acceptodds
Under review as a conference paper at ICLR 2027

RPG Dreamer: Towards Agent-Driven Generation of Long-Horizon Interactive Worlds

Abstract

Games, as a special category of software, create interactive worlds governed by their own specific rules. Given the recent successes of coding agents, increasing attention is devoted to automatic game generation. We observe that existing methods fail to create long-horizon games, as they entangle game mechanics with content and lack effective playtesting capabilities. In this work, we present RPG Dreamer, a framework capable of generating games with hours-long coherent playthroughs. We use role-playing games (RPGs) to exemplify the long-horizon generation capabilities of the framework, as RPGs feature branching narratives, non-linear progressions, and extended playtime. RPG Dreamer consists of an Agent-Oriented Game Engine (AOGE) and a Reviewer-centric Generation Loop (RGL). AOGE separates mechanics from content and exposes an agent interface with uniform observation and semantic action, enhancing both the reliability and coverage of automated gameplay. RGL is a closed-loop pipeline that enables agents to iteratively generate, review, and evolve games through a multi-reviewer system, including a team of gameplay reviewers who conduct playtesting to expose latent bugs. Due to the absence of benchmarks for evaluating hour-scale game generation, we further propose RPG-Bench, a benchmark of 150 evaluation instances spanning diverse themes, whose metrics are grounded in dynamic gameplay trajectories. Results on RPG-Bench demonstrate that RPG Dreamer improves the scalability of state-of-the-art coding agents on long-horizon games.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Language Models, Agents, Game

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fZ52OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RPG Dreamer, an agent-based framework for automatically generating long-horizon role-playing games (RPGs). The two main challenges in existing coding agents are the entanglement of game mechanics and content, and insufficient capabilities for effective game testing. To address these challenges, the paper introduces two components: the Agent-Oriented Game Engine (AOGE) and the Reviewer-centric Generation Loop (RGL). AOGE separates reusable game mechanics from game-specific content and provides a unified Agent Interface. RGL uses code, content, and gameplay reviewers to iteratively inspect and refine generated games. The authors also introduce RPG-Bench, comprising 150 game-generation tasks. The benchmark evaluates generated games through dynamic gameplay trajectories along four dimensions: requirement fidelity, content richness, visual and interaction quality, and playability. Experiments show that RPG Dreamer substantially outperforms existing coding agents and game-generation approaches across several metrics.

Critical strengths
  1. Current coding agents struggle with complex narratives, branching quests, and long-horizon state dependencies. This paper focuses on addressing this challenge by focusing on complete games that can be played over extended periods, rather than short demos.
  2. Dynamic evaluation based on actual gameplay trajectories can reveal inconsistencies in game logic that static evaluations may overlook.
Critical weaknesses
  1. Completion rate alone does not adequately support the motivation that game-generation quality deteriorates with complexity. In Section 2.2, the authors observe that route completion rates decrease substantially as game complexity increases from L1 to L5. However, completion rate is not a direct measure of generation quality. A breakdown of unsuccessful routes by failure type would provide stronger support for the stated motivation.
  2. The manually constructed game mechanics create a gap between the claimed task of complete game generation and what the framework actually generates. The paper emphasizes generating complete, complex, long-horizon games from natural-language requirements. However, Section 3 and Appendix F.1 indicate that the core mechanics in AOGE are manually implemented in advance and remain fixed during subsequent generation. The task therefore appears closer to automatic game-content generation on top of a pre-engineered framework than fully end-to-end game generation. Since game mechanics and their interactions are central to game design and playability, providing them in advance may substantially reduce task difficulty. This also raises concerns about the fairness of the baseline comparison where RPG Dreamer benefits from prebuilt, tested mechanics, whereas baseline coding agents must implement corresponding systems themselves. The authors should clarify the scope of generation supported by the framework and ideally compare methods under the same prebuilt-mechanics setting.
  3. The playability metric primarily measures functional correctness rather than the broader quality of gameplay. Playability (PL) receives a weight of 30% in the overall score. However, according to Section 4.2 and Appendix G.1, it is largely assessed through quest completion, game-state consistency, and basic functional operations. These measures capture functional correctness and progression reliability, but do not fully represent playability. This concern is particularly relevant given that the main game mechanics are manually predefined. The authors should more clearly delimit what PL measures and, if possible, provide complementary evidence on gameplay quality and real playtime.
Questions for authors

Q1. In the complexity-scaling experiment in Section 2.2, could the authors provide a breakdown of why routes were not completed? How many failures were caused by implementation bugs, as opposed to legitimate gameplay difficulty or insufficient guidance? Q2. Which mechanics in AOGE are manually implemented? If a user requests a new gameplay mechanic or a game system not supported by the current AOGE, can RPG Dreamer autonomously design and implement it? Q3. If all baselines were provided with the same prebuilt game mechanics as RPG Dreamer, would the proposed framework still achieve substantial improvements? Q4. What motivates the decision to measure playability primarily through quest completion, state consistency, and operational correctness? Is there additional evidence of gameplay depth, difficulty balance, or player experience?

4 comments

Reviewer fZ52

W1 & Q1: Causes of Uncompleted Routes

We thank the reviewer for raising this question regarding the failure causes in our complexity-scaling study. As described in Section 2.2 and Appendix D.1.1, for every generated game we extract its complete branching route structure from the source code in advance to serve as a walkthrough guide for the evaluators, thereby eliminating the influence of gameplay difficulty and insufficient in-game guidance. In fact, we inspected the failure causes of all uncompleted routes and found that every failure stems from implementation bugs, including unreasonable HP and combat stat configurations, erroneous dialogue branch conditions, and unreachable areas. Although some battles are challenging and require proper skill usage to win, evaluators can pass them by following the walkthrough guide, so combat difficulty is not an issue. We clarify this finding in the revised Appendix D.1.1.

W2 & Q2: Scope of Task and Constructed Mechanics

We thank the reviewer for this thoughtful question regarding the role of AOGE and the scope of our task.

Scope of the task. Our claim about generating complete, complex, long-horizon games refers to the fact that, compared with prior methods that produce only isolated code snippets, single levels, or text-only prototypes, our generated games are complete, self-contained games encompassing images, sprite animations, audio, and multi-chapter narratives. Meanwhile, rather than focusing on having agents invent intricate or novel game mechanics, our work focuses on reliably generating playable long-horizon games. Other systems for end-to-end game generation also use templates: OpenGame builds on project templates, while AutoUE builds on Unreal Engine gameplay modules and code templates.

Scope of the constructed mechanics. In our experimental setup, manually implemented means that we defined and fixed the required set of foundational RPG mechanics, prompted an agent to implement these requirements, manually verified that all required mechanics were properly implemented, and then kept these mechanics fixed when generating games across different requirements. Because games with different mechanics require different evaluation perspectives, we fixed the mechanics in our experiments to establish consistent evaluation rubrics.

Role of mechanics. While the constructed mechanics determine how a player interacts with the game, they do not guarantee playability. In fact, especially in games, many bugs arise from content rather than mechanics. For example, the mechanics define how objects, collision boxes, and interactive zones operate on a map, whereas the content specifies their exact positions and sizes; in practice, the bugs players frequently encounter are not game crashes upon entering a map, but offset positions or incorrect sizes that prevent players from obtaining required items to advance the progression. AOGE only specifies the overall architecture without restricting the game genre: users can either modify our provided one to add custom mechanics or prompt an agent to build a new version for other genres, such as visual novels or platformers.

We have revised Section 3 in the updated manuscript to clarify the construction of AOGE and the respective roles of mechanics and content.

W2 & Q3: Fair Comparison with Prebuilt Mechanics

We thank the reviewer for raising this concern regarding comparison fairness. To test whether RPG Dreamer's gains come merely from prebuilt game mechanics, we have added an ablation in Appendix B.5 (Table 8) evaluating multiple methods when provided with the exact same prebuilt mechanics used by RPG Dreamer:

HarnessModelPrebuilt MechanicsRFCRVIPLOverall
Kimi-codeKimi K3No92.045.575.940.557.8
Kimi-codeKimi K3Yes90.049.077.244.060.1
CodexGPT 5.6 SolNo88.057.956.263.862.2
CodexGPT 5.6 SolYes88.065.362.067.167.1
RPG DreamerGPT 5.6 SolYes92.075.472.686.579.6

As shown in the table, providing prebuilt mechanics alone brings only minor overall gains to the baselines (+4.0% for Kimi-code and +7.9% for Codex), whereas RPG Dreamer still outperforms Codex by 18.6% overall and 28.9% in playability under the same model and prebuilt mechanics. This confirms that prebuilt mechanics alone do not substantially improve baseline agents on long-horizon games without our agentic interface and reviewer-centric loop.

W3 & Q4: Scope and Motivation of the Playability Metric

We thank the reviewer for the comment on our playability metric. Rather than measuring subjective player enjoyment, our playability metric primarily evaluates whether a player can completely experience the created long-horizon game content, including progressing through quests across all branches, using all designed game functions, and maintaining reasonable behavior under edge-case interactions. We view this as the fundamental prerequisite for a game to be playable, whereas higher-level aspects of the gameplay experience, such as gameplay depth, difficulty balance, and overall engagement, are highly subjective and require more complex, dedicated investigation, which we discuss as future work. We have updated Section 4.2 in the revised manuscript to clarify the scope of the Playability metric.

If you have any further questions, we would be more than happy to continue the discussion.

I thank the authors for making the effort to address my questions. However, the paper will be much stronger if the authors can include more in-depth playability metric instead of leaving it as future work.

To the best of our knowledge, this is the first work to build a framework for long-horizon agentic game generation. We have put substantial effort into establishing this new direction:

  1. Decoupled Game Generation Framework: We design AOGE to decouple reusable game mechanics from game-specific content while providing unified observation and action interfaces for agents.
  2. Closed-Loop Generation Pipeline: We introduce RGL to iteratively evolve game content and target latent bugs via multi-reviewer playtesting.
  3. Long-Horizon RPG Dataset: We construct RPG-Bench, comprising 150 evaluation instances for generating hour-scale RPGs across diverse genres.
  4. Trajectory-Based Evaluation: We introduce an evaluation protocol grounded in dynamic gameplay trajectories across multiple modes, where playability specifically evaluates long-horizon stability.

We fully agree that a more in-depth playability metric would be valuable. We would like to respectfully clarify that our decision to leave this for future work is based on the following two considerations:

  1. Scope and Page Limit: Our work already encompasses a broad scope covering both the generation framework and the benchmark (as Reviewer WKF2 also noted). Adding extra metrics would make it difficult to explain everything clearly within the page limit.
  2. Time and Resource Constraints: Assessing higher-level aspects of the gameplay experience requires extensive subjective human evaluations. Conducting such evaluations reliably involves substantial time and annotation costs, making it infeasible to complete in the short term.
Reviewer_WKF2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce RPG Dreamer, a pipeline to generate long-horizon playable games. The pipeline consists of 2 phases which are meant to separate mechanics generation and content generation. The first stage constructs as Agent-Oriented Game Engine (AOGE), which contains reusable mechanics, content slots, and an interface through which an agent can observe game state. The second stage is called Reviewer-centric Generation Loop (RGL) that automatically populates game specific content. This stage is a self-evolving loop that generates assets, writes code and designs the game, and eventually reviews it. Lastly, authors introduce RPG-Bench, a benchmark containing 150 instances with diverse themes. The results suggest that the RPG Dreamer beats the baselines quite comfortably.

Critical strengths

The strengths are:

  • Very nice introduction.
  • Explanation of methodology is very clear
  • Overall method makes sense. Once stage for mechanics and the other for the content design is the right way to go about.
  • Benchmark is an addition to long-horizon evaluation tasks.
  • Ablations are thorough.
Critical weaknesses

The weaknesses are:

  • Introducing such a big pipeline with a benchmark overpopulates the paper. Due to this, the related work had to be moved to the appendix, which is not a good sign. Why? Because related work is one of the most important sections for researchers who are interested in the domain. This has to be fixed.
  • There is no game-specific baseline harness. All baselines are LLMs with the provider harness. There should be at least one specialised harness.
  • Figure 1 is very hard to follow. A clear, human-made example here would be great for the paper
  • Figure 4 feels completely unnecessary.
Questions for authors
  • Can you move related works section in the main paper? A suggestion is to remove Figure 4 and compress Figure 1 to make some space for the Related Works section. You can also remove equations 1 and 4, as they do not add any value.

  • What specialised harness can you implement as a baseline?

  • Can you create an example diagram that replaces Figure 1 yourself?

  • Can you remove Figure 4?

This is a very strong paper. I am happy to make it a "clear acceptance" if the above queries are resolved.

3 comments

W1 & Q1: Related Work in the Main Paper

We thank the reviewer for this constructive suggestion of related work. Following the reviewer's recommendation, we have moved a Related Work section to the main text.

W2 & Q2: Game-Specific Baseline Harness

We thank the reviewer for highlighting the importance of comparing against a specialized game-generation harness. Because using coding agents to generate complete games with assets is a very new research direction, most related studies are still at the preprint stage. Nevertheless, we have included a comparison against the pioneering work OpenGame in Table 2, which is a specialized agent harness tailored specifically for end-to-end game generation. To avoid any ambiguity, we have revised the manuscript to highlight this comparison.

W3, W4, Q3 & Q4: Presentation and Figures

We thank the reviewer for the helpful feedback on improving the presentation of the paper. In the revised manuscript, we have redesigned and replaced Figure 1 with a clearer, more compact human-crafted diagram, and we have removed the former Figure 4 from the main text.

Please feel free to reach out if you have any additional suggestions. We are always happy to engage further.

I am happy that the authors put in effort for all the changes. I am happy to change the score to 3. I still feel that having RPG-Bench makes the paper a bit too crowded. I understand the need of the benchmark. If it was submitted to a journal then it might have made more sense to have this much information in the main text. Paper is in a great shape and should be published somewhere.

We sincerely thank the reviewer for their encouragement and thoughtful suggestions.

We greatly value ICLR's emphasis on ambitious and comprehensive research, and we believe our work aligns well with this vision by establishing a foundation for long-horizon game generation through both our framework and benchmark. We look forward to sharing these contributions with the community and fostering further research in this direction. Meanwhile, we are already preparing an extended version for a journal that supports additional game genres and more comprehensive evaluations.

Reviewer_jBKsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper covers an approach to automated game generation of RPGs by splitting up the generation process into content and mechanics, and giving agents the ability to test the generated games for playability. The paper also introduces RPG-Bench, a novel benchmark to evaluate generated RPGs. The authors find that their approach does better on their benchmark than various baseline LLM agentic systems.

Critical strengths

The inclusion of multiple different LLMs for comparison is beneficial, and I applaud not using the same LLM for generation and evaluation, which I find to be a common mistake.

Critical weaknesses

The current paper has a number of substantial weaknesses holding it back.

One major weakness is a lack of engagement with prior work. The most crucial of these are claims not supported by citations, for example "Next, according to game design theory (Salen & Zimmerman, 2003), games can be decomposed into two parts: mechanics, which define reusable and underlying rules, and content, which instantiates the rules with specific assets, levels, and other artifacts". This is not what "Rules of Play" says at all. A closer citation might have been Zubek's Elements of Game Design, which separates out mechanics and game objects. There are major omissions when covering related work. The idea to separate out mechanics and content in game generation is over a decade old [1]. Even if the authors wish only to engage with LLM work this has been done over two years ago [2].

More substantially, the contributions are not of sufficient technical strength to warrant inclusion. A novel harness for agents is not a sufficient technical contribution for this venue. The benchmark is, in comparison, very promising. However, there is almost no detail on its construction and it is not separately verified to actually measure what the authors claim it measures. While the authors twice mention a user study and provide a table of results from it, there is no description of said study in the paper nor any attempt to ensure that the user study results correlate with RPG-Bench.

  1. Cook, Michael, Simon Colton, and Jeremy Gow. "The angelina videogame design system—part i." IEEE Transactions on Computational Intelligence and AI in Games 9.2 (2016): 192-203.
  2. Zammit, Marvin, Antonios Liapis, and Georgios N. Yannakakis. "CrawLLM: theming games with large language models." 2024 IEEE Conference on Games (CoG). IEEE, 2024.
Questions for authors
  1. What was the methodology of the user study?
4 comments

W1: Prior Work and Citations

We thank the reviewer for the helpful feedback on our discussion of game design theory and related work in automated game generation.

Game design theory references. Regarding our reference to Rules of Play, we cited this book because Chapter 11 examines what constitutes game rules and Chapter 12 further discusses how different levels of rules jointly define a game; through the respective examples of playing-card suit designs and the reward-and-punishment narrative on a Chutes and Ladders board, these two chapters illustrate that a game's visual and narrative content can change without altering its formal rules. We also appreciate the suggestion to reference Elements of Game Design, which we have read and cited in the revised manuscript.

Prior decoupling approaches. As shown in the two works mentioned by the reviewer, ANGELINA and CrawLLM, we acknowledge that separating mechanics from content has already been reflected in earlier research, and we do not claim this general concept as our core innovation. Instead, our innovation in AOGE lies in how we operationalize this separation for long-horizon coding-agent generation and closed-loop playtesting. While prior pipelines mainly apply this separation to one-shot asset re-theming or single-level assembly, AOGE exposes configurable Content Slots for multi-chapter branching progression and pairs reusable mechanics with a content-agnostic Agent Interface for intent-level automated playtesting.

In the revised manuscript, we have enriched the Related Work section to discuss these prior methods in detail.

W2: Technical Contribution Strength

We thank the reviewer for raising this concern regarding the technical contribution of our work. To the best of our knowledge, within the domain of agent-based game generation, our work is the first to decompose a game into three distinct parts: mechanics, content, and an agent interface. Furthermore, unlike prior approaches that focus predominantly on the forward generation process, our content generation stage places reviewers at the center and prioritizes the feedback channel for uncovering defects, as bugs in games are much harder to detect through static checks alone compared to conventional software. Both comparisons with SOTA methods and controlled ablation studies confirm that our proposed approach is highly effective, and we believe these findings and our framework are well worth sharing with the community.

W3 & Q1: Benchmark and User Study

We thank the reviewer for recognizing the value of RPG-Bench and for the constructive comments regarding our benchmark details and user study.

Details of benchmark construction. RPG-Bench consists of two components: an evaluation methodology and user requirement prompts. For the evaluation methodology, detailed descriptions and prompts for the gameplay modes, evaluation metrics, and evaluation procedure are provided in Appendix F.1.2–F.1.4. To further clarify the user requirement prompts, we have additionally included Appendix F.1.1 in the revised manuscript to summarize their construction and distribution, covering theme distribution, specification granularity, and contributor sources.

Verification of the evaluation methodology. To verify the reliability and effectiveness of our automatic evaluation, Appendix B.1 presents inter-judge agreement across different VLM evaluators as well as correlation analyses between automatic scores and human ratings.

Details of the user study. The user-study protocol and full questionnaire are documented in Appendix F.2. Rather than relying on overall impressions, we design a multi-mode, behaviorally anchored evaluation with specific questions about what evaluators directly observe during play, covering the same four dimensions as RPG-Bench. Specifically, fifteen evaluators participate in a blind study: after a standardized briefing, each evaluator plays the assigned games across all five gameplay modes before answering questions. Each evaluator plays the games generated by all compared methods under each shared requirement before submitting scores to calibrate personal rating standards.

To make these materials easier to locate, we have added clearer pointers in the main text to the corresponding appendix sections.

If you have any further questions about our framework, benchmark details, or user study, please feel free to let us know. We would be very glad to discuss them further.

Reviewer_jBKsOpenReview ↗

Re: Official Comment by Authors

Thanks to the authors for adding additional methodology detail as requested, but this is still missing a fair amount of information about the methodology. In particular, there's no mention of where the users were recruited from and there's no specifics about what information they were given about the study in terms of instructions or tutorialization saying only "we brief every evaluator on the four dimensions, the required gameplay modes, and the overall evaluation procedure to ensure a consistent standard across evaluators."

I retain my same concern about limited technical novelty. Being the first agentic approach to model the game generation process in terms of separating out these different components is not sufficient novelty for inclusion.

AuthorsOpenReview ↗

Details on Human Evaluation

Thank you for asking us to clarify the details of our human evaluation. Below is the complete protocol, which we will incorporate into Appendix F.2:

  1. Recruitment, Background, and Compensation. We recruited evaluators outside the research team through open calls across universities and companies. All participants provided informed consent, had no conflicts of interest with the study outcomes, and received fair compensation independent of their ratings. All 15 evaluators included in the final analysis had prior experience playing video games, and approximately half had experience in conventional game development (programming, art, or game design).

  2. Controlled Environment, Briefing, and Tutorialization. The evaluation was conducted in a controlled laboratory environment. Before formal scoring, we held a standardized group briefing covering the workflow, the five required gameplay modes (main-quest completion, quest coverage, sequence breaking, out-of-order interaction, and operation verification), and the four scoring dimensions (requirement fidelity, content richness, visual and interaction quality, and playability).

    • Demonstration and Reference Examples: Using a held-out game (excluded from formal evaluation), we walked evaluators through the interface and demonstrated how to test each mode—such as attempting to enter a late-game area before meeting its prerequisites (sequence breaking) and revisiting an earlier scene after world-state changes (out-of-order interaction). We also showed reference examples illustrating different rating tiers for representative questionnaire items so that evaluators shared a calibrated interpretation of the four-level, behaviorally anchored options.
    • Evaluation Rules: Evaluators were instructed to work independently, refrain from discussing the games or their ratings with other participants, and evaluate the games purely through gameplay without inspecting or modifying the source code.
  3. Blinded Gameplay and Scoring Protocol. All games were anonymized to conceal their generating methods and presented in randomized order. As detailed in Appendix F.2, each evaluator was assigned three requirement instances and played all compared methods for a given requirement before submitting any scores. Evaluators were required to spend at least 30 minutes playing each assigned game across all five modes, while the evaluation platform automatically logged in-game actions, active playtime, and screen recordings. Evaluators then completed the behaviorally anchored questionnaire based solely on direct gameplay observations.

  4. Objective, Blinded Quality Checks and Replacement. We initially enrolled 15 evaluators. To ensure data quality without introducing selection bias, members of the research team audited session logs for sustained, active gameplay strictly blind to the evaluators' ratings. Exclusion decisions were based solely on objective behavioral logs (e.g., prolonged inactivity or failure to execute the required gameplay modes within the minimum 30-minute window). When a session failed this check, we discarded that evaluator’s ratings and recruited a replacement who completed the exact same briefing, tutorialization, gameplay, and quality-check protocol. Across initial and replacement recruitment, we contacted 21 individuals to obtain 15 validated evaluators.

We will revise Appendix F.2 to include these recruitment, tutorialization, and quality-control details alongside the full questionnaire.

AuthorsOpenReview ↗

Clarification on Technical Novelty

Thank you for sharing your perspective on our work's technical contributions.

In our framework, modular separation in AOGE serves as an architectural prerequisite to enable our core technical contributions: closed-loop, execution-grounded verification for long-horizon interactive worlds and dynamic trajectory-based benchmarking:

  1. Mechanics-Grounded Semantic Agent Interface & Closed-Loop Playtesting (AOGE + RGL): A fundamental bottleneck preventing frontier coding agents from scaling to hour-long games is not code generation itself, but the inability to discover latent state-progression bugs deep within branching playthroughs (Section 2.2). Prior automated testing methods either drive games through low-level GUI and keyboard-mouse inputs, which struggle to progress into late-game stages, or co-generate test endpoints biased toward already-verified paths. By constructing a content-agnostic, intent-level Agent Interface prior to content generation and coupling it with RGL's multi-reviewer system (code, content, and gameplay reviewers), agents autonomously execute multi-branch routes and adversarial probes to iteratively diagnose and repair late-stage failures, raising latent bug detection rates from 10% (static code/content inspection) to 82% (Appendix B.4, Figure 9).

  2. First Benchmark & Dynamic Trajectory Evaluation for Hour-Scale Game Generation (RPG-Bench): To the best of our knowledge, this is the first work to formalize and benchmark long-horizon agentic game generation. We introduce a trajectory-grounded evaluation protocol spanning five complementary gameplay modes that systematically evaluate the coherence of the interactive RPG experience across extended playthroughs, where static checks often fail.

  3. Contribution to ICLR's Applications Track: We submitted this paper under ICLR's primary area of "Applications to computer vision, audio, language, and other modalities." Scaling foundation coding and multimodal agents from short-horizon demos to stateful, hour-long interactive games introduces distinct challenges. We hope this problem formulation, our proposed pipeline and benchmark can help facilitate future work.

We will revise our manuscript to sharpen our discussion of technical novelty and clarify this positioning in the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.