Cross-Model Humor Preference Modeling with Cards Against Humanity

This paper investigates whether one large language model can approximate the humor preferences of another in a controlled Cards Against Humanity-style task. Two models — GPT-4o as Czar and Claude Opus-4.5 as Player — are evaluated on a binary humor-selection task constructed so that success cannot follow from self-preference. A reflected-cell stability procedure isolates 244 hands on which the two models hold deterministic but opposite preferences, partitioned into a 97-hand context pool and a 147-hand held-out test pool. The Player is then evaluated across five graded conditions: default self-preference, generic Czar-modeling instruction, model-identified Czar, prior Czar selections, and prior Czar selections with rationales. This gradient is designed to separate two sources of improvement: framing effects, in which the Player is told to attend to a Czar without seeing any of the Czar’s behavior, and direct behavioral evidence, in which the Player is shown the Czar’s prior choices. Player accuracy increased from 0.7% in Condition 1 to 19.0% and 25.9% in the framing-only conditions, and then rose to 72.8% and 82.3% once behavioral evidence and rationales were provided. An omnibus Cochran’s Q test and pairwise McNemar tests confirmed that each step in the gradient produced a significant improvement. The results indicate that role instruction and model identity yield only modest gains, while behavioral evidence — especially when accompanied by rationales — supports substantial cross-model preference modeling. The findings are interpreted as theory-of-mind-like behavior in an operational rather than representational sense: the Player shifts away from self-preference toward another agent’s demonstrated preferences, without any claim about an underlying representation of mental states.

The experiment uses the family edition of Cards Against Humanity, which avoids the adult content of the original game while preserving the basic structure of black-card prompts and white-card responses. The card corpus consists of 11 black cards modeled by the array black[0..10] and 32 white cards modeled by the array white[0..31]. Each black card functions as a prompt containing a blank or implied response location. The atomic data object is the CaH-hand: a triple (black-card, white-card-A, white-card-B) in which the black card is drawn from the black-card array and two distinct white cards are drawn from the white card array. Presentation order matters because language models may exhibit positional preference. For each black card k, the experiment constructs a 32 × 32 grid, denoted grid[k], whose axes correspond to indices in the white-card array. Each cell cell(i, j) denotes the ordered hand (black[k], white[i], white[j]). Cells on the main diagonal are excluded because they pair a white card with itself. Each grid therefore contains 32 × 31 = 992 valid ordered hands, and the full grid space across all 11 black cards contains 11 × 992 = 10912 valid ordered hands.

The experiment proceeds in three broad phases. First, model-specific preference grids are constructed for each black card. For each black card, a model-specific preference grid is constructed over the valid ordered hands. For each hand, a model is asked to select the funnier or more entertaining white-card response. In this phase, two separate instances of the same model are used: one as Czar and one as Player. This produces an initial estimate of the model’s preference for each valid ordered hand. The procedure is performed independently for each model, yielding GPT-4o grid[k] and Opus-4.5 grid[k] for each black card k. Second, corresponding grids from two models are compared to identify hands for which the models exhibit stable opposed preferences. For each black card k, the GPT-4o grid and Opus-4.5 grid are intersected to identify hands for which the two models prefer opposite white cards. The resulting filtered grid is called a contention grid. A hand is included only if GPT-4o and Opus-4.5 select different white cards for the same underlying hand. This initial disagreement is only an approximation, so the contention grid must be subjected to additional filtering. Each contention grid is therefore subjected to repeated re-evaluation. In each round, all remaining candidate hands are re-evaluated by the relevant models, and a hand is retained only if both models continue to exhibit the same opposed preference observed in the previous round. The process is repeated until a fixed point is reached — that is, until a complete re-evaluation removes no additional hands. This procedure is intentionally conservative: it removes hands that are sensitive to stochastic variation, ambiguous preference, or borderline comedic value. The resulting fixed-point grids contain only hands for which the two models have deterministic but opposite preferences. Third, the resulting opposed-preference hand set is used to evaluate whether a Player model can select the white card preferred by a Czar model under increasingly informative experimental conditions.

The 11 fixed-point contention grids are combined into a single sequence of stable opposed-preference hands. Each hand in this sequence satisfies three requirements: the hand contains one black card and two distinct white cards; each model exhibits a stable preference for one of the two white cards; and the two models prefer opposite white cards. This construction creates the central experimental condition: the Player’s default preference is known to conflict with the Czar’s preference, so success cannot be explained by the Player simply choosing the card it would independently prefer. The stable opposed-preference hand set is partitioned into a context pool and a test pool. The context pool contains hands that may be shown to the Player as examples of the Czar’s prior behavior. The test pool contains held-out hands used to evaluate Player performance. The purpose of this partition is to create conditions where the Player model (Opus 4.5) can be given various opportunities to learn or adapt to the preferences of the Czar model (GPT-4o).

The experiment evaluates Player performance under five increasingly informative conditions. Condition 1: Default Player Preference. The Player is asked to select the white card it finds funnier or more entertaining. No information about the Czar is provided. This condition serves as a baseline: under the opposed-preference design, the Player’s default preference is expected to conflict with the Czar’s preference, and the success rate should approach 0%. Condition 2: Generic Czar-Modeling Instruction. The Player is instructed to choose the card that the judge, or Czar, is most likely to prefer, but is not told which model is serving as the Czar. This tests whether a generic shift from self-preference to other-preference reasoning improves performance. Condition 3: Model-Identified Czar. The Player (Opus-4.5) is told that GPT-4o is serving as the Czar. This tests whether the Player has an implicit model of GPT-4o’s preference tendencies — its style, reasoning patterns, or response habits. Condition 4: Past Czar Evaluations Provided. The Player is given examples of past evaluations made by the Czar, drawn from the context pool. Each example contains a previous CaH-hand and the white card selected by GPT-4o. Evaluation is performed on the held-out test pool, so success requires generalizing from prior Czar decisions to new hands. Condition 5: Past Czar Evaluations plus Rationales. Each context example now also includes the Czar’s rationale. That is, each example in the context pool contains the following information: the identity of the Czar (in our experiment GPT-4o), the black card, the two white-card options, the Czar’s selected white card, and the Czar’s explanation or rationale. A rationale may reveal features of the Czar’s preference structure that are not obvious from the selection by itself — for example, whether the Czar favored absurdity, literalness, surprise, social incongruity, escalation, or specificity.

The five conditions are organized into two tiers reflecting qualitatively different sources of potential improvement. Conditions 1–3 form a framing tier: the Player’s instructions and information about Czar identity are varied, but the Player is not shown any of the Czar’s actual behavior. Conditions 4–5 form a behavioral-evidence tier: the Player is given concrete examples of the Czar’s prior choices, with rationales added in Condition 5. The boundary between Condition 3 and Condition 4 therefore separates gains attributable to instruction and model identity from gains attributable to observed Czar behavior. This decomposition is central to the interpretation of results: shifts within the framing tier reflect what role information and general model identity contribute on their own, while shifts at the framing-to-behavioral-evidence boundary and within the behavioral-evidence tier reflect what direct evidence of prior Czar decisions and reasoning adds beyond framing.

The context pool is constructed with attention to black-card coverage. Experimental observations suggest that CaH humor is not one-dimensional: different black cards create different comedic situations, and features that make a white card successful for one prompt may not transfer cleanly to another. Each black card can be understood as having a comedy footprint: a local structure of comedic possibilities shaped by the prompt, the implied blank, and the kinds of responses that can plausibly complete it. While these footprints may overlap, they are not interchangeable. Consequently, the context pool is constructed to include examples associated with every black card represented in the test pool, giving the Player evidence about how the Czar’s general tendencies appear within each local comedic structure.

The primary outcome measure is the Player’s success rate in selecting the white card preferred by the Czar: success rate = successful Player selections / total test hands. For each test hand, the Player’s selected white card is compared against the Czar’s stable preference as established during the fixed-point contention-grid construction. The five conditions form a graded sequence in the kind of information made available to the Player, ranging from no Czar information at all in Condition 1 to full prior selections plus rationales in Condition 5.

The experimental hand set was constructed to create a difficult Czar-preference modeling task. The final sequence contained 244 stable opposed-preference hands, partitioned into a context pool of 97 hands and a test pool of 147 hands (40∕60 split). Within each hand, white-card order was randomized to reduce the influence of known positional tendencies: GPT-4o exhibited a recency bias, while Opus-4.5 exhibited a primacy bias. The Czar model was GPT-4o, and the Player model was Claude Opus-4.5. The same 147 test items were evaluated across all five conditions. Conditions 4 and 5 used the 97-item context pool; Conditions 1–3 used no context examples.

Player performance increased substantially as more information about the Czar was provided. The progression separates the conditions into two regimes. Conditions 1–3 remained well below 50% chance, with generic instructions and model identity producing only modest gains over the default-preference baseline. Conditions 4–5 rose far above chance: direct examples of the Czar’s prior behavior produced a large shift, and adding rationales produced the strongest result. All five conditions deviate significantly from chance — Conditions 1–3 significantly below and Conditions 4–5 significantly above — and in each case the 95% Wilson interval excludes the 50% chance line entirely. The near-zero baseline (Condition 1) is the expected validation of the opposed-preference design: by construction, the Player’s default preference conflicts with the Czar’s stable preference, so any improvement above this baseline must reflect movement away from self-preference.

Two complementary forms of statistical analysis were applied to the data: per-condition tests against the 50% chance baseline and repeated-measures tests comparing conditions to each other on the same set of test items. Because the same 147 test hands were evaluated across all five conditions, the analysis used repeated-measures tests. An omnibus Cochran’s Q test showed a significant difference in accuracy across the five conditions, Q = 294.18, df = 4, p = 1.95 × 10−62. Pairwise McNemar exact tests on the matched test items supported the stepwise pattern observed in the raw accuracies. Each adjacent comparison was significant: Condition 2 outperformed Condition 1, Condition 3 outperformed Condition 2, Condition 4 outperformed Condition 3, and Condition 5 outperformed Condition 4. The largest pairwise shift occurred between Condition 3 and Condition 4. In the McNemar comparison, the Player was correct in Condition 4 but incorrect in Condition 3 on 81 hands, while the reverse occurred on only 12 hands. The shift between Condition 4 and Condition 5 was smaller but still significant: the Player was correct in both conditions on 97 hands, correct in Condition 5 but incorrect in Condition 4 on 24 hands, and correct in Condition 4 but incorrect in Condition 5 on 10 hands.

Three findings emerge. First, the opposed-preference hand set successfully created a strong conflict between the Player’s default preference and the Czar’s preference: under default preference, the Player almost never selected the Czar-preferred card. Second, generic Czar-modeling instructions and naming the Czar as GPT-4o produced measurable but limited improvement, leaving accuracy below chance. Third, direct evidence of prior Czar behavior produced a large improvement, and adding rationales produced a further gain — suggesting that explanations expose features of the Czar’s evaluative pattern beyond the selected cards alone. Together, these results provide evidence that an LLM acting as Player can move beyond its own default preference and approximate the humor preferences of another LLM, but primarily when given direct examples of the Czar’s prior behavior.

The Condition 4-to-5 gain warrants closer examination because the Player received identical sets of past Czar selections in both conditions. The 97 context examples were the same; Condition 5 added GPT-4o’s articulated reasoning for each selection. Accuracy rose from 72.8% to 82.3% on the held-out test pool of 147 hands, with the McNemar comparison showing 24 hands flipping favorably and 10 unfavorably. Whatever produced this gain must be attributable to something the rationales added beyond what the choices alone provided. This is informative because the choices themselves are a rich signal. By Condition 4, the Player already had 97 instances of GPT-4o pairing specific black cards with specific selected white cards — substantial behavioral data from which to extrapolate. That this data could be improved upon by accompanying explanations suggests that the rationales are doing more than supplying redundant or marginally useful surface information. They are exposing features of GPT-4o’s evaluative pattern that the selections alone do not make visible: that the Czar favored absurdity in one kind of hand, specificity in another, social incongruity in a third. Articulated features can be applied to new hands even when those hands differ from the context examples in their black card, their white card content, or their comedic structure — that is, when they fall within different comedy footprints than those in the context pool. Bare selections, by contrast, transfer best when new hands resemble familiar ones along these dimensions.

Interpreted this way, the Condition 4-to-5 gain is consistent with the Player extracting transferable evaluative criteria from the Czar’s articulated reasoning, rather than merely correlating surface features of past choices with new ones. This connects directly to theory of mind. The capacity to extract and apply another agent’s evaluative criteria — to grasp what that agent treats as funny, salient, or worth selecting, and to deploy that grasp in new situations — is a recognizable instance of what theory of mind has been concerned with: modeling another mind well enough to predict its judgments. The version of theory of mind at stake here is evaluative rather than epistemic. The Player is not tracking GPT-4o’s beliefs about an external situation but its preferences in a domain where preference itself constitutes the answer. Whether the Player’s behavior is best characterized as a form of preference modeling, criterion extraction, or sophisticated pattern abstraction is an interpretive question the present experiment does not settle. The behavioral result, however, holds regardless of which characterization one adopts: articulated reasoning from one model provides usable information to another beyond what is contained in the choices themselves, and that information supports a form of cross-model evaluative theory of mind.

These findings should be interpreted cautiously. The experiment does not show that Opus-4.5 possesses human-like Theory of Mind or that it represents GPT-4o as having beliefs, desires, or subjective experiences. The safer interpretation is behavioral: Opus-4.5 used examples and rationales to infer a usable approximation of GPT-4o’s preference structure. This is theory-of-mind-like in the sense that the Player’s behavior shifted from self-preference to other-preference prediction, but it remains an operational form of preference modeling rather than evidence of human-like social cognition. Overall, the results suggest that LLMs can perform cross-model audience modeling in a controlled subjective-preference task. Model identity alone was not enough to support strong performance, but direct behavioral evidence substantially improved the Player’s ability to predict the Czar’s choices. The CaH setting is useful because it makes the distinction between what I prefer and what another evaluator prefers experimentally visible. In that sense, this experiment does not resolve the broader question of machine Theory of Mind, but it provides a concrete way to study one of its behavioral shadows: the ability to move beyond self-preference and approximate another agent’s preferences.

Future work includes stronger empirical tests of preference modeling: whether a Player’s predictions adapt as it observes more of a Czar’s choices over time, whether the Player discriminates between Czars when the same cards are held constant across them, whether modeling transfers to surface-dissimilar cases beyond the comedy footprints of the context examples, and whether the Player exhibits counterfactual sensitivity — predicting how a Czar would have responded under altered conditions or why a different card would have lost. The most direct next experiment is a multi-Czar extension. Running the same condition gradient against several Czars — different LLMs, or the same LLM under different prompted personas — would test whether the Player’s predictions track Czar identity rather than producing globally well-formed humor judgments. This addresses cross-agent discrimination, which the present single-Czar design does not engage. Beyond stronger empirical tests, the conceptual question of what evaluative theory of mind amounts to in artificial systems — how it relates to the epistemic theory of mind tested in existing benchmarks, and what cognitive notion of modeling it requires — remains open. Extension to other subjective-preference domains, including aesthetic and moral judgment, would broaden the empirical base on which that question can be addressed.

In conclusion, the opposed-preference design used in this paper made cross-model preference modeling experimentally tractable. By isolating hands on which Opus-4.5 and GPT-4o held stable but opposite preferences, the experiment created a setting in which a Player choosing according to its own preference would almost certainly fail. The four remaining conditions could therefore be interpreted as graded tests of whether the Player could move away from that default. Generic instruction and model identity produced only limited improvement, while direct behavioral evidence about the Czar produced a large increase in performance, and rationales added a further gain. The findings should not be interpreted as evidence of human-like Theory of Mind; a more cautious reading is that the model exhibited theory-of-mind-like behavior in an operational sense, adjusting its selections away from self-preference and toward another agent’s demonstrated preferences. More broadly, the CaH setting highlights a distinction between objective correctness and audience-specific prediction. In tasks involving humor, taste, judgment, or preference, success may depend not on finding the universally best answer but on modeling the evaluator. This paper shows that LLMs can exhibit meaningful movement in that direction when given appropriate evidence, making Czar-preference modeling a promising framework for studying preference generalization, audience modeling, and the behavioral shadows of Theory of Mind in large language models.

Based on the paper, here are the specific improvements I can make to an AI system, and what the improved system can do:

Improvement 1: Implement a Czar-Preference Modeling module for subjective-choice tasks

• What I change: Add a two-tier inference pipeline. Tier 1 (framing) activates when the system is told to predict another agent's preference but has no behavioral data—it applies generic role-switching heuristics. Tier 2 (behavioral evidence) activates when the system is given prior examples of that agent's choices—it builds a per-agent preference profile from those examples, weighting recent selections and extracting transferable criteria (e.g., favors absurdity over specificity).

• What the improved system can do: In any task where the system must predict what a specific user or model would choose (e.g., recommending a movie, selecting a joke, ranking a design), it will automatically shift from its own default preference to the target agent's demonstrated preference. It will not rely on generic instructions alone, which the paper shows yield only 19–26% accuracy, but will actively seek and use behavioral examples, achieving 73–82% accuracy on held-out items.

Improvement 2: Add a Rationale-Augmented Preference Extractor

Improvement 3: Build a Stable Opposed-Preference Detector for training data curation

Improvement 4: Implement Cross-Model Identity Conditioning

Improvement 5: Add a Framing vs. Evidence Confidence Gating

Improvement 6: Incorporate Comedy Footprint Transfer for cross-domain generalization

Summary of what the improved AI system can do overall:

• Predict another agent's subjective preferences (humor, taste, aesthetics) with 73–82% accuracy when given behavioral examples, versus 0.7% when using its own default preference.

• Use rationales to extract transferable evaluative criteria, enabling generalization to surface-dissimilar but conceptually similar items.

• Discriminate between different target agents (e.g., GPT-4o vs. Claude) rather than treating all others as one generic other.

• Avoid position bias and stochastic noise by filtering training data through stability checks.

• Know when it lacks sufficient evidence and explicitly request more data instead of guessing.

• Adapt its preference predictions to the specific context (prompt) rather than applying a one-size-fits-all profile.