acceptodds
Under review as a conference paper at ICLR 2027

PREFER: From Player Experience to Personalized Game Refinement

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 8.7% higher mean post-play preference ratings than baseline.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Personalized game refinement, Preference-guided optimization, LLM agents, Assume–guarantee contracts, Game code generation.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yjBHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces an automated game designer that changes the game based on persona. The paper focused on a few dimensions like control driven or goal driven.

Critical strengths

The paper grounds the automatic game design process with respect to persona.

Critical weaknesses

The paper does not compare key LLM providers like openai and anthropic. I believe the idea to enforce persona into the designer is somewhat flawed. I like the idea to understand traces through these persona, but it is quite limiting. Fundamentally this paper needs greater scale, it mostly compares older models, and also simpler card games. No complex video games, nor board games. I would suggest to the authors to look into scaling the work to get better conclusive.

Questions for authors

Here are my raw notes:

ICLR 59742

I can read the abstract! Wow! Ok we want to understand how different LLM / VLM play games and what is it they do in terms of from established psychometric questionnaires.

Introduction. citations for PXI and CORGIS? I guess Vanden and Denisova. shared experience dimensions sounds cool, but also empty, what does it mean? l049 is repeated from the abstract, I know that. please tell me more about these. why limiting yourself to card games, could have been more general board games as well? or even simpler video games.

method. the math is confusing. The method is hard to parse.

Reviewer_wyauOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper covers an approach to game generation by having an LLM optimize an existing game according to different personas. They include an experiment that first investigates how different LLMs evaluate existing games then how these LLMs could be used to optimize these games. They find that the LLMs demonstrate a slight increase in ratings for the updated games.

Critical strengths

Including an experiment investigating how different LLMs evaluate existing games is useful. It may have been useful to also include a bit more detail or investigation into why these ratings differed or how consistent the ratings were, but the experiment is interesting.

Critical weaknesses

Unfortunately, the current paper draft has many substantial weaknesses.

The first weakness is the lack of clarity in the text. Sentences like "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." and "To examine how the same dimensions can guide refinement for players with different preferences, we use personas to simulate distinct player profiles." follow each other and say almost the same thing. This makes reading the paper very difficult.

The second major weakness is the lack of technical novelty. There are many papers investigating using an LLM to update a game, as illustrated by the authors' own citations. The technical novelty in the paper appears to arise from the use of personas but this has unclear value and is also not novel. Personas are common in the LLM-as-a-judge paradigm [1].

The third major weakness is the lack of replicability. There are no prompts included in the appendix, no code release for the harness, and generally speaking not enough detail to replicate the work.

The fourth major weakness is the results. While I appreciate the first experiment, the authors see only a marginal increase according to the LLM raters. Given that the same LLM is making the change and judging the result of the change this is expected given the known self-preference bias in LLMs [2].

  1. Guerdan, Luke, et al. "Doubly-robust LLM-as-a-judge: Externally valid estimation with imperfect personas." International Conference on Learning Representations. Vol. 2026. 2026.
  2. Wataoka, Koki, Tsubasa Takahashi, and Ryokan Ri. "Self-preference bias in llm-as-a-judge." arXiv preprint arXiv:2410.21819 (2024).
Questions for authors
  1. Can the authors share the details of their agentic harness/framework?
Reviewer_LS3cOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors introduce PREFER, a pipeline that rates card-game trajectories on twelve dimensions. These dimensions are taken from 2 previous works. The dimensions are used to cover how a game supports and engages the player. Separately, there is an overall preference score. It uses those ratings from the twelve dimensions in two ways. First, five models judge the same trajectories, and the paper compares dimension–rating associations, agreement on which games score high or low, and similarity to five personas. Second, a ridge model maps a persona’s dimension scores to preference, ranks dimensions by coefficient magnitude, and asks an LLM to revise one mechanic in that direction. Function contracts check the code, and a candidate is kept when the target dimension moves and the same linear model predicts a preference gain. On LLM persona playtests, PREFER’s mean preference is 3.333, against 3.067 for the strongest adapted baseline and 2.926 for the original games.

Critical strengths
  • The pipeline makes sense. The split between dimension scores and the overall rating feels like the right way to go about it. The two calls don't see each other, so the association analysis isn't an artefact of the model.
  • Introduction and methodology are very solid.
  • Figures are clear and easy to follow.
Critical weaknesses
  • I believe Related Works section can improve a bit.
  • A spelling mistake at line 384: Reflexion instead of Reflecxion.

No other weakness. The paper is a very strong paper.

Questions for authors

There are no questions from my side. My suggestion is to add citations like GameTraversalBenchmark[1] and lmgame-bench[2] instead of GameBench and TextArena, as they are only preprints. There are quite a few preprints that you are using. I suggest you cite papers that are peer-reviewed as I believe the paper is very strong and it needs to have proper citations.

I am happy to push my rating to 4 if the citations are fixed.

[1] Nasir, M.U., James, S. and Togelius, J., 2024. Gametraversalbenchmark: Evaluating planning abilities of large language models through traversing 2d game maps. Advances in Neural Information Processing Systems, 37, pp.31813-31827. [2] Hu, L., Huo, M., Zhang, Y., Yu, H., Xing, E.P., Stoica, I., Rosing, T., Jin, H. and Zhang, H., 2026, April. lmgame-bench: How good are llms at playing games?. In International Conference on Learning Representations (Vol. 2026, pp. 81308-81356).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.