Paper deep dive
PREFER: From Player Experience to Personalized Game Refinement
Qi Chai, Zhang Zheng, Lu Yang, Zixun Sun, Hao Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 87%
Last extracted: 10/15/2026, 5:20:26 AM
Summary
The paper introduces PREFER, a framework that utilizes large language models (LLMs) as virtual playtesters to evaluate gameplay experience using shared psychometric dimensions (adapted from PXI and CORGIS). PREFER benchmarks model evaluations and guides persona-specific game refinement by identifying experience-preference relationships, selecting refinement objectives, and implementing rule changes via function-level contracts. Experiments demonstrate that PREFER achieves higher post-play preference ratings compared to baselines.
Entities (17)
Relation Signals (13)
Prefer → guides → Game Refinement
confidence 95% · PREFER... guide persona-specific game refinement.
Prefer → uses → Experience Dimensions
confidence 95% · PREFER uses shared experience dimensions to benchmark how models evaluate gameplay and to guide persona-specific game refinement.
Experience Dimensions → adaptedfrom → CORGIS
confidence 90% · We adapt experience dimensions from established psychometric questionnaires... PXI and CORGIS
Experience Dimensions → adaptedfrom → PXI
confidence 90% · We adapt experience dimensions from established psychometric questionnaires... PXI and CORGIS
Prefer → benchmarks → LLM Evaluation
confidence 90% · benchmark how models evaluate gameplay
Persona → guides → Rule Refinement
confidence 90% · For each persona, PREFER uses experience-preference relationships to select refinement objectives, revises game rules...
Prefer → outperforms → Baseline
confidence 90% · PREFER achieves 9.2% higher mean post-play preference ratings than the strongest evaluated baseline.
GLM → evaluatedby → Prefer
confidence 85% · We benchmark GLM... on 91 card games
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Understanding how large language models (LLMs) evaluate gameplay experience is important for using their feedback to guide game design. However, overall ratings alone provide limited insight into which aspects of experience are associated with these judgments and how they can inform concrete game changes. We introduce PREFER, a framework that uses a shared set of experience dimensions adapted from established psychometric questionnaires to benchmark how models evaluate gameplay and to guide persona-specific game refinement. Through these dimensions, we characterize the perspectives models offer as virtual playtesters by examining how their overall assessments relate to specific aspects of experience and to the gameplay they produce. To examine how the same dimensions can guide refinement for players with different preferences, we use personas to simulate distinct player profiles. For each persona, PREFER uses experience-preference relationships to select refinement objectives, revises game rules and their implementations, and playtests candidate games to assess experience changes and predicted preference gains. Experiments with five personas show that PREFER achieves 9.2% higher mean post-play preference ratings than the strongest evaluated baseline.
Tags
Links
- Source: https://arxiv.org/abs/2610.19213v1
- Canonical: https://arxiv.org/abs/2610.19213v1
Trouble viewing inline? Open PDF directly →
Full Text
76,981 characters extracted from source content.
Expand or collapse full text
PREFER: FROM PLAYER EXPERIENCE TO PERSONALIZED GAME REFINEMENT Qi Chai ∗ Zhang Zheng ∗ Lu Yang Zixun Sun Hao Wang † ABSTRACT Understanding how large language models (LLMs) evaluate gameplay experience is important for using their feedback to guide game design. However, overall ratings alone provide limited insight into which aspects of experience are associ- ated with these judgments and how they can inform concrete game changes. We introduce PREFER, a framework that uses a shared set of experience dimensions adapted from established psychometric questionnaires to benchmark how models evaluate gameplay and to guide persona-specific game refinement. Through these dimensions, we characterize the perspectives models offer as virtual playtesters by examining how their overall assessments relate to specific aspects of experience and to the gameplay they produce. To examine how the same dimensions can guide refinement for players with different preferences, we use personas to simulate distinct player profiles. For each persona, PREFER uses experience-preference relationships to select refinement objectives, revises game rules and their implemen- tations, and playtests candidate games to assess experience changes and predicted preference gains. Experiments with five personas show that PREFER achieves 9.2% higher mean post-play preference ratings than the strongest evaluated baseline. 1INTRODUCTION Large language models (LLMs) are increasingly used in game development, supporting the generation of playable prototypes and the evaluation of player experience (Sun et al., 2024; Lankes & Stöckl, 2023). Gameplay experience involves multiple aspects, such as challenge, mastery, and immersion (Sweetser & Wyeth, 2005; Vanden Abeele et al., 2020). Understanding how these aspects relate to models’ overall judgments is therefore important for characterizing the perspectives they offer as virtual playtesters and for using their feedback to guide game refinement. Established player-experience instruments, such as PXI and CORGIS, provide structured dimensions for describing functional, psychosocial, and challenge-related aspects of gameplay (Vanden Abeele et al., 2020; Denisova et al., 2020). Building on such instruments, Dutta et al. (2026) map player reviews to experience constructs and examine their associations with review scores. In parallel, MeepleLM generates ratings and critiques from game rulebooks under different player personas, illustrating how LLMs can provide feedback from multiple player perspectives (Li et al., 2026). Building on these advances, we ask how different LLMs evaluate the experience reflected in actual gameplay. An overall rating alone does not reveal whether models agree on specific aspects of an experience, or how those assessments relate to their overall judgments. Characterizing their perspec- tives therefore calls for comparing both levels of evaluation and examining how these judgments relate to the gameplay the models produce. Beyond this analysis, a further question is whether the same experience dimensions can guide game refinement for players with different preferences, by specifying which aspects to change and in which direction. To address these questions, we introduce PREFER, a framework that uses shared experience dimen- sions to benchmark how models evaluate gameplay and to guide persona-specific game refinement. We adapt experience dimensions from established psychometric questionnaires to assess gameplay trajectories, collecting overall ratings separately from dimension-level assessments. Different models evaluate the same trajectories, allowing us to examine how their overall judgments relate to specific ∗ Equal contribution. † Corresponding author. 1 arXiv:2610.19213v1 [cs.AI] 13 Oct 2026 a Game suite b Persona-guided refinement c Model benchmarking A ♠ 8 ♦ K ♣ Card gamesRules + code Playground sas′ StateActionNext state PersonasExperience targetsRule refinementReplay & select Efficiency Narrative Social System Thrill Curiosity + Autonomy + Challenge - Contract-guided edits BeforeAfter Personalized game Feedback Playing models Gameplay M1M2M3 No persona prompt A ♠ 8 ♦ K ♣ Trajectories × EfficiencyNarrativeSocialSystemThrill M1M2M3 Cross-model evaluation Persona evaluation Evaluation records Rate Rate ExperienceOverall ratingReasons D1D2D3 D4D5D6 D7D8D9 D10D11D12 1–10 Game: Actor: Evaluator: Scorer ♣ Figure 1: Overview of PREFER. (a) The game suite provides rules, code, and a playground. (b) Persona preferences guide rule refinement, implementation, and playtesting. (c) Models and persona- based evaluators assess the same gameplay trajectories using shared experience dimensions and overall ratings. aspects of experience and how their assessments agree or differ. We further compare their evaluations of trajectories produced through their own play and that of other models, and examine agreement in game-level rankings. However, players differ in what they value in a game (Tondello et al., 2019). To better simulate diverse real-world players, we follow MeepleLM (Li et al., 2026) in using personas to represent distinct player profiles. For each persona, PREFER relates experience dimensions to overall preference to select which aspects to change and in which direction. These objectives guide rule revisions, which are translated into function-level contracts that specify required behavior and guide implementation checks and repair. Candidate games are then playtested under the target persona, and acceptance requires both the intended experience changes and a positive predicted preference gain. Feedback from unsuccessful candidates guides further rule revisions. We evaluate PREFER on familiar card games and their rule variants. The model benchmark identifies shared patterns in game rankings alongside differences in how experience dimensions relate to overall ratings. Across five personas, PREFER achieves 9.2% higher mean post-play preference ratings than the strongest baseline evaluated. A separate implementation study also shows a higher success rate in producing runnable games and higher rule-consistency scores than baseline. Our contributions are threefold: • We introduce PREFER, a framework that adapts experience dimensions from psychometric questionnaires to gameplay trajectories, providing a shared basis for model evaluation and persona-specific game refinement. •We establish a model benchmark that characterizes the perspectives LLMs offer as virtual playtesters by comparing their experience judgments, overall ratings, and game rankings, and relating these evaluations to the gameplay they produce. •We develop a persona-specific game refinement method that uses experience-preference relationships to select modification targets, guides rule revisions with playtesting feedback, and supports implementation through function-level contracts and execution checks. 2RELATED WORK Game-Based Agent Evaluation. Games let researchers observe how agents follow rules, choose actions, and respond to other players over several turns. clembench tests instruction following and task success through dialogue games (Kranti et al., 2023). GameBench studies strategic reasoning across nine games (Costarelli et al., 2024). TextArena offers competitive text games that involve skills such 2 as negotiation and deception (Guertler et al., 2025). Our benchmark uses shared experience measures to compare models without persona prompts against reference players given fixed persona prompts. We compare their evaluation profiles on shared gameplay evidence and relate these judgments to the play the models produce. Personas and Player Experience. Persona prompts ask an agent to follow a given role or set of preferences. TRAIT measures model personality through multiple-choice questions about every- day situations (Lee et al., 2025). PERIL studies how persona prompts affect game strategies. It maps a model’s answers to personality questions into strategy parameters (Licato & Steinle, 2025). MeepleLM uses player personas to rate games and explain their judgments from rule descriptions (Li et al., 2026). PXI and CORGIS offer scales for player experience and perceived challenge (Van- den Abeele et al., 2020; Denisova et al., 2020). We use these scales to describe gameplay. Personas guide our game revisions and provide reference players for model comparison. Game Refinement and Verification.Game refinement requires both useful feedback and code that follows the revised rules. Self-Refine uses model feedback to improve generated outputs (Madaan et al., 2023) and systems (Yuksekgonul et al., 2025; Ye et al., 2026). Cardiverse creates card game variants and playable implementations (Li et al., 2025). Fly, Fail, Fix uses play records from a reinforcement learning agent to guide changes to game settings (Zook et al., 2025). For code repair, LDB checks execution step by step to locate errors (Zhong et al., 2024). PolyVer uses function contracts to verify programs written in different languages (Chen et al., 2025). In our method, a target player’s experience guides rule changes. Fixed function contracts guide code repair, and playtesting checks the resulting experience. 3METHOD PREFER uses shared experience dimensions to study how LLMs evaluate gameplay and to guide game refinement for a target persona. We first describe the experience assessment. We then explain how persona preferences guide rule changes, how function contracts guide their implementation, and how playtesting provides feedback for further revision. 3.1EXPERIENCE ASSESSMENT We describe gameplay through selected dimensions from PXI (Vanden Abeele et al., 2020) and CORGIS (Denisova et al., 2020). These dimensions cover how a game supports and engages the player, as well as the demands of play. For example, challenge fit measures whether difficulty suits the player. Cognitive challenge measures the reasoning and planning required during play. We keep the dimension definitions and scoring scales fixed across models and personas. Appendix B gives the full definitions and scoring details. LetRdenote the rulebook andτa completed gameplay trajectory. The trajectory records the focal player’s observations, available actions, chosen actions, and outcomes. An LLM evaluator uses this evidence to rate each experience dimension. A separate assessment gives an overall rating of the same playthrough. We write the two assessments as x =E m (R,τ,p), y =S m (R,τ,p),(1) wheremis the evaluation model,x∈ R d contains thedexperience scores, andyis the overall rating. The optional personapspecifies the player’s preferences. Each experience score includes a short explanation based on the trajectory. The two assessments do not receive each other’s outputs, soyis not calculated from x. For model comparison, we omit the persona prompt and compare judgments of the same gameplay evidence. The pairs(x,y)let us examine how each model’s dimension scores relate to its overall ratings. For persona-guided refinement, the player agent and both assessments use the same target persona p. In this setting, y measures the persona’s preference for the played game. 3.2RULE REFINEMENT Rule refinement starts from a game with rulebookR, codeC, and a target personap. We collect paired experience and preference measurements across games under this persona. A regularized linear predictorf p relates the experience vector to the preference score. We fit the predictor on other games before using it to refine the current game. Appendix B.1 gives the predictor and fitting procedure. 3 Contract designer Programmer Evidence-guided locate Rule edit Choose draw order A♠6♠9♣ run f1f2f3 Contracts A Preconditions G Postconditions Frozen Code C′ Local: check behavior Valid A f2 Check G Joint: check prerequisites f1f2 Providers Check A f3 Pass Checked code Fail Localize G fails → f2 A fails → f1, f2 Min-cost scope f1f2f3 Cover failures Mark f2 Trace Re-run Calling Graph Code Implement Figure 2: Contract-guided rule implementation. Revised rules are translated into function contracts that guide code edits and execution checks. Failure traces guide the selection of functions to repair, and the updated code is rechecked under the same contracts. The predictor assigns a weightw pj to each experience dimensionj. In the original scoring units, this weight estimates the preference change for a one-point increase in dimensionj, with other dimensions fixed. We use the sign ofw pj to choose the direction of change. A positive weight suggests increasing the dimension, and a negative weight suggests decreasing it. We denote this direction by d pj ∈−1, +1 and skip dimensions with zero weight. Given the current experience vectorx, we remove targets without enough room for change on the scoring scale. We rank the remaining targets by|w pj |. This ranking compares the predicted gains from equal-sized changes in individual dimensions. It sets the search order, while playtesting checks the effects of an actual rule change. For each selected target, an LLM proposes a revised rulebookR ′ usingR,p, and the target dimension and direction. The proposal changes a game mechanic while keeping the game objective and unaffected rules fixed. The next stage implements this rulebook and checks the resulting code. 3.3RULE IMPLEMENTATION Contract construction. We compareR ′ withRto identify changed requirements and relevant rules that must remain unchanged. Each requirement keeps a link to its source rule text. We also extract function definitions and direct calls from the code’s abstract syntax tree. This information helps assign each required outcome to an owner function and identify functions that may supply its prerequisites. Following Chen et al. (2025), we express function responsibilities through assumptions and guarantees. We add provider links and timing conditions to form each contract clause κ = (f,A,G,P,α,δ),(2) wherefis the owner function,Ais its required entry condition, andGis its required outcome. The setPlists possible prerequisite providers. The triggerαactivates the obligation, andδsets its deadline. For argumentsu, statesbefore the call, states ′ after the call, and return valuer, the clause requires A(u,s) =⇒ G(s,s ′ ,r).(3) BothAandGare executable checks. We check that the plan covers the rule requirements, has valid function links, and can execute its predicates. Semantic review provides advice on whether the clauses match the rules. Once the plan is accepted, its predicates and clause identities stay fixed during code generation and repair. Local and joint checks. An implementation model updates the affected function bodies using the revised rules and accepted contracts. Local checks test an owner function in isolation. Each test supplies arguments and state that satisfyA, independently of the candidate provider functions. It then calls the owner and testsG. A failed guarantee or an exception under valid entry conditions identifies an owner failure. If the test input fails A, the clause remains untested. 4 0.12 0.44 0.57 0.49 0.06 0.45 0.42 0.47 0.56 0.03 0.51 0.15 0.06 0.20 0.41 0.18 0.42 0.32 0.32 0.28 0.15 0.01 0.19 0.03 -0.01 0.17 0.48 0.27 0.00 0.26 0.18 0.15 0.21 -0.02 0.20 -0.03 0.03 0.24 0.43 0.33 0.38 0.31 0.31 0.38 0.37 0.34 0.22 0.09 0.12 0.35 0.56 0.42 0.20 0.49 0.41 0.54 0.49 0.11 0.32 0.08 Goals Autonomy Mastery Curiosity Challenge fit CognitionDecisions Meaning Immersion Emotion Feedback Control GLM DeepSeek Kimi MiniMax Qwen 12D −0.650.000.65 Spearman correlation Efficiency Narrative Social System Thrill GLM DeepSeek Kimi MiniMax Qwen 24.3 22.5 23.0 21.7 24.5 22.4 22.3 24.2 22.0 21.8 10.1 13.2 9.9 8.3 7.7 25.3 23.3 25.2 22.1 25.8 17.9 18.7 17.7 25.9 20.2 (a)Dimension associations(b)Persona similarity shares Figure 3: Experience dimensions characterize model evaluation. (a) Associations with overall ratings reveal shared emphasis on mastery and differences across other dimensions. (b) Normalized similarity shares summarize resemblance to the five persona profiles; bold marks the largest share for each model. Joint checks test whether gameplay supplies the required entry conditions. A runtime monitor follows each obligation from its triggerα. It checksAwhen the owner is called and reports a failure if Ais false. A required call missing at deadlineδalso causes a failure. The monitor uses provider calls observed during this period to identify possible causes. For both checks, a lack of evidence is recorded as unknown. Repair scope. We use execution feedback to select which functions to repair (Zhong et al., 2024; Xia et al., 2025). Each supported failureigives a setE i of candidate functions. A local failure gives E i =f. A joint failure gives the providers supported by the execution record. We select a set that covers every failure while limiting edit cost S ∗ = arg min S⊆F X f∈S c f subject to S∩ E i ̸= ∅ for all i,(4) whereFis the set of editable functions andc f > 0is the number of source lines inf. Failures without supported candidate functions remain unresolved. LetUcontain the failure sets not yet covered. IfUis nonempty, we choose a set with the fewest candidates, E † ∈ arg min E∈U |E|. The minimum remaining cost is V (U) = 0,U = ∅, min f∈E † [c f + V (E ∈U | f /∈ E)], otherwise. (5) Selectingfcovers every remaining set that contains it. We solve this recurrence by branch-and-bound search. A branch is dropped when its accumulated cost exceeds the best complete solution. A completed search returns S ∗ . Appendix D gives the search limits and tie-breaking rules. The repair model receives the rules, fixed contracts, code, failure records, and selected functions. It changes only the selected function bodies and preserves their signatures. We keep a patch only if rechecking improves the results without new failures or loss of earlier evidence. Repair stops when the required checks pass, no supported repair scope remains, or the budget is exhausted. Only candidates that pass the required checks proceed to playtesting. 3.4PLAYTESTING AND FEEDBACK The player agent plays each candidate with the target personap. We assess the resulting trajectory using the same dimensions and scoring procedure as for the original game. Letxbe the original experience vector andx ′ k the vector for candidatek. A candidate needs valid gameplay and complete experience scores before it can be accepted. A rule change can improve the target dimension but harm other parts of the experience. We therefore check both the target change and the predicted preference gain. For candidatektargeting dimension 5 00.30.6 Goals Autonomy Mastery Curiosity Challenge fit Cognition Decisions Meaning Immersion Emotion Feedback Control 0.17 0.34 0.50 0.40 0.45 0.40 0.40 0.53 0.54 0.05 0.45 0.28 Efficiency 00.30.6 -0.11 0.34 0.39 0.38 0.29 0.30 0.30 0.14 0.16 0.04 0.31 0.10 Narrative 00.30.6 0.18 0.11 0.07 0.16 0.22 -0.04 0.07 0.08 -0.02 0.02 0.24 0.14 Social 00.30.6 0.13 0.33 0.48 0.40 0.33 0.45 0.46 0.36 0.48 -0.04 0.32 0.05 System 00.30.6 0.14 0.29 0.38 0.49 0.57 0.35 0.45 0.57 0.56 0.50 0.47 0.09 Thrill 012 Mean baseline Shared model Persona-specific model 1.46 1.08 1.01 Efficiency 012 0.50 0.47 0.39 Narrative 012 0.79 0.56 0.55 Social 012 0.87 0.61 0.55 System 012 1.82 1.21 1.21 Thrill (a)Persona profiles Spearman correlation (b)Preference prediction error Held-out MAE (lower is better) Figure 4: Persona evaluation profiles and preference prediction. (a) Experience dimensions relate differently to each persona’s ratings. (b) Preference prediction from each persona’s own experience scores, comparing a persona-mean baseline, shared dimension slopes with persona indicators, and separate persona models. Lower MAE indicates better prediction; intervals show uncertainty. j k , acceptance requires ( d pj k x ′ k,j k − x j k ≥ ε, f p (x ′ k )− f p (x) > 0, (6) whereε > 0is the minimum required change in the target dimension. The first condition checks progress in the selected direction. The second uses all experience dimensions to estimate the total preference change. We try targets in their ranked order and generate each candidate from the same original game. A failed implementation, incomplete playtest, or rejected candidate leads to the next target. We accept the first candidate that passes all checks. If no candidate passes within the search budget, we keep the original game and enter the feedback stage. Following Madaan et al. (2023), we use feedback to revise a failed proposal. We select the highest- priority rejected candidate with valid gameplay and complete experience scores. The feedback includes its rules, score changes, acceptance decision, and an explanation based on the trajectory. The generator changes the mechanic while keeping the persona, target dimension, and direction fixed. The new proposal goes through the same implementation checks and playtesting. Equation 6 compares it with the original game. If no suitable candidate exists or the feedback revision also fails, we retain the original game. 4EXPERIMENTS Our experiments connect three questions: what experience dimensions reveal about model evaluation, how persona-specific preference relationships define refinement targets, and whether the resulting game revisions receive higher post-play preference ratings. We examine each connection separately before following the complete process in a concrete case. 4.1EXPERIMENTAL SETUP Model evaluation. We benchmark GLM (Zeng et al., 2026), DeepSeek (Xu et al., 2026), Kimi (Team et al., 2026), MiniMax (Chen et al., 2026), and Qwen (Qwen Team, 2026) on 91 card games with fixed rules and implementations. The models play against random opponents without persona prompts. Five neutral judges and five persona references assess the recorded trajectories with actor identities hidden. Overall ratings and 12 experience ratings are collected in separate calls. Appendix A details model configurations and analysis sets. 6 Table 1: Agreement between adjacent game groups. Group Gap is the mean rating difference, and Pair Agreement is the percentage of correctly ordered cross-group pairs. Judge ModelGLM-5.3 DeepSeek-4.1flash Kimi-K2.7 MiniMax-M2.5 Qwen3.8-Max Group Gap1.871.101.691.611.59 Pair Agreement96.078.488.091.593.7 GLMDeepSeekKimiMiniMaxQwen 7.0 5.46.8 7.2 5.2 Euchre 7.06.67.0 6.2 5.8 Pinochle 2 7.0 6.06.86.8 5.8Rummy Rum 7.07.07.0 7.0 6.6Sixty Six 6.8 6.66.8 7.4 5.8 500 Rum 6.0 4.2 5.0 6.0 4.6 Peep Nap 5.8 5.0 4.6 6.24.8Play Or Pay 5.6 4.6 6.0 5.4 4.8Anaconda Poker 5.4 5.0 5.0 6.2 4.2 Palace 5.2 5.0 5.2 5.4 5.0 Sir Garnet 3.03.03.0 2.0 2.0Bingo 3.03.03.03.02.0Left Center Right 3.03.03.03.03.0Ship Captain And Crew 2.82.02.42.4 1.6 Klondike 2.81.82.2 2.6 1.8 Napoleon At St Helena Pooled-score group HighMiddleLow Figure 5: Shared game preferences and local ranking differences. Persona prediction and refinement. We adopt five profiles from MeepleLM (Li et al., 2026): Efficiency Essentialist, Narrative Architect, Social Lubricator, System Purist, and Thrill Seeker. They provide reference evaluations for model comparison and serve as target players in game refinement. We validate preference predictions on held-out games and assess game revisions through persona playtesting and code implementation checks. Appendix B describes the evaluation protocols. 4.2MODEL EVALUATION THROUGH EXPERIENCE DIMENSIONS We use experience dimensions to characterize how models evaluate recorded gameplay. Figure 3 presents two complementary views: the associations between individual dimensions and overall ratings, and the similarity of these association profiles to persona references. We first examine whether models broadly agree on which games they rate highly. With complete rankings, four judges define high-, middle-, and low-rated groups for the fifth judge to compare. Each group spans ten rank positions. Each judge is held out in turn. All five favor the higher-rated group. Averaging the high–middle and middle–low comparisons gives mean gaps of 1.10-1.87 points and pairwise agreement of 78.4-96.0%, shown in Table 1. Figure 5 illustrates this pattern with the highest, middle, and lowest five games by pooled score. Each column orders the displayed games by one judge’s mean rating over actor sources, and lines connect the same game across judges. High-scoring games generally remain above low-scoring ones. At the level of overall ratings, the models show broad agreement. We then compare how their overall ratings relate to experience dimensions and use persona references to interpret the resulting association profiles. Figure 3(a) shows positive mastery associations across models, alongside differences in immersion and challenge fit. The displayed mastery correlations range from 0.41 to 0.57, placing mastery among the two largest correlations for every model. The other prominent associations differ: GLM has an immersion correlation of 0.56, DeepSeek a challenge-fit correlation of 0.42, and Qwen a meaning correlation of 0.54. These patterns identify a common association with mastery while showing that the remaining dimensions relate differently to each model’s overall ratings. The persona references provide a way to interpret these differences in evaluation emphasis. Figure 3(b) expresses these evaluation patterns relative to the five persona references. System has the largest estimated similarity share for GLM, DeepSeek, Kimi, and Qwen, while Thrill has the largest share for 7 Table 2: Refinement and implementation outcomes. Original games provide the preference reference. Bold marks the best method value, including ties. MethodPreference Rate Success Rate (%) Execution Rate (%) ECon PCon Original games2.927– Reflexion (Shinn et al., 2023)2.78783.3399.960.656 6.514 PromptBreeder (Fernando et al., 2024)2.74073.3399.860.670 5.407 Cardiverse (Li et al., 2025)2.82076.6799.520.658 5.991 ReCode (Yu et al., 2025)3.05383.3399.960.659 6.648 PREFER3.33386.6799.960.661 7.377 MiniMax. Similarity describes a combination of evaluation emphases for each model. For example, MiniMax’s shares are 25.9% for Thrill, 22.0% for Narrative, and 22.1% for System. These shares describe relative resemblance between evaluation profiles, without implying positive raw correlations. Social has the smallest normalized share for every neutral model. DeepSeek yields the largest share for Social, 13.2%, whereas the corresponding values for the other models range from 7.7% to 10.1%. This suggests that the pool of neutral judges may provide uneven coverage of different reference perspectives, highlighting the importance of persona-based perspectives. 4.3VALIDATING EXPERIENCE-BASED PREFERENCE MODELS The persona profiles in Figure 4(a) characterize differences in evaluation emphasis on common benchmark inputs. For example, emotional challenge is positively associated with Thrill’s ratings and negatively associated with System’s. These patterns motivate preference predictors that combine experience dimensions for each persona. We next examine whether the same experience dimensions predict preference when games are played under the target persona. This analysis uses a separate dataset of 455 persona-played observations from the same 91 games and five personas, with experience scores and preference ratings collected separately. Preference prediction.We compare separate regularized linear predictors for each persona with a persona-mean baseline and a shared predictor that uses common dimension coefficients with persona- dependent score levels. Figure 4(b) reports persona-wise prediction errors. Averaged equally over the five personas, the separate predictors achieve a mean absolute error (MAE) of 0.740, compared with 1.090 for the persona-mean baseline and 0.785 for the shared predictor. As shown in Figure 4(b), our persona-specific experience-based preference predictor has lower MAE than the shared predictor in four of the five personas, with nearly identical errors for Thrill. Appendix B.1 gives the procedures. Experience and persona ablations. Table 3 compares the experience sources and persona con- ditioning using identical observations and held-out-game folds. MAE andR 2 are averaged over personas; Order measures accuracy for within-persona pairs with different observed ratings, with both games held out. The nine PXI dimensions achieve MAE 0.761 and ordering accuracy 81.8%, close to the combined representation’s 0.740 and 82.0%. Using only the three CORGIS challenge dimensions yields MAE 0.905 and ordering accuracy 72.6%. Removing both sources gives the persona-mean baseline. The - persona ablation retains all 12 dimensions but shares both coefficients and intercept across personas, without persona indicators (MAE 0.880; Order 81.3%). PREFER uses the relationships fitted on the original 49 games to select which dimensions to change and in which direction for each persona, guiding the game revisions evaluated in Section 4.4. 4.4POST-PLAY PREFERENCE AND IMPLEMENTATION QUALITY We compare PREFER with Cardiverse’s game-generation workflow (Li et al., 2025), Reflexion’s self-feedback refinement (Shinn et al., 2023), ReCode’s bounded recursive planning (Yu et al., 2025), and PromptBreeder’s mutation (Fernando et al., 2024). We report mean post-play preference and implementation metrics in Cardiverse (Li et al., 2025): Success Rate, Execution Rate, code-rulebook embedding similarity (ECon), and play-rule consistency (PCon). Appendices B and D detail the adaptations, datasets, and protocols. As shown in Table 2, PREFER achieves a mean preference of 3.333, a paired increase of 0.407 over the original games and 9.2% above the best baseline’s mean of 3.053. It also achieves the highest implementation success rate (86.67%) and PCon (7.377), with execution rate tied at 99.96%. The 8 Table 3: Experience and persona ablations in preference prediction. PredictorMAE↓ R 2 Order(%) PREFER0.740 0.43582.0 - CORGIS0.761 0.42681.8 - persona0.880 0.03081.3 - PXI0.905 0.22272.6 - CORGIS&PXI 1.090 -0.02250.0 Table 4: Mean paired preference change by persona. PersonaDirected refinement With feedback Efficiency+0.333+0.400 Narrative+0.067+0.067 Social+0.233+0.267 System+0.100+0.133 Thrill+1.167+1.167 All personas+0.380+0.407 8 8 Game: Crazy Eights Persona: Thrill Seeker Design target: Challenge fit↑ Diagnose the failure Redesign the decisionVerify in play Choice + immediate consequence First attempt Long drawReveal Challenge fit (1–5) 22 OriginalFirst rule Too little pressure in draw decisions Gameplay evidence Gamble Draw · max 3 draws Draw a card Playable Not playable 3rd draw Not playable Draw 1 or 2 Play it now Bust MAX Highest value Stock bottom Face-down Stop or Continue Draw again Challenge fit (1–5) 23 OriginalRevised Predicted gain +3.28 Selected Rated after selection Preference (1–10) 26 OriginalRevised Figure 6: Feedback-guided revision in Crazy Eights. The target persona and experience dimension remain fixed while playtesting feedback guides a new drawing mechanic. different method rankings under ECon and PCon highlight the value of checking actual gameplay alongside code-rulebook similarity. Table 4 reports paired preference changes from the original games for both directed refinement and its feedback extension, averaged over 30 games per persona. Directed refinement achieves a mean gain of 0.380 points. Persona-level mean gains range from 0.067 for Narrative to 1.167 for Thrill. These differences may reflect how readily persona preferences can be addressed through local rule edits. Such edits can reshape the immediate risks and consequences relevant to Thrill, but may offer fewer opportunities to develop thematic progression for Narrative or improve strategic control and balance for System. The feedback extension yields a small additional improvement, increasing the mean paired preference gain from 0.380 to 0.407 points, an additional 0.027 points. Overall, PREFER improves mean post-play preference in the evaluated card games, with gains varying across personas. 4.5CASE STUDY: FROM AN EXPERIENCE TARGET TO A RULE CHANGE Figure 6 follows a Crazy Eights revision for Thrill Seeker with a fixed challenge-fit target. Revealing a player’s hand after extended drawing leaves challenge fit unchanged. Feedback then guides a stop-or-continue drawing mechanic with a three-draw limit. In the resulting playtest, challenge fit rises from 2 to 3 and post-selection preference from 2 to 6. The case illustrates how feedback guides a concrete mechanic change toward the same experience target. Appendix C.1 gives the full rule changes and evaluation record. 5CONCLUSION PREFER uses shared experience dimensions to connect model evaluation with persona-guided game refinement. The dimensions reveal differences hidden by overall ratings, predict persona preference on held-out games, and define concrete modification targets. Delivered revisions receive higher mean post-play preference than the original games and evaluated baselines. The implementation evaluation also reports higher success and play-rule consistency. Together, these results demonstrate how experience measurements can inform concrete, executable game revisions. 9 AI USE STATEMENT Generative AI tools assisted manuscript polishing, game code generation, and icon creation. The authors reviewed the manuscript content, checked the generated game code for executability, and reviewed the use of the generated icons. The use of LLMs as game-playing agents, experience evaluators, persona-conditioned judges, and rule and code generators within the research is described in Sections 3 and 4. The authors take responsibility for the final content of this work, including AI-assisted text, code, and visual assets. REFERENCES Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, Binyang Jiang, Boji Dan, Changhao Zhang, Changqing Yu, Chao Wang, Cheng Ma, et al. The minimax-m2 series: Mini activations unleashing max real-world intelligence. arXiv preprint arXiv:2605.26494, 2026. Pei-Wei Chen, Shaokai Lin, Adwait Godbole, Ramneet Singh, Elizabeth Polgreen, Edward A Lee, and Sanjit A Seshia. Polyver: A compositional approach for polyglot system modeling and verification. In Proceedings of the 25th Conference on Formal Methods in Computer-Aided Design–FMCAD, 2025. Anthony Costarelli, Mat Allen, Roman Hauksson, Grace Sodunke, Suhas Hariharan, Carlson Cheng, Wenjie Li, Joshua Clymer, and Arjun Yadav. Gamebench: Evaluating strategic reasoning abilities of llm agents. arXiv preprint arXiv:2406.06613, 2024. Alena Denisova, Paul Cairns, Christian Guckelsberger, and David Zendle. Measuring perceived challenge in digital games: Development & validation of the challenge originating from recent gameplay interaction scale (corgis). International Journal of Human-Computer Studies, 137: 102383, 2020. Supriya Dutta, Joel Oksanen, Jaakko Väkevä, Shamit Ahmed, Markus Kirjonen, and Perttu Hämäläi- nen. Mining player experience trends from game reviews using large language models. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–29, 2026. Chrisantha Fernando, Dylan Sunil Banarse, Henryk Michalewski, Simon Osindero, and Tim Rock- täschel. Promptbreeder: Self-referential self-improvement via prompt evolution. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 13481–13544. PMLR, 2024. URLhttps://proceedings.mlr. press/v235/fernando24a.html. Leon Guertler, Bobby Cheng, Simon Yu, Bo Liu, Leshem Choshen, and Cheston Tan. Textarena. arXiv preprint arXiv:2504.11442, 2025. Chalamalasetti Kranti, Jana Götze, Sherzod Hakimov, Brielen Madureira, Philipp Sadler, and David Schlangen. clembench: Using game play to evaluate chat-optimized language models as conversa- tional agents. In Proceedings of the 2023 conference on empirical methods in natural language processing, p. 11174–11219, 2023. Michael Lankes and Andreas Stöckl. Game reviews reviewed: A game designer’s perspective on ai-generated game review analyses. In 2023 IEEE conference on games (CoG), p. 1–8. IEEE, 2023. Seungbeen Lee, Seungwon Lim, Seungju Han, Giyeong Oh, Hyungjoo Chae, Jiwan Chung, Minju Kim, Beong-woo Kwak, Yeonsoo Lee, Dongha Lee, et al. Do llms have distinct and consistent personality? trait: Personality testset designed for llms with psychometrics. In Findings of the Association for Computational Linguistics: NAACL 2025, p. 8412–8452, 2025. Danrui Li, Sen Zhang, Samuel S Sohn, Kaidong Hu, Muhammad Usman, and Mubbasir Kapadia. Cardiverse: Harnessing llms for novel card game prototyping. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 29723–29750, 2025. 10 Zizhen Li, Chuanhao Li, Yibin Wang, Jianwen Sun, Yukang Feng, Jiaxin Ai, Fanrui Zhang, Mingzhu Sun, Yifei Huang, and Kaipeng Zhang. Meeplelm: A virtual playtester simulating diverse subjective experiences. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 18678–18722, 2026. John Licato and Stephen Steinle. Do persona-infused llms affect performance in a strategic reasoning game? In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, p. 3497–3528, 2025. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36:46534–46594, 2023. Qwen Team. Qwen3.8-Max: A new bar for coding and cowork, August 2026. URLhttps: //qwen.ai/blog?id=qwen3.8. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36:8634–8652, 2023. Fan-Yun Sun, SI Harini, Angela Yi, Yihan Zhou, Alex Zook, Jonathan Tremblay, Logan Cross, Jiajun Wu, and Nick Haber. Factorsim: Generative simulation via factorized representation. Advances in Neural Information Processing Systems, 37:87438–87472, 2024. Penelope Sweetser and Peta Wyeth. Gameflow: a model for evaluating player enjoyment in games. Computers in Entertainment (CIE), 3(3):3–3, 2005. Kimi Team, Yifan Bai, Yiping Bao, Y. Charles, and Cheng Chen et al. Kimi k2: Open agentic intelligence, 2026. URL https://arxiv.org/abs/2507.20534. Gustavo F Tondello, Karina Arrambide, Giovanni Ribeiro, Andrew Jian-lan Cen, and Lennart E Nacke. “i don’t fit into a single type”: A trait model and scale of game playing preferences. In Ifip conference on human-computer interaction, p. 375–395. Springer, 2019. Vero Vanden Abeele, Katta Spiel, Lennart Nacke, Daniel Johnson, and Kathrin Gerling. Development and validation of the player experience inventory: A scale to measure player experiences at the level of functional and psychosocial consequences. International Journal of Human-Computer Studies, 135:102370, 2020. Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. Demystifying llm-based software engineering agents. Proceedings of the ACM on Software Engineering, 2(FSE):801–824, 2025. Anyi Xu, B Li, Bangcai Lin, Bing Xue, BingCheng Xian, Bingzheng Xu, Bochao Wu, Bowei Zhang, Boyi Deng, C Yu, et al. Deepseek-v4. 1-flash: Pushing the limits of kv cache compression. arXiv preprint arXiv:2609.19969, 2026. Deheng Ye, Zheng Zhang, Hao Wang, and Chunyan Miao. A survey on ai for ai: When the improver becomes the improvee. Preprints, September 2026. doi: 10.20944/preprints202609.0591.v2. URL https://doi.org/10.20944/preprints202609.0591.v2. Zhaoyang Yu, Jiayi Zhang, Huixue Su, Yufan Zhao, Yifan Wu, Mingyi Deng, Jinyu Xiang, Yizhang Lin, Lingxiao Tang, Yuyu Luo, Bang Liu, and Chenglin Wu. ReCode: Unify plan and action for universal granularity control. arXiv preprint arXiv:2510.23564, 2025. URLhttps://arxiv. org/abs/2510.23564. Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. Optimizing generative ai by backpropagating language model feedback. Nature, 639(8055):609–616, 2025. Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763, 2026. 11 Li Zhong, Zilong Wang, and Jingbo Shang. Debug like a human: A large language model debugger via verifying runtime execution step by step. In Findings of the Association for Computational Linguistics: ACL 2024, p. 851–870, 2024. Alex Zook, Josef Spjut, and Jonathan Tremblay. Fly, fail, fix: Iterative game repair with reinforcement learning and large multimodal models. arXiv preprint arXiv:2507.12666, 2025. 12 ABENCHMARK DETAILS A.1DATA, EVALUATORS, AND ANALYSIS SETS The neutral benchmark contains 91 fixed games and five actor configurations. We match requests by the exact rule and focal-player evidence texts, deduplicate identical inputs, and require every instrument needed for a comparison. The common set has 363 unique inputs from 91 games with overall and 12-dimensional ratings from all five neutral judges and all five persona references. Overall replies with insufficient evidence remain missing; they are not replaced by zero. Persona scoring includes the completed supplementary questionnaires. Completed play includes trajectories stopped at the decision or time limit. The focal-player evidence records visible states, legal actions, selected actions, and outcomes. Initial play fixes seat 0 and seed 0, with random agents at other seats, limits of 500 focal decisions and four hours, and at most one model-selected recovery from an invalid action. Evaluation requests hide actor identity and collect dimensions and overall scores separately. The overall prompt concerns the recorded experience, excluding replay intention and recommendation. The actor configurations are GLM-5.3, DeepSeek-V4.1-Flash, Kimi-K2.7-Code, MiniMax-M2.5, and Qwen3.8-Max. The Qwen judge is Qwen3.8-Max-0902. The original external-rating batch requests temperature 0 for Kimi, DeepSeek, MiniMax and Qwen. Reasoning settings and output limits differ across services. A.2DIMENSION ASSOCIATIONS AND WITHIN-GAME CHECKS For each evaluator, we correlate its dimension ratings with its separately obtained overall scores, using Spearman correlation with average ranks for ties. Table 5 gives the values underlying the model rings and persona profiles in the main text. These pooled associations can reflect differences between games. To inspect variation within games, we rank scores and dimensions across the common set, subtract the corresponding game means from their ranks, and correlate the centered ranks. Table 6 reports these within-game associations. Intervals use 2,000 whole-game bootstrap samples. Simultaneous intervals account for multiple model-pair and dimension comparisons. Table 5: Associations between experience dimensions and overall ratings. Spearman correlations characterize evaluation emphasis across models and personas. E, N, S, Y, and T denote Efficiency, Narrative, Social, System, and Thrill. DimensionGLMDS Kimi M QwenENSYT Goals/rules0.12 0.06 -0.01 0.030.12 0.17 -0.110.180.13 0.14 Autonomy0.44 0.200.17 0.240.35 0.340.340.110.33 0.29 Mastery0.57 0.410.48 0.430.56 0.500.390.070.48 0.38 Curiosity0.49 0.180.27 0.330.42 0.400.380.160.40 0.49 Challenge fit0.06 0.420.00 0.380.20 0.450.290.220.33 0.57 Cognitive challenge0.45 0.320.26 0.310.49 0.400.30 -0.040.45 0.35 Decision challenge0.42 0.320.18 0.310.41 0.400.300.070.46 0.45 Meaning0.47 0.280.15 0.380.54 0.530.140.080.36 0.57 Immersion0.56 0.150.21 0.370.49 0.540.16 -0.020.48 0.56 Emotional challenge0.03 0.01 -0.02 0.340.11 0.050.040.02 -0.04 0.50 Progress feedback0.51 0.190.20 0.220.32 0.450.310.240.32 0.47 Ease of control0.15 0.03 -0.03 0.090.08 0.280.100.140.05 0.09 Model and persona profiles.Across the shared inputs, mastery is positively associated with overall ratings for every model (ρ = 0.413-0.574). Immersion has correlation 0.560 for GLM, compared with 0.148 for DeepSeek and 0.212 for Kimi. Challenge fit has correlations 0.419 for DeepSeek and 0.380 for MiniMax, compared with−0.003for Kimi. For the Thrill persona, challenge fit, meaning, immersion, and emotional challenge correlate with overall ratings at 0.567, 0.571, 0.562, and 0.496, respectively; System’s emotional-challenge correlation is−0.044. These values describe pooled associations across the recorded games and trajectories. 13 Table 6: Experience associations within games. Correlations of game-centered ranks describe how experience ratings and overall judgments vary among trajectories of the same game. DimensionGLM DeepSeekKimi MiniMaxQwen Goals/rules0.174-0.0010.1140.0310.139 Autonomy0.1320.1500.0250.1170.144 Mastery0.3210.2330.2450.2100.270 Curiosity0.2360.0420.1210.1220.062 Challenge fit0.0800.251 -0.0050.134 -0.036 Cognitive challenge0.2100.1120.0190.0900.152 Decision challenge0.1360.1750.0040.1040.046 Meaning0.2590.1350.0590.1020.211 Immersion0.2610.0900.0350.0950.140 Emotional challenge -0.0120.0150.0120.134 -0.001 Progress feedback0.3590.1990.1180.0900.124 Ease of control0.1800.1440.0590.0670.158 GLM DeepSeek Kimi MiniMax Qwen 20 24 85 85 55 20 74 25 76 Goals 24 37 33 47 58 42 23 56 27 42 59 Autonomy 45 26 24 25 47 47 70 59 42 38 20 Mastery 58 73 35 24 64 40 24 59 46 24 20 Curiosity GLM DeepSeek Kimi MiniMax Qwen 77 20 27 20 64 20 78 68 57 25 Challenge fit 19 44 50 47 68 50 41 31 35 26 Cognition 51 21 20 53 40 64 54 42 31 29 27 Decisions 29 25 83 84 62 26 71 3821 Meaning 050100 GLM DeepSeek Kimi MiniMax Qwen 50 68 49 23 71 47 30 38 4518 Immersion 050100 96 69 84 31 91 30 4720 Emotion 050100 41 19 39 53 74 50 32 56 34 61 Feedback 050100 80 87 42 29 86 52 67 Control Experience rating 12345 Figure 7: How evaluators use the experience rating scale. Rating distributions reveal differences in scoring levels across models, personas, and experience dimensions. Colors identify the five rating levels. A.3SCORING LEVEL AND SAME-INPUT JUDGMENTS We compare all ten judge pairs on each common input, yielding 3,630 comparisons. Dimension distance is the mean absolute difference across the 12 ratings. We estimate each pair’s usual signed overall-score difference from all other games, weighting games equally, subtract this offset from the held-out comparison, and then take the absolute adjusted gap. Bootstrap samples refit the offset while excluding every copy of the held-out game. Applying thresholds of dimension MAE at most 0.5 and overall gap at least three identifies 63 pairs from 43 inputs and 30 games. For the 1,834 pairs with dimension MAE at most 0.5, the share with gap at least two falls from 31.5% to 8.8% after adjustment. The reduction is 22.6 percentage points. For a gap of at least three, the corresponding shares are 3.4% and 1.4%. In Authors, 26 successful requests produce a 13-0 win. GLM assigns 7 for smooth progress, whereas Qwen assigns 2 and emphasizes limited challenge and uncertainty. In Palace, Kimi assigns 3 and emphasizes the loss and slow hand reduction; MiniMax assigns 7 and values varied, coherent interaction. In Cold Hands Poker, one forced ante ends at a focal payoff of−5: DeepSeek assigns 5 for an empty but not actively frustrating experience, whereas Qwen assigns 2 for a lack of meaningful 14 RawAdjusted 0 10 20 30 40 Gap ≥ 2 (%) 31.5% 8.8% (a) Scoring level 13579 Overall rating GLM DeepSeek Kimi MiniMax Qwen 7 3 8 5 2 (b) Authors 13579 Overall rating GLM DeepSeek Kimi MiniMax Qwen 5 5 3 7 4 (c) Palace Figure 8: Scoring levels and judgments of the same play. (a) Accounting for habitual scoring differences reduces large overall-score gaps among judges with similar dimension ratings. (b,c) Authors and Palace illustrate differing evaluations of identical gameplay. play. In Arlington, GLM assigns 8 and values agency and progress, whereas DeepSeek assigns 5 because procedural evidence offers limited support for strong subjective engagement. A.4PERSONA REFERENCE SIMILARITY Letv r contain evaluatorr’s 12 dimension-overall correlations. We center each vector across dimen- sions and compare it with each persona profile: c r = v r −v r 1, s m,p = c ⊤ m c p ∥c m ∥c p ∥ , q m,p = 1 + s m,p P p ′ (1 + s m,p ′ ) . (7) The raw similarity is Pearson correlation between association profiles and may be negative. We map similarity from[−1, 1]to[0, 1]using(1 + s)/2, then normalize across the five personas to obtainq. Negative similarities are retained by this transformation. The shares describe relative resemblance among the five reference profiles; equal similarities yield a 20% share for each persona. Constant profiles and a zero denominator remain undefined. Whole-game bootstrap samples recompute the associations and similarities; 2,000 samples provide the intervals shown in the figures. Figure 10 also compares gameplay behavior. It averages within-game rank proximity among the ten player configurations across all 91 games. Ties receive average ranks, including when all configurations share the same descriptor value. Table 7: Relative resemblance to persona evaluation profiles. Normalized similarity shares (%) summarize the relative emphasis of each persona within a model’s evaluation profile. JudgeEfficiency Narrative Social System Thrill GLM24.322.410.125.317.9 DeepSeek22.522.313.223.318.7 Kimi23.024.29.925.217.7 MiniMax21.722.08.322.125.9 Qwen24.521.87.725.820.2 15 0%20%40% Efficiency Narrative Social System Thrill 24.30 22.44 10.15 25.26 17.86 GLM 0%20%40% 22.46 22.29 13.22 23.31 18.72 DeepSeek 0%20%40% 22.97 24.18 9.93 25.23 17.68 Kimi 0%20%40% 21.68 22.01 8.32 22.12 25.86 MiniMax 0%20%40% 24.47 21.83 7.67 25.83 20.20 Qwen Normalized similarity weight (%) Figure 9: Relative persona resemblance and its uncertainty. Normalized similarity shares show each model’s relative resemblance to the five reference profiles. The 20% line indicates equal shares; intervals show uncertainty. Efficiency Narrative Social System Narrative Social System Thrill 0.40 0.290.29 0.490.400.15 0.490.330.330.28 (a) Rating order 01 Efficiency Narrative Social System Narrative Social System Thrill 0.60 -0.020.03 0.860.71-0.22 0.480.35-0.160.43 (b) Profile similarity −11 Efficiency Narrative Social System Narrative Social System Thrill 0.80 0.780.77 0.810.800.78 0.720.790.730.74 (c) Play summaries 01 Figure 10: Relations among the five personas. The panels compare agreement in overall ratings, resemblance of dimension-association profiles, and proximity of gameplay behavior. These comple- mentary views describe different aspects of persona similarity. A.5PERSONA SCORES RECONSTRUCTING NEUTRAL RATINGS This supplementary analysis predicts a neutral judge’s overall score from the five persona scores on the same input. It is distinct from predicting a target persona’s preference from experience dimensions in Section 4.3. The reconstruction uses 363 inputs from 91 games. Within each training fold, persona scores and the target are standardized with means and standard deviations that weight games equally. Nonnegative mixture weights sum to one: min β j X i a i t i,j − X p β j,p z i,p ! 2 , β j,p ≥ 0, X p β j,p = 1,(8) wherea i is the reciprocal of the number of inputs in its game. Each game is held out in turn, and predictions are returned to the judge’s original scale. Baselines use the training mean, equal persona weights, the single persona selected by training error, and a fitted Efficiency-Thrill mixture. MSE is averaged first within game and then over games. Error-contrast intervals resample the 91 fixed held-out game losses 2,000 times; coefficient intervals use 1,000 whole-game refits. 16 Table 8: Reconstructing neutral ratings from persona scores. Held-out MSE compares alterna- tive combinations of persona evaluations. Lower values indicate better prediction. JudgeMean Equal Single Five personas Two personas GLM2.292 1.116 1.6021.1121.306 DeepSeek 1.507 1.192 1.5921.2271.398 Kimi2.560 1.668 2.3931.6841.982 MiniMax 2.412 1.534 2.3051.5551.951 Qwen1.606 0.603 0.8450.5520.735 0.00.81.6 GLM DeepSeek Kimi MiniMax Qwen 1.18 0.28 0.88 0.86 1.05 (a) vs. mean 0.00.30.6 0.19 0.17 0.30 0.40 0.18 (b) vs. two personas Reduction in held-out MSE Figure 11: Predictive value of persona score combinations. Positive error reductions favor the five-persona fit over (a) the mean baseline and (b) the Efficiency–Thrill fit. Intervals show uncertainty. 050100 Efficiency Narrative Social System Thrill 31.8 13.6 5.9 24.4 24.2 GLM 050100 26.1 9.1 15.8 28.6 20.4 DeepSeek 050100 23.0 16.6 6.1 27.9 26.4 Kimi 050100 19.7 22.4 9.8 26.4 21.7 MiniMax 050100 39.3 18.0 4.3 24.4 14.1 Qwen Reconstruction weight (%) Figure 12: Persona contributions to neutral-score reconstruction. Fitted weights describe how persona ratings combine to predict each neutral judge’s scores. Intervals show uncertainty in these relative contributions. A.6OWN-SOURCE RATING CONTRASTS For each game and judge, we subtract its average rating of other actors from its rating of its own actor’s trajectory. We then subtract the panel’s mean contrast for that actor. This adjustment accounts for actor outputs that all judges prefer; actor identities remain hidden during evaluation. All adjusted intervals shown in Figure 13 include zero. The Qwen judge and actor have different service configurations, so “own source” denotes their paired family label, not an identical endpoint. Table 9: Ratings of a model’s own gameplay relative to other actors. Adjusted contrasts ac- count for actor preferences shared by the judge panel. Judge Raw mean Adjusted mean GLM+0.159+0.048 DeepSeek -0.206+0.026 Kimi+0.113+0.059 MiniMax +0.080+0.147 Qwen+0.080-0.056 −202 Adjusted rating difference GLM DeepSeek Kimi MiniMax Qwen +0.056 +0.031 +0.069 +0.172 -0.065 Mean Figure 13:Own-source rating differences across games. Adjusted contrasts vary across games, while average effects remain compatible with zero. Diamonds indicate means and lines show uncertainty intervals. 17 A.7GAME RANKINGS AND BROAD CONSENSUS Game scores average the five actor sources equally within each of 91 complete games. Figure 14 compares rank similarity at two levels on matched data. The left panel correlates the ordering of the 12 dimension associations; the right correlates game scores. Figure 5 displays 15 games selected from the high, middle, and low pooled-score groups. Pairwise game-ranking correlations range from 0.599 to 0.847. For GLM and DeepSeek, the game-ranking correlation is 0.647 and the dimension-profile rank correlation is 0.259 on the matched set. For Table 1, the other four judges define high, middle, and low groups, each spanning ten rank posi- tions. Fractional weights preserve ties at group boundaries. Group Gap and Pair Agreement average the high–middle and middle–low comparisons; tied scores receive half credit in pair agreement. GLM DeepSeek Kimi MiniMax DeepSeek Kimi MiniMax Qwen 0.26 0.830.49 0.310.480.36 0.800.550.800.64 (a) Profiles −11 GLM DeepSeek Kimi MiniMax DeepSeek Kimi MiniMax Qwen 0.65 0.850.60 0.730.630.61 0.750.630.700.66 (b) Game ranks −11 Figure 14: Similarity in evaluation emphasis and game rankings. Models can agree on which games score highly while differing in how experience dimensions relate to their overall judgments. A.8EVALUATION RESEMBLANCE AND PRODUCED PLAY Evaluation resemblance and output endorsement are matched on all 91 benchmark games, covering 363 unique inputs. For a model-persona pair, evaluation resemblance is the Spearman correlation of their overall ratings, centered over personas within the model. Output endorsement is the persona’s rating of that model’s play after removing game-specific model and persona mean effects. This resemblance measure uses overall ratings, rather than the dimension-profile shares in Figure 3. The joint game-bootstrap analysis recomputes both directions. The correlation across the 25 model- persona cells is 0.046. −0.30.00.3 −0.2 0.0 0.2 1 2 3 4 5 GLM −0.30.00.3 1 2 3 4 5 DeepSeek −0.30.00.3 1 2 3 4 5 Kimi −0.30.00.3 1 2 3 4 5 MiniMax −0.30.00.3 1 2 3 4 5 Qwen Play endorsement Centered rating correlation 1 Efficiency2 Narrative3 Social4 System5 Thrill Figure 15: Evaluation resemblance and endorsement of model-produced play. Similarity in rating patterns shows no clear overall association with relative endorsement of a model’s gameplay. Colors and numbers identify personas. A separate exploratory analysis relates log-derived play descriptors to game ratings on all 91 bench- mark games. We average the five actor sources equally per game, then correlate game-average scores with log-transformed decision count, forced-action share, action-type diversity, and the mean 18 log number of choices at non-forced decisions. For trajectories with no free decisions, free-action diversity and choice breadth are set to zero. Figure 16 reports individual and simultaneous uncertainty. −101 GLM DeepSeek Kimi MiniMax Qwen +0.28 +0.05 +0.34 +0.15 +0.05 Decision length −101 -0.10 +0.05 -0.17 +0.04 -0.08 Forced actions −101 +0.11 -0.02 +0.23 +0.06 +0.18 Action diversity −101 +0.24 +0.04 +0.09 +0.06 +0.19 Choice breadth Spearman correlation 95% intervalSimultaneous 95% interval Figure 16: Observed gameplay characteristics and overall ratings. Associations summarize how decision length and action-choice characteristics relate to scores. Thick and thin lines show individual and simultaneous uncertainty intervals. BEXPERIENCE REPRESENTATION AND EVALUATION DETAILS Evaluation datasets. Preference-model validation uses 91 games played under each of the five personas, giving 455 observations. The refinement records cover 30 games and five personas (150 conditions); the preference comparison in Table 2 uses the same 150 conditions, retaining the original- game score when a revision fails. Its implementation columns come from a separate set of runs on 30 shared-rule tasks. The preference runs evaluate games delivered to the target persona, while the implementation runs compare code generated for fixed target rulebooks. Reflexion and ReCode means use the reported aggregate scores for the evaluated conditions and the original-game scores for failed revisions. Appendix C describes the paired preference outcomes, and Appendix D details the implementation protocol. Preference predictors used to guide these historical revisions remain fitted on the original 49-game dataset; the expanded validation set does not change the recorded revisions. Table 10 groups the selected constructs by their source instrument. PXI contributes functional and psychosocial properties (Vanden Abeele et al., 2020), and CORGIS contributes three forms of challenge (Denisova et al., 2020). We adapt these concepts to trajectory-grounded scores on a 1-5 scale, allowing the same preference-model family to operate across game-specific rules and action spaces. Preference ratings.The post-play evaluator receives the persona profile, rule context, and concrete events from the played trajectory. The anchors are: 1, actively avoid another play; 3, unlikely to replay; 5, neutral or mixed; 7, willing to replay; 9, strongly prefers replaying or choosing this game; and 10, exceptional fit. A rating includes trajectory-specific evidence. The median of three temperature-zero evaluations of each trajectory is the reported preference score. The five personas are LLM player agents. Experience measurement during refinement. The measurement prompt receives the persona profile, rulebook, focal player, and played trajectory and returns a 1-5 score with trajectory evidence for each dimension. The selected attributes are measured in two fixed prompt batches. Three temperature-zero measurements are combined by the median for each dimension. Long play records are divided at turn boundaries, with observation-count-weighted scores across the selected chunks within each call. Refinement protocol.Each condition starts from its frozen original game and preference score. A candidate’s identity comprises game, persona, target dimension, and direction. Unchanged historical candidates reuse their corresponding evaluated outcome; newly selected candidates receive post- selection evaluation. The complete result includes every condition, assigning the original score when the procedure retains the original game. The same fallback applies to failed baseline revisions, which 19 Table 10: Experience constructs selected from two instruments. PXI supplies functional and psychosocial properties; CORGIS distinguishes sources of challenge. The selected constructs are adapted to the persona’s played trajectory. Selected constructTrajectory-based interpretation Player Experience Inventory (PXI) Goals and rulesClarity of objectives and the rules needed to pursue them Challenge fitFit between the challenge and the intended player’s experience Progress feedbackPerceptible information about progress and consequences Ease of controlEase of expressing the intended game action AutonomyMeaningful opportunities to choose how to act MasteryOpportunities to apply and develop competence CuriosityInterest in discovering game states, outcomes, or possibilities MeaningPerceived purpose and significance of play ImmersionAbsorption in the unfolding game Challenge Originating from Recent Gameplay Interaction Scale (CORGIS) Cognitive challengeDemands on reasoning and understanding Decision challengeDemands created by consequential alternatives Emotional challengeEmotional stakes and demands during play contribute zero preference change in the 30-game comparison. One gameplay trajectory per selected condition supplies the three preference ratings. B.1PREFERENCE PREDICTOR For persona p, the preference predictor takes the form f p (x) = b p + w ⊤ p x,(9) whereb p is the intercept andw p contains the dimension-specific coefficients. The model is fitted with regularization on paired experience and preference measurements from other games before being used to guide refinement of the current game. We express the coefficients in the original scoring units, so thatw pj represents the predicted preference change associated with a unit increase in dimensionj, holding the remaining dimensions fixed. Data and validation. The validation dataset contains 455 complete experience-preference ob- servations: one persona-played record for every combination of 91 games and five personas. A separate ridge model is fitted for each persona with regularization coefficient 10. Standardization and median imputation use training-fold statistics only; the current data have complete dimensions. Absolute-score prediction holds out the evaluated game across all personas. Pairwise validation excludes both compared games and evaluates within-persona pairs with distinct observed preference labels. Both predictions in a pair come from the same fitted model; a tied prediction receives half credit. Regression metrics average the five persona-specific results equally. Ordering accuracy pools 14,785 unequal-label pairs. Shared relationship baseline. The shared model is a pooled ridge regression with one common coefficient per dimension and five persona one-hot indicators. All feature columns, including the indicators, are standardized within each training fold; all coefficients use the same ridge penalty 10. Thus, it permits persona-dependent score levels while sharing the dimension slopes. The mean baseline predicts each persona’s training mean. The - persona ablation uses one pooled ridge model with the same 12 experience dimensions but no persona indicators, so both slopes and intercept are shared. Every model excludes the same held-out game or game pair. Figure 4(b) compares separate models with the shared and mean baselines; Table 3 reports the ablations. Uncertainty shown in the figures for absolute errors resamples whole games from the fixed held-out predictions (2,000 replicates). The pooled comparator and separate models use the same fixed penalty but differ in training sample size and feature scaling. Prediction results.Separate persona models achieve MAE 0.740 versus 1.090 for the persona-mean baseline, a paired error reduction of 0.350 points. The shared model achieves MAE 0.785; the paired 20 error reduction from shared to separate models is 0.046 points. Mean persona-wiseR 2 is 0.435 for the separate models, and their ordering accuracy is 82.0%, compared with 81.6% for the shared model. Both games in each ordering comparison are excluded from fitting. Without persona indicators, MAE is 0.880, mean persona-wise R 2 is 0.030, and ordering accuracy is 81.3%. Experience-source comparison.PXI uses goals/rules, challenge fit, progress feedback, ease of con- trol, autonomy, mastery, curiosity, meaning, and immersion; CORGIS uses cognitive, decision, and emotional challenge. All source comparisons retain identical observations, labels, fitting procedures, and validation folds. Adding CORGIS to PXI reduces MAE by 0.021 points. Target frequencies. Figure 17 counts the first-ranked feasible target over 30 games per persona in the historical refinement experiment. Efficiency first targets higher ease of control in 24 games, Narrative higher curiosity in 27, Social higher immersion in all 30, and System higher meaning in 27. Thrill first targets lower cognitive challenge in 22 games. The ranking uses the conditional coefficients of f p in the original scoring units after filtering for feasible changes. Control + Curiosity + Immersion + Meaning + Challenge fit + Feedback + Cognition − Emotion + Efficiency Narrative Social System Thrill 2451 273 30 273 2226 Figure 17: Preferred refinement directions across personas. First-ranked feasible targets reveal different priorities in the refinement experiment. Plus and minus signs denote intended increases and decreases in an experience dimension. Verification of actual refinement targets.The original refinement records fix three feasible targets per game-persona condition, ranked by the magnitude of the original-unit coefficient, with direction given by its sign. We independently refit the leave-current-game models on the original 49-game dataset (245 observations), divide each standardized coefficient by its training feature scale, and match the saved coefficients for all 150 conditions and all 450 ranked targets to numerical tolerance 10 −8 . Figure 17 counts the first-ranked feasible target, not the eventually accepted target. In Crazy Eights, for example, Efficiency first targets meaning (w = 0.583), Narrative curiosity (0.332), Social immersion (0.843), System meaning (0.422), and Thrill lower cognitive challenge (−0.721). Thrill’s challenge-fit target is third in that game’s frozen search order (w = 0.643). The target is not changed using an observed post-intervention preference score. CSUPPLEMENTARY REFINEMENT RESULTS Paired outcomes across 30 games. Each of the 30 games contributes all five personas. With the feedback extension, PREFER improves 39 of the 150 original-game conditions, ties 107, and worsens four. Directed refinement delivers 57 total preference points over the original games across 150 paired conditions, giving a mean of 0.380. Its 93 selected revisions comprise 32 first-ranked, 31 second- ranked, and 30 third-ranked candidates. The remaining 57 conditions retain their original games. The outcome counts are 35 increases, 111 ties, and four decreases. 21 Table 11: Sensitivity of refinement outcomes to the admission rule. Alternative rules trade off the number of adopted revisions and observed preference changes. Italics identify the rule used in the refinement experiment. Admission gate on the fixed poolAdopted / 150 From feedback Mean ∆y Imp. / Tie / Dec. Target≥ 0.25 and prediction > 010310+0 .40739 / 107 / 4 Target≥ 0.25 only10613+0.40040 / 104 / 6 Prediction > 0 only12128+0.48048 / 97 / 5 Target non-decreasing and prediction > 011522+0.43344 / 101 / 5 Validity only14350+0.42056 / 82 / 12 Systematic diagnostic feedback. The feedback experiment revisits 52 eligible retained-original conditions. In each condition, the source is the lowest numerical rank with valid gameplay and exact matching experience records. One feedback-driven rule revision preserves that source candidate’s target and direction. The batch produces 50 valid scored candidates. Ten satisfy the deployed admission rule. Updating the selected outputs gives 103 adopted revisions and 61 total preference points, for a mean of61/150 = 0.406. The dedicated Crazy Eights case has a separate record and is excluded from these aggregates. Sensitivity of the admission rule.Table 11 applies alternative admission rules to the fixed pool of delivered candidates: the 93 selected revisions and 50 valid feedback candidates. Every row uses the same executability requirement. This post-hoc analysis describes selection behavior on that pool. The deployed gate gives the main feedback mean of 0.407; the prediction-only rule selects 121 candidates with a mean of 0.480. These values expose a practical choice between requiring the intended target movement and accepting a broader set of predicted preference gains. C.1CASE RECORD AND VERIFICATION The Crazy Eights example is the separate feedback pilot for Thrill Seeker and the challenge-fit target. It is excluded from both the systematic feedback aggregate and the method comparison, each covering 150 conditions. We checked the revised rulebook, completed-play status, original experience responses, median preference record, saved gate decision, and frozen target coefficients. The core experience responses give challenge fit 3; the three separately collected preference evaluations have median 6. The saved baseline preference is 2. Summing the original-unit coefficient contributions reproduces the recorded predicted gain of 3.276. The gate record explicitly excludes observed preference and marks preference assessment as occurring after selection. The first proposal reveals a player’s hand after extended drawing. Its challenge-fit responses are all 2, and the saved critique emphasizes weak immediate consequences of drawing. The revised rule limits drawing to three cards, permits stopping after an unsuccessful draw, and buries the highest-value card on a bust. This reduces both hand size and remaining hand value. DIMPLEMENTATION AND BASELINE DETAILS Illustrative obligation. For a private-inspection rule, the observation function assumes that the intended recipient and card are recorded in the pending-inspection state. Its guarantee requires delivering the card to that recipient while preserving other players’ information boundaries. The provider set identifies functions that can establish this pending state. The inspection event activates the obligation, and the recipient’s next decision sets its deadline, expressing that delivery must occur in time to inform that decision. A function may own multiple clauses, each remaining linked to its supporting rule requirement. Planning. A contract plan maps each rule requirement to its owner function and lists possible prerequisite providers. Every clause has a stable identity, executable assumption and guarantee predicates, and an obligation trigger and deadline. Structural checks and requirement coverage determine plan acceptance; semantic-review advice informs plan revision. The preparation stage permits four attempts and records semantic-review advice. Accepted predicates and clause identities remain frozen through the implementation loop. 22 Local execution.A fixture begins from a reference reset-state copy and constructs entry arguments satisfying the owner’s assumption. The candidate owner executes on that input. The guarantee evaluates the actual before-state, after-state, and return value. Fixtures establish their prerequisites independently of the candidate provider functions, giving the local result a direct interpretation as an owner check. Joint execution.The runtime observes real function-entry events and associates them with activated obligations. At a matching owner call, it applies the same assumption predicate used locally. A reached deadline with a missing required call is recorded as a failed obligation. Candidate providers are drawn from declared relationships and execution evidence in the obligation window. Check outcomes distinguish pass, fail, and unknown, according to the observed predicate and execution coverage. Scope selection and repair.In Equation 4, the cost of a function is its inclusive source-line count, computed as its end line minus its start line plus one. The search in Equation 5 explores candidates in ascending cost order. Among scopes with equal total cost, it prefers fewer functions and then breaks ties by sorted function names. Each search is bounded to two seconds; reaching this limit leaves scope selection unresolved and returns no supported repair scope. The responsibility sets contain candidate functions supported by the observed execution, and failures without an evidence-supported responsibility set remain unresolved. Source edits are restricted to selected existing function bodies with uniquely matched replacements; signatures, rule text, and checking predicates are preserved. Up to two repair rounds test improvement while preserving prior passing cases. Baseline adaptations. Cardiverse follows its rulebook-to-code workflow using implementation templates (Li et al., 2025). Reflexion adapts self-feedback refinement (Shinn et al., 2023): feedback on the original persona play guides rule revision, and a reflection on the required changes precedes code editing and execution checks. ReCode adapts recursive code generation (Yu et al., 2025) with a plan limited to two levels and four executable leaf steps. Code is updated and tested after each step. PromptBreeder adapts prompt mutation (Fernando et al., 2024) using a fixed mutation prompt to generate one rule variant per game for all five personas. In the shared-rule implementation tasks, the same prompt supplies an implementation strategy for the prescribed target rulebook. Implementation evaluation protocol.The implementation evaluation uses the shared target rule- books described below. Preference refinement uses localized edits with execution checks; its playtest- ing and outcome aggregation are described in Appendix B and Appendix C. The benchmark contains 30 game-level rule-revision tasks, with a frozen target assigned to each task and personas rotated across games. Each program is evaluated with a random policy over seeds 52000-52009. The PCon evaluator uses the final six decisions and the game outcome to assess agreement with the revised rulebook. Outcome aggregation.Success rate counts final implementations that complete all ten evaluation games, with all 30 tasks as the denominator. For PCon, terminal non-delivery contributes zero; delivered outputs use valid paired seeds. We average within games and then weight games equally. This yields 29 comparable games and 286 paired seed positions. The implementation summary reports each system’s final output. Execution and embedding metrics. Exec replays each successful final implementation on 100 additional random seeds, 53000-53099, and averages its error-free fraction across games. Each rollout uses the same random-policy worker as the ten-seed evaluation, with a 4,000-turn limit and a 150- second timeout. ECon uses Qwen3-Embedding-0.6B to embed the complete final code and its shared structured target rulebook. Their cosine similarity is averaged over successful implementations. EFROM ASSESSMENT TO AN EXECUTABLE DESIGN OBJECTIVE A shared vocabulary makes different priorities inspectable. The model benchmark shows that similar game rankings can coexist with different dimension associations and scoring levels. Persona 23 references interpret these patterns; refinement uses conditional experience-preference models to choose targets. Held-out prediction connects these measurements to preference on unseen games. A fixed target connects feedback to a mechanic.Keeping the persona, dimension, and direction fixed lets feedback address the same design objective. The Crazy Eights record illustrates how an unchanged measurement guides a new drawing mechanic. The 30-game comparison tests delivered revisions across games. The results concern LLM personas playing card games; human-player studies can examine how the approach transfers to other players and settings. Implementation quality supports rule refinement. The intended experience depends on the behavior players encounter. Function contracts specify how cooperating functions should realize a revised rule. The implementation evaluation reports more runnable outputs and higher play-rule consistency for PREFER than for Cardiverse. 24