acceptodds
Under review as a conference paper at ICLR 2027

ReactiveGWM: Flexible Control and NPC Reactivity in Game World Models

14 views

Abstract

Existing game world models typically adopt role-specific interactions, where player and NPC roles are bound to fixed characters. This limits their flexibility in multi-character games, where different characters may receive external control while NPCs must react to interactions triggered by players. This setting raises two key challenges: *how to flexibly assign control roles to individual characters, and how to support direct player control and reactive NPC behavior within a unified model*. These challenges are particularly pronounced in shared-view 2D games, where multiple, potentially visually identical characters share the same viewpoint, making camera cues insufficient to distinguish their roles. To address these challenges, we introduce **ReactiveGWM**, a reactive game world model that flexibly assigns control modes at initialization and jointly simulates externally controlled players and reactive NPCs. Specifically, ReactiveGWM introduces *Spatial Role Binding*, which grounds learned character handles to their corresponding regions in the initial frame using instance masks. Building on these handles, *Unified Agency Conditioning* unifies heterogeneous control signals across characters by encoding player actions and conditional NPC rules into character-specific control-token groups. Each group is then bound to its corresponding character handle, enabling the model to apply each control signal to its designated character. Meanwhile, causal self-attention restricts temporal context to the current and preceding latent frames when generating player actions and NPC responses. Experiments on two multi-character 2D games demonstrate that ReactiveGWM supports flexible character control across different player/NPC role assignments while jointly generating accurate player-controlled behaviors and reactive NPC responses, enabling more configurable and richer multi-character interactions. Anonymous project page is available at https://reactivegwm.github.io/ReactiveGWM/.

Primary area:
generative models
Keywords:
Game World Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Qe8jOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents ReactiveGWM, a video world model for controlling multiple characters and generating NPC reactions in games. At the start of each rollout, users choose which characters to control directly and which should follow given NPC rules. Spatial Role Binding links characters in the initial frame to their control inputs, while Unified Agency Conditioning handles both player action sequences and conditional NPC rules. The model is fine-tuned from a pretrained video generator and evaluated in two shared-view 2D games. The experiments assess action execution, character selection, and NPC reactions, showing separate control of characters with similar appearances and NPC responses triggered by player actions. Ablations also show the contribution of role binding and NPC rule conditioning.

Critical strengths

The paper explores world models with multiple interacting NPCs, which I think is a promising direction. Judging from the examples, these are among the better demonstrations of multi-NPC interaction I have seen. They show NPC reactions triggered by player actions, as well as multiple characters performing different actions in the same scene.

There are also some interesting design choices. The authors use the same character handle to connect a character’s visual region with its control inputs, and handle player actions and NPC rules through the same control mechanism. This allows users to assign direct control or rule-based behavior even to characters that look identical. The design is straightforward, and the ablations show that it helps.

Critical weaknesses

I work on both coding agents and video generation, and my impression is that coding agents are progressing faster at the moment. This leaves me with some questions about the practical value of this approach. I am not yet sure which would be cheaper and more stable: using a coding agent through its API to build a similar game from scratch, or building a video generation framework, collecting data, and training a world model. The paper evaluates temporal consistency, but the experiments are limited to five-second clips. It also does not report training and inference costs, interaction latency, or behavior over longer runs. I find the work interesting from a research perspective, but for actual game development, the additional benefits over implementing similar interactions in code still need to be explained.

Questions for authors

Please refer to the Weaknesses section.

5 comments
AuthorsOpenReview ↗

Author Response

>>>Q1 Research Contributions vs. Practical Value of coding agents Approaches

>>>A1 Thank you for the insightful question! We agree that coding agents can implement predefined game mechanics more efficiently and reliably. However, video world models offer the potential to enable more flexible and fine-grained interactive experiences, expanding the possibilities of game creation beyond fixed, explicitly programmed mechanics. For example, while code-based engines can efficiently define game logic and coarse scene structures, video world models can generate fine-grained character motions and detailed visuals.

Moreover, recent research has begun exploring hybrid approaches that combine code-based engines with video world models, aiming to leverage the complementary strengths of both paradigms (e.g., Code World Model and Matrix-Game 4.0).

Our work focuses specifically on advancing video world models. We take a step toward this vision by exploring flexible multi-character control within a unified video world model, enabling characters to be either controlled by players or exhibit conditional NPC reactivity. We hope our research contributes to the development of video world models toward richer, more flexible, and interactive virtual worlds.

>>>Q2 Longer video and Training/inference cost

>>>A2 Thank you for the suggestion! For longer video, we further train a streaming variant using the Causal Forcing++ algorithm. Appendix F presents two continuous 30-second demonstrations at approximately 10 FPS: (1) two externally controlled players executing multiple actions; (2) NPC repeatedly approaching the player after each retreat. Both demonstrations illustrate sustained character control and repeated interactions well beyond the five-second clips. The videos are shown in the Anonymous Project Page.

For training and inference costs, training for 20K updates takes approximately 160 H200 GPU-hours, while the separate streaming adaptation requires approximately 300 H200 GPU-hours. For inference, the original model runs at approximately 2 FPS, whereas the distilled streaming model achieves approximately 10 FPS, corresponding to average per-frame processing times of approximately 500 ms and 100 ms, respectively, both including VAE decoding.

Thank you for providing the longer demonstrations, training costs, and inference speeds. These additions address some of my concerns about missing experimental details, but they do not resolve my central concern about the practical value of this approach. I still find coding-agent-based game development more promising in terms of stability, extensibility, and sustained gameplay. I would appreciate clarification on four concrete scenarios:

  1. During prolonged simultaneous two-player control, especially when characters overlap or exchange positions, do identity swaps or control errors accumulate over time?

  2. Can NPCs combine multiple rules into more complex gameplay strategies or follow condition–response combinations not represented in training?

  3. Can users control previously unseen characters using only a new initial image and character masks, or is additional training required?

  4. How does the model retain persistent changes caused by earlier interactions, such as damage to characters or changes to scene objects, once the relevant events fall outside its current video context?

My broader concern is that the paper primarily demonstrates training a video model on data collected from specific 2D games, while providing limited evidence that the model reliably captures interaction rules and maintains persistent game state. Once NPCs and multiple players are introduced, consistent interactions and reliable rule execution become central to the gameplay experience. The additional 30-second demonstrations are helpful, but extending a demonstration does not establish reliability during sustained play. The current results strike me as a promising research demonstration, rather than sufficient evidence for the more flexible game-development possibilities described in the response.

I also remain unconvinced that the authors have adequately justified using a heavyweight video-generation backbone for these settings. Code-based games can explicitly maintain state and rules while supporting targeted debugging and incremental feature development. The video-generation approach additionally has to address visual degradation, long-term memory, and consistency over extended rollouts. Although the response highlights the potential advantages of fine-grained motion and rich visuals, it does not establish what concrete experiences these advantages enable in the demonstrated 2D settings that would be difficult to achieve through code.

Cost further strengthens this concern. For the relatively limited 2D scenarios demonstrated here, using a coding agent to develop a playable game with comparable interactions may not require substantial development effort, and the resulting code can be reused and continuously modified. By contrast, the video-generation approach requires upfront data collection and training, followed by ongoing generation costs every time the game is played. Code-based games also incur execution and rendering costs, but the response does not establish that the additional generation costs deliver corresponding improvements in interaction capabilities or gameplay experience.

The response therefore has not persuaded me to change my overall assessment: the current results appear to reproduce interactions that code can already implement reliably, using a complex and computationally expensive generative framework, without demonstrating sufficient additional value to justify that choice. I remain pessimistic about the practical prospects of this approach to 2D game development. I maintain my rating of 3: Weak acceptance in recognition of the paper’s specific research contribution to multi-character control and conditional NPC reactivity, rather than agreement with the application prospects presented in the authors’ response.

Thank you for your timely reply and for recognizing the scientific value of our work. We first address your four new questions:

  1. We added an 30-second rollout with separate control streams for both characters, shown in Anonymous Project Page. It shows an initial position exchange and further interaction, but also ghosting, unintended jumps by the non-target character, and failed later crossings. These failure cases indicate that control errors indeed accumulate during prolonged interaction.
  2. We tested the distilled model on combined rules without additional training. The results suggest that reliable multi-rule generalization has not yet been established. This limitation is understandable, as training relies on single-rule rollouts with only ten predefined NPC rules per game, offering limited generalization to unseen condition–response combinations. We emphasize that our work primarily focuses on flexible character control and local NPC reactivity, while more complex NPC planning and compositional rule generalization remain outside the scope of this work.
  3. We tested the original ReactiveGWM with a new initial frame replacing both characters and the scene, together with corresponding masks, as shown in Anonymous Project Page. The background and the new character portraits in the upper-left and upper-right corners remain largely stable. However, the characters become blurred or distorted when performing actions.
  4. Our model does not include an explicit game-state or event-memory module. Our focus is flexible character control and local NPC reactivity and long-term state and event consistency remains a limitation outside the scope of this work.

Regarding application value, we agree with your point: code agents can implement these 2D games with explicit mechanics, NPC logic, and persistent state. Our paper focuses on how a learned video world model can bind external actions and conditional NPC behaviors to individually specified characters. We use 2D games primarily because their clearly defined actions, character identities, and interaction rules provide a controlled setting for evaluating this research question. We hope this character-level interaction interface can support future interactive video generation and visual prototyping. We do not claim a demonstrated development-cost or reliability advantage over code-based implementations.

Thank you again for your thoughtful feedback. If you have any further questions or concerns, please let us know.

Reviewer_YydgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReactiveGWM addresses flexible player/NPC role assignment in multi-character game world models. Spatial Role Binding associates characters with learned handles through initial-frame instance masks, while Unified Agency Conditioning binds player actions and conditional NPC rules to these handles. The model uses causal temporal attention and is evaluated on two shared-view 2D games, demonstrating promising character selection and action control alongside preliminary evidence of reactive NPC behavior.

Critical strengths

Relevant problem and coherent design. The framework supports assigning externally controlled and reactive roles independently of character appearance, addressing an important challenge when visually identical characters share the same scene.

Critical weaknesses

Missing comparative baselines limit the evidence for methodological novelty. The main results compare against ground-truth videos, and the ablations remove entire components. These experiments do not establish advantages over existing approaches or simpler conditioning mechanisms. In particular, a response-only baseline that removes the condition clause would help distinguish conditional reasoning from action generation driven by response keywords.

Questions for authors

Could the authors provide a controlled counterfactual evaluation that holds the initial scene, NPC rule, and generation seed fixed while delaying or removing the triggering player event, and measures whether the NPC correspondingly delays or suppresses its response using events observed in the generated video?

4 comments
AuthorsOpenReview ↗

Author Response

>>>Q1 Comparison with Response-Only variant

>>>A1 Thank you for the suggestion! We have conducted an additional comparison between ReactiveGWM and its Response-Only variant on the HNM game, as shown in the table below.

MethodJR (Gemini) ↑JR (GPT) ↑
ReactiveGWM76.67±0.0076.67±0.00
Response-Only56.67±3.3348.89±1.92

Specifically, for the Response-Only variant, we remove the if condition and retain only the response component during training. The results show that removing the conditional component leads to a substantial decline in Joint Reactivity (JR), demonstrating the effectiveness of our condition-response prompting strategy in training NPCs to react appropriately to player actions. Details and further discussion are provided in Section 4.4 (Lines 517–523).

>>>Q2 Counterfactual Analysis

>>>A2 Thank you for this valuable suggestion. We have added a controlled counterfactual evaluation in Appendix E, comprising six triplets (18 generated videos), to examine whether NPC responses adapt to changes in the timing or presence of Player-triggered events.

Experimental Setup. For each triplet, we consider three conditions: (1) Original, where the Player action is triggered at 1 s; (2) Delayed, where the same action is delayed to 2 s; and (3) Removed, where the triggering action is replaced with no-op inputs. Across all three conditions, we hold the initial scene, complete NPC rule, generation seed, and sampling settings fixed, varying only the Player action. Importantly, we manually annotate the actual Player-event onset and NPC-response onset directly from the generated videos, rather than relying on the specified trigger times.

Quantitative Results.

ConditionMean observed Player-event timeMean observed NPC-response time
Original1.075 s1.367 s
Delayed2.067 s2.517 s
RemovedNot observedNot observed

The results reveal three key findings:

  1. Delayed triggers lead to delayed responses. Delaying the Player action shifts the observed Player-event onset by 0.992 s and the NPC-response onset by 1.150 s on average, demonstrating that NPC responses follow changes in event timing.
  2. Removed triggers suppress responses. In all six removal videos, neither the triggering Player event nor a corresponding NPC response occurs throughout the 5 s clip.
  3. NPC responses remain rule-consistent. In all 12 triggered videos, NPC responses occur after the observed Player event, satisfy the complete NPC rule, and begin within 1.5 s of the event.

Overall, these results demonstrate that NPC responses adapt to the observed timing and presence of Player events, rather than following a fixed response schedule. Details and discussion are provided in Appendix E. The videos are shown in the Anonymous Project Page.

Thank you to the authors for their thoughtful response. Could you also provide some qualitative video results on the counterfactual evaluation?

Thank you for your timely reply!

We have added the corresponding counterfactual evaluation videos to our anonymous project page, whose link is provided in the paper’s abstract. In this project page, the Controlled Player interventions section contains six groups of videos covering the original, delayed, and removed trigger conditions, together with annotated timelines showing the triggering player events. These videos allow direct comparison of how NPC responses change when the triggering event is delayed or removed.

Thank you again for your feedback! If you have any further questions, please let us know.

Dear Reviewer Yydg:

Thank you again for your feedback!

As the discussion period is coming to an end, we would like to kindly ask if you have any remaining questions or concerns regarding our work or responses. We would be happy to provide further clarification if needed.

We sincerely appreciate your time and consideration!

Best regards,

The Authors

Reviewer_fKviOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing game world models typically adopt role-specific interactions, where player and NPC roles are bound to fixed characters. This limits their flexibility in multi-character games, where different characters may receive external control while NPCs must react to interactions triggered by players. This setting raises two key challenges: how to flexibly assign control roles to individual characters, and how to support direct player control and reactive NPC behavior within a unified model. These challenges are particularly pronounced in shared-view 2D games, where multiple, potentially visually identical characters share the same viewpoint, making camera cues insufficient to distinguish their roles.

To address these challenges, we introduce ReactiveGWM, a reactive game world model that flexibly assigns control modes at initialization and jointly simulates externally controlled players and reactive NPCs. Specifically,Spatial Role Binding grounds learned character handles to their corresponding regions in the initial frame using instance masks. Building on these handles, Unified Agency Conditioning unifies heterogeneous control signals across characters by encoding player actions and conditional NPC rules into character-specific control-token groups. Each group is then bound to its corresponding character handle, enabling the model to apply each control signal to its designated character. Meanwhile, causal self-attention restricts temporal context to the current and preceding latent frames when generating player actions and NPC responses.

Experiments on two multi-character 2D games demonstrate that ReactiveGWM supports flexible character control across different player/NPC role assignments while jointly generating accurate player-controlled behaviors and reactive NPC responses, enabling more configurable and richer multi-character interactions.

Critical strengths
  1. In extremely ambiguous scenarios where camera cues are absent and visually identical characters frequently appear (shared-view 2D games), the paper achieves dynamic role assignment for arbitrary characters (externally controlled players vs. autonomous NPCs). By injecting learnable handles only into the initial frame, it completely eliminates the need for frame-by-frame masks or trajectory guidance during subsequent generation. Meanwhile, in Unified Agency Conditioning, handles are injected exclusively into the Keys of cross-attention (responsible for addressing), while Values preserve pure control semantics (responsible for behavior). This design thoroughly decouples "control addressing" from "action semantics," solving the challenge of precise multi-character control with minimal inference overhead and establishing an inspiring new paradigm for multi-agent interactive world models.

  2. Under causal temporal attention constraints, the paper enables the diffusion model to directly internalize natural language "If-Then" conditional rules (e.g., "parry when a projectile approaches," "pursue and attack when a third party is hit"), exhibiting event-monitoring and reactive capabilities analogous to traditional game state machines. This pushes the research boundary of world models from "passive video renderers" toward "interactive logic-decision engines."

Critical weaknesses
  1. Across the entire experimental evaluation (Table 1 and main experiments), the paper only compares against Ground Truth and its own ablation variants (w/o Binding / w/o Conditioning), lacking horizontal comparisons with reasonable Baselines, which substantially undermines the persuasiveness of its empirical conclusions.

  2. The experiments are restricted to short 5-second (100-frame) clips within isolated 2D shared-view arena settings, where NPC behaviors are limited to simple, single-step static atomic rules (Single-step Atomic If-Then). Lacking long-horizon interactive capabilities, the paper fails to demonstrate the model's practicality as a streaming, playable game engine.

Questions for authors
  1. External comparison models need to be introduced; it remains unclear whether the currently achieved metrics represent a qualitative breakthrough brought by the proposed Spatial Role Binding & UAC architecture, or purely benefit from the fitting capacity of the underlying strong 5B pretrained video foundation model (Wan2.2).

  2. In fighting games, characters frequently experience cross-ups, close-range grappling, and overlapping occlusions during throw animations. After multiple frames of complete occlusion or spatial overlap, what mechanism ensures that the internal handles do not suffer from "ID switching" or "control drift"? Is there a quantitative error analysis specifically targeting these extreme crossing/occlusion scenarios (hard cases)?

  3. JR (Joint Reactivity) and AA (Action Accuracy) rely entirely on GPT-6-astra and Gemini 3.5 Flash for evaluation. Have the authors calibrated the consistency between the LLM judges' verdicts and human expert annotations (Human-LLM Agreement Rate)? Do the LLMs suffer from systematic hallucinations regarding subtle execution errors?

3 comments
AuthorsOpenReview ↗

Author Response [1/2]

>>>Q1 More comparisons with baselines

>>>A1 Thanks for the valuable suggestion! We have added ActionParty, which supports multi-character control in 2D games, as an external baseline. The results are presented in the following table:

MethodAA ↑ GeminiAA ↑ GPTSA ↑ GeminiSA ↑ GPTJR ↑ GeminiJR ↑ GPTVQ ↑ GeminiVQ ↑ GPT
HNM
ActionParty0.00±0.000.00±0.000.00±0.000.00±0.0035.56±5.0933.33±0.0076.47±0.3575.60±0.14
ReactiveGWM88.89±0.0096.30±0.0095.00±4.41100.00±0.0076.67±0.0076.67±0.0079.31±0.3778.80±0.53
SF3
ActionParty2.02±3.500.00±0.000.00±0.000.00±0.0042.22±6.9416.67±0.0040.93±0.8957.18±0.30
ReactiveGWM98.99±1.7596.97±0.00100.00±0.00100.00±0.0091.11±1.9291.11±3.8571.85±1.2782.49±0.52

ReactiveGWM outperforms this baseline across all four metrics under both judges. This gap is consistent with the different conditioning pathways. ActionParty conveys actions through coordinate-anchored subject states before rendering them into video, which can make pose-sensitive fighting actions harder to control when displacement is small.

ReactiveGWM directly conditions video features on temporally aligned, handle-bound action and rule tokens, making who should act, what to do, and when explicit. These differences offer a plausible explanation for the observed gap. The implementation details are provided in Section 4.1 (line 316-318), and the corresponding qualitative comparisons are presented in Section 4.2. The comparison videos are shown in the Anonymous Project Page.

>>>Q2 Evaluation Scope and Long-horizon interactive ability

>>>A2 Thanks for the question! Firstly, for the evaluation scope, you are right that our setting is restricted to simple, single-step static atomic rules with 2D shared-view arena settings. This is because the primary goal of ReactiveGWM is to enable flexible and unified control of multiple characters, rather than complex NPC planning or game environments. We intentionally adopt simple If-Then rules to isolate and evaluate NPC reactivity to player actions. Despite their simplicity, these rules require the model to correctly identify triggering events and generate appropriate responses for the designated NPCs, as supported by our experimental results. Moreover, we choose 2D shared-view games as a controlled yet challenging testbed for multi-character control, where multiple characters share the same camera view, may have identical appearances, and require explicit instance-level control binding. Therefore, these settings directly target the challenges addressed by our Spatial Role Binding and Unified Agency Conditioning. We acknowledge that more complex NPC behaviors and diverse game environments remain important directions for future work.

Secondly, for long-horizon interactive capabilities, we further train a streaming variant using the Causal Forcing++ algorithm. Appendix F presents two continuous 30-second demonstrations at approximately 10 FPS: (1)two externally controlled players executing multiple actions; (2) NPC repeatedly approaching the player after each retreat. Both demonstrations illustrate sustained character control and repeated interactions well beyond the five-second clips. The videos are shown in the Anonymous Project Page.

>>>Q3 Ablation on Contributions

>>>A3 We thank the reviewer for this suggestion. We have added original Wan2.2-5B to Table 2 to assess the pretrained backbone’s native control and reactivity capabilities. It achieves only 1.11/0.00 SA and 20.00/8.89 JR, compared with 95.00/100.00 SA and 76.67/76.67 JR for our method (Gemini/GPT). These results indicate that the pretrained backbone alone does not reliably provide the required capabilities. Furthermore, the existing ablations in Table 2 show that removing Spatial Role Binding substantially reduces SA, while removing NPC conditioning substantially lowers JR. Together, these findings support the effectiveness of our method design for character-specific control and conditional NPC reactivity.

AuthorsOpenReview ↗

Author Response [2/2]

>>>Q4 Quantitative error analysis on crossing/occlusion scenarios

>>>A4 Thank you for the insightful question. Regarding the mechanism for preventing control drift, our initial-mask handle binding, handle-conditioned action keys, and causal visual history jointly support identity-consistent control during interactions. For example, in an SF3 throw sequence, Player and NPC exchange positions but subsequently respond correctly to jump commands for Player. However, this capability remains imperfect, with some examples exhibiting control drift. For quantitative error analysis, in a fixed-scene diagnostic across ten samples, we identified four candidate cases exhibiting correct responses without control drift, while the remaining six showed continuity or control failures.

We acknowledge that these hard cases (close-range grappling and prolonged occlusion) remain particularly challenging. While our approach supports flexible and unified character control, maintaining consistent character identity under **severe occlusion introduces additional research challenges and remains an important direction for future work. Details and further discussion are provided in Appendix H. The videos are shown in the Anonymous Project Page.

>>>Q5 Human-LLM Agreement and LLM Hallucinations

>>>A5 Thanks for the suggestion! Firstly, we audited all 120 main-model videos (60 AA and 60 JR across SF3 and HNM). Two experienced human annotators independently evaluated the videos, and a third annotator resolved the three disagreements. Raw human–human agreement was 97.50% (117/120). For each video, we compared each LLM’s majority verdict across its three original evaluation rounds with the final human reference.

MetricNGPT–Human agreementGemini–Human agreement
AA6096.67%95.00%
JR6085.00%81.67%

These results use the original binary scoring (pass=1; fail/unknown=0). They provide direct human calibration evidence, showing high agreement for AA and more remaining disagreements for JR, supporting the value of human review alongside automated evaluation.

Secondly, we agree that LLM judges may hallucinate subtle execution details. Although averaging three evaluation rounds reduces run-to-run variability and LLM judges show high agreement with human evaluations, it cannot eliminate such errors. For example, both judges consistently accepted an SF3 blocking case across all three rounds, whereas human evaluation found no blocking action. We therefore agree that automated evaluation should be complemented by human review, particularly for subtle actions. Our human audit provides this complementary assessment and helps quantify the limitations of LLM-based evaluation. Further details are provided in Appendix G.

Dear Reviewer fKvi:

Thank you again for your feedback!

As the discussion period is coming to an end, we would like to kindly ask if you have any remaining questions or concerns regarding our work or responses. We would be happy to provide further clarification if needed.

We sincerely appreciate your time and consideration!

Best regards,

The Authors

Comments on the paper

2 comments

We sincerely thank all reviewers for their constructive feedback and encouraging comments. We are particularly grateful for their recognition of the following strengths of our work:

  • Promising research direction and important challenge: fKvi, Yydg, Qe8j
  • Coherent and interesting method design: Yydg, Qe8j
  • Promising experimental results: Yydg, Qe8j particularly noted, "Judging from the examples, these are among the better demonstrations of multi-NPC interaction I have seen."
  • Effective ablation studies: Qe8j
  • An inspiring new paradigm: fKvi

We address the reviewers’ questions and concerns below. Meanwhile, we have incorporated the necessary changes into the revised manuscript. All modifications are highlighted in blue for ease of reference.

AuthorsOpenReview ↗

Summary of Discussion

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

Thank you for your time and effort in managing the review process for our paper. To facilitate your review, we would like to provide a brief summary of the discussion period.


  • In the pre-discussion state, the reviewers provided positive feedback on several key aspects:
    • Promising research direction and important challenge: fKvi, Yydg, Qe8j
    • Coherent and interesting method design: Yydg, Qe8j
    • Promising experimental results: Yydg, Qe8j particularly noted, "Judging from the examples, these are among the better demonstrations of multi-NPC interaction I have seen."
    • Effective ablation studies: Qe8j
    • An inspiring new paradigm: fKvi

  • During the rebuttal phase, we actively addressed all the concerns with additional experiments, analyses, and clarifications. The major concerns are listed below:
    • External baseline (fKvi). We compared ReactiveGWM with ActionParty, an existing method for multi-character control in 2D games. ReactiveGWM scored higher on all metrics in both games (Sections 4.1–4.2).
    • Response-only ablation (Yydg). Removing the condition clause while retaining the response text reduced Joint Reactivity in both games. This result shows the effectiveness of our conditional prompting approach for NPC reactivity (Section 4.4).
    • Trigger dependence (Yydg). We held the initial scene, NPC rule, and generation seed fixed while varying the player input. The results show that NPC responses adapt to the observed timing and presence of player events. Following the reviewer’s request, we also added annotated videos to the anonymous project page.
    • Longer videos (fKvi, Qe8j). We added two 30-second streaming demonstrations at approximately 10 FPS, showing independent two-player control and repeated NPC reactions.

In addition to these, we have addressed all specific concerns raised by each reviewer and posted detailed responses in the comment sections. Meanwhile, we have incorporated the necessary changes into the revised manuscript. All modifications are highlighted in blue for ease of reference.


Finally, we are grateful to all reviewers for their active engagement during the discussion. We thank all reviewers, ACs, SACs, and PCs for their time and dedication to the review process.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.