching converges to optimal stable matching (Gale-Shapley).
Validation on Columbia Speed Dating dataset (initial chemistry) and Divorce Prediction dataset (long-term stability).

Results¶

Speed Dating (initial chemistry):
LLM Love Observer: F1=0.67, AUC=0.60 (Stage 1, pre-date info only); F1=0.67, AUC=0.56 (Stage 2, with during-date signals)
LLM Mixed: F1=0.64, AUC=0.57 (Stage 1); F1=0.67, AUC=0.57 (Stage 2)
Logistic regression baseline: F1=0.66, AUC=0.61 (Stage 1); F1=0.61, AUC=0.60 (Stage 2)
Similarity baseline: F1=0.55, AUC=0.54 (Stage 1); F1=0.65, AUC=0.50 (Stage 2)
LLM methods modestly exceed or match baselines; overall performance is limited by dataset sparsity.
Divorce Prediction (long-term stability):
Logistic regression: F1=0.95, AUC=1.00 (benefits from >0.9 feature correlation)
LLM Love Observer (10 ICL examples): F1=0.90, AUC=0.92
LLM methods perform comparably without personalized fine-tuning, validating feasibility of the critical events hypothesis.

Limitations¶

Text-only simulation omits non-verbal cues (tone, gesture, expression).
Divorce prediction logistic regression is an extremely strong baseline due to highly correlated survey features (>0.9), making the comparison favorable to the baseline.
Dataset size is small (170 couples for divorce; 8,378 dates but outcome sparsity noted).
Current framework is dyadic; multi-person dynamics (group dating, family integration) require game-theoretic extensions.
Convergence theorem relies on hypotheses (sparse rewards, deterministic critical states) that are empirically motivated but not directly verified.
No fine-tuning of base LLMs; persona-approximation quality is unquantified.

Relevance to Agentic AI / LLM Agents¶

This work demonstrates a concrete application of persona-driven LLM agents operating within a self-contained simulated environment—a design pattern directly relevant to social simulation and multi-agent interaction research. The dual-capacity architecture (agent + environment in one LLM) is a novel structural choice that informs how agentic systems can bootstrap world modeling without separate environment simulators. The reward-modeling framing with inverse RL connects to agent alignment and preference learning, showing how human behavioral signals can supervise agent policies in non-task-oriented domains. The critical events hypothesis—that only sparse, high-stakes moments drive outcomes—is a generalizable principle for designing efficient agentic evaluation and rollout strategies.