Cross-Model Humor Preference Modeling with Cards Against Humanity
An arXiv paper (2608.07481) explores the ability of one large language model to mimic the humor preferences of another within a structured Cards Against Humanity-like experiment. This study, presented as a cross-type submission, assesses two models: GPT-4o acts as the Czar while Claude Opus-4.5 serves as the Player in a binary humor-selection challenge, ensuring that self-preference cannot lead to success. Researchers utilized a reflected-cell stability method to identify 244 hands where the models exhibited deterministic yet opposing preferences, dividing them into a context pool of 97 hands and a test pool of 147 hands. The Player's performance is analyzed across five graded conditions, aiming to distinguish between framing effects and direct behavioral evidence. This innovative research, available on arXiv, sheds light on cross-model alignment in subjective tasks like humor, with potential implications for AI personalization and human-AI interactions.
Key facts
- The paper is available on arXiv with ID 2608.07481.
- The study uses a Cards Against Humanity-style task to test humor preference modeling.
- Two models are evaluated: GPT-4o as Czar and Claude Opus-4.5 as Player.
- The task is binary humor-selection, designed to prevent self-preference success.
- A reflected-cell stability procedure isolated 244 hands with deterministic opposite preferences.
- The hands are partitioned into a 97-hand context pool and a 147-hand held-out test pool.
- The Player is evaluated across five graded conditions.
- The conditions include default self-preference, generic instruction, model-identified Czar, prior selections, and prior selections with rationales.
- The study aims to separate framing effects from direct behavioral evidence.
- The paper is a cross-type submission (announce type: cross).
Entities
Institutions
- arXiv