acceptodds
Under review as a conference paper at ICLR 2027

ReWAM: Reciprocal World Action Models for Interactive Autonomous Driving

Abstract

In interactive scenarios, an autonomous driving system is required to generate ego actions under the influence of other agents' behaviors. Existing World Action Models (WAMs) typically model other agents as components of the world model rather than as decision-makers that fundamentally shape the action of the ego agent, which impairs their performance in dense interaction scenarios. We introduce Reciprocal World Action Models (ReWAM), a game-theoretic world action modeling framework that captures the reciprocal influence between the ego agent and other agents by representing them as conditional responders whose actions are mutually influenced. We instantiate this framework with a Level-k response hierarchy, where role-specific ego and other action DiTs exchange compact strategy tokens through cross-agent attention while remaining grounded in a shared representation of the future driving world. To learn the response policy of the ego agent from demonstrations, we formulate expert actions as samples from the best response distribution and jointly optimize the entire hierarchy using conditional flow matching. Our framework is evaluated on the NAVSIM dataset and achieves state-of-the-art performance compared to baselines. The improvement is particularly significant in interactive scenarios, validating that modeling reciprocal responses provides a more effective foundation for interaction-aware world action generation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, Autonomous Driving, Interaction, Game Theory

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_XUYbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to address a structural limitation of World Action Models (WAMs) which is they treat other agents as part of the world model rather than as decision-makers that shape the ego’s action. ReWAM addresses this by modeling ego and other agents as mutually responsive conditional responders through a Level-k game hierarchy. On the NAVSIM benchmark, ReWAM achieves PDMS of 90.6, with gains concentrated in the most interaction-intensive scenarios.

Critical strengths
  • The paper articulates the gap between WAMs and interaction-aware prediction/planning, and proposes a principled bridge.
  • The interaction stratification indicates ReWAM’s advantage over baselines grows monotonically with interaction intensity
Critical weaknesses
  • NAVSIM evaluates open-loop planning against cached logged context, and other agents do not actually respond to the ego’s predicted actions, and the method uses ground-truth scene annotations for current agent states. Thus the reciprocal-response gains are measured where other agents never adapt to the ego plan. So the core claim of mutually responsive interaction is not directly demonstrated in a reactive, closed-loop setting.
  • The method learns from demonstrations by assuming the logged joint policy components are reward-maximizing best responses to one another, but real driving logs are not necessarily game-theoretically optimal or level-k consistent. The optimal reasoning depth, the weight schedule α_k ∝ k^p, and K_train are also chosen somewhat ad hoc and shown to yield diminishing (even negative) returns beyond Level-4, suggesting dataset-dependence without principled selection.
  • The gains are marginal on aggregate and partly lower than the closest SOTA on some metrics. Nevertheless, the benchmark has already been updated recently.
  • All experiments are on NAVSIM only with no cross-dataset validation Long-horizon, multi-camera, or occluded interactions where reciprocal reasoning matters most, are not tested.
  • The method builds on DriveLAW and contributes the Level-k interaction block and a parameter-shared Other Action DiT. The reciprocal benefit may partly inherit the backbone’s world-modeling strength, and no ablation isolates the contribution of the strategy encoder / parameter sharing from the cross-agent attention itself.
Questions for authors
  • Does the reciprocal-response benefit survive true closed-loop interaction? An evaluation to directly substantiate mutually responsive interaction should be provided.
  • Is the aggregate SOTA lead statistically significant and robust?
  • How are the Level-k training depth and weight schedule α_k ∝ k^p selected, and how sensitive are results to them?
  • How much of the gain comes from the reciprocal-reasoning design vs. the added capacity?
4 comments
AuthorsOpenReview ↗

Performance Under Reactive Traffic

Q1/W1: Do the benefits of reciprocal conditioning persist when other agents can respond in closed loop?

Response: Thank you for asking whether the results extend to traffic that responds to the ego vehicle.

To test ReWAM beyond non-reactive log replay, we evaluated it on NAVSIM v2 with a reactive traffic policy. In this setting, simulated surrounding vehicles can change their motion in response to the ego vehicle. As shown in Table 1, ReWAM achieves the highest EPDMS among the compared methods in this reactive setting. Compared to DriveLaW, ReWAM demonstrates significant improvements in most metrics, including EPDMS, EP, TTC, and DAC. These results show that ReWAM remains competitive when surrounding traffic can react to the ego vehicle. This evaluation tests the full planner under reactive traffic; it does not isolate the gain from reciprocal conditioning under continuous closed-loop replanning.

We appreciate your emphasis on closed-loop evaluation.

Table 1. The Performance on NAVSIM v2 (higher is better). For each metric, the highest score is in bold and the second-highest distinct score is underlined.

MethodNCDACDDCTLCEPTTCLKHCECEPDMS
TransFuser96.989.997.899.7‾\underline{99.7}87.195.492.798.387.276.7
DiffusionDrive98.295.999.4‾\underline{99.4}99.887.5‾\underline{87.5}97.396.8‾\underline{96.8}98.387.7‾\underline{87.7}84.5
DriveSuprim97.596.599.4‾\underline{99.4}99.688.496.695.598.377.083.1
ARTEMIS98.3‾\underline{98.3}95.198.699.881.597.4‾\underline{97.4}96.598.389.183.1
DriveVLA-W098.599.198.099.7‾\underline{99.7}86.498.193.297.9‾\underline{97.9}58.986.1‾\underline{86.1}
DriveLaW97.396.499.598.084.296.996.296.384.785.4
ReWAM (ours)98.3‾\underline{98.3}97.6‾\underline{97.6}99.299.686.298.197.397.583.186.3
AuthorsOpenReview ↗

Evidence for ReWAM’s SOTA Lead

Q2: Is the aggregate lead over baselines statistically significant and robust?

Response: Thank you for comments on how well the benchmark results support our claim.

On NAVTEST dataset, which comprises more than 12,000 scenarios, ReWAM ranks first among the compared methods with a PDMS of 90.6, surpassing the previous best reported score of 90.4 from ReWorld. The evidence extends beyond this aggregate comparison: ReWAM outperforms both DriveLaW and DiffusionDrive in all five interaction-intensity groups on NAVTEST. It also ranks first among the methods compared on NAVSIM v2, with an EPDMS of 86.3 under a reactive traffic policy (Table 1). Taken together, the comprehensive evaluation results provide strong support for ReWAM's state-of-the-art performance.

We appreciate the opportunity to clarify the evidence behind this claim.

AuthorsOpenReview ↗

Sensitivity to Training Depth and Level Weights

Q3/W2: How are the training depth and the weight schedule αk∝kp\alpha_k \propto k^p selected, and how sensitive are the results to these choices?

Response: Thanks for your comments on how we chose the reasoning depth and the level weights, and how much these choices affect performance.

Training depth. The submitted paper compares training depths from Ktrain=1K_{\mathrm{train}}=1 to 55. PDMS rises from 90.03 at Level-1 to 90.48 at Level-2 and 90.55 at Level-3, reaching its highest reported value of 90.64 at Level-4. We therefore selected Ktrain=4K_{\mathrm{train}}=4: it achieves the best observed performance without requiring extra reasoning depths.

Level weights. The submitted paper uses quadratic level weights (αk∝k2\alpha_k \propto k^2). To assess sensitivity to the exponent pp, we also trained models with linear weights (αk∝k\alpha_k \propto k, i.e., p=1p=1). At our selected depth of Level-4, PDMS is 90.62 for p=1p=1 and 90.64 for p=2p=2, a difference of only 0.02 points. Across Levels 1–4, the largest absolute difference at any matched depth is 0.05 points at the reported precision (Table 2). These small differences indicate that PDMS is largely insensitive to the choice between linear and quadratic weighting.

We appreciate your suggestion to explore the influence of training depth and level weights on the performance of our framework.

Table 2. Sensitivity to training depth and the level-weight exponent on NAVTEST.

KtrainK_{\mathrm{train}}ppNCDACTTCComfortEPDDCPDMS
1199.0697.3996.2599.9783.2898.1290.03
1299.0697.3996.2599.9783.2898.1290.03
2199.2197.7697.0899.9883.4698.1690.49
2299.2597.6897.1399.9783.4498.1790.48
3199.2497.7197.0599.9883.5298.1490.50
3299.2697.7497.1399.9883.5498.1590.55
4199.2897.7697.1999.9783.5898.1590.62
4299.2597.8197.1899.9683.6098.1490.64
AuthorsOpenReview ↗

Reciprocal Reasoning Versus Additional Prediction Capacity

Q4/W5: How much of the gain can be attributed to reciprocal reasoning rather than additional model capacity?

Response: Thank you for concerns on whether the additional prediction branch can explain ReWAM's gain over DriveLaW.

To test this, we augmented DriveLaW with the same Other Action DiT used in ReWAM. The augmented baseline and ReWAM receive identical inputs during training and inference, including the same annotated current states of surrounding agents. Additionally, both use the same future trajectories for auxiliary supervision and are fine-tuned for the same number of steps.

As shown in Table 3, the augmented baseline raises PDMS from 89.1 to only 89.4, whereas ReWAM achieves 90.6. Thus, 1.2 of the 1.5 PDMS points gained over DriveLaW remain when both models include the additional prediction branch. This gap supports the joint contribution of cross-agent attention and Level-kk updates beyond predicting other agents' trajectories as an auxiliary task.

We appreciate your suggestion to test this distinction under matched conditions.

Table 3. NAVSIM ablation. DriveLaW + Other Action DiT and ReWAM use identical inputs and auxiliary supervision.

MethodNCDACTTCComfortEPPDMS
DriveLaW99.097.196.7100.081.389.1
DriveLaW + Other Action DiT99.097.095.8100.081.789.4
ReWAM (ours)99.397.897.2100.083.690.6
Reviewer_Rh3COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReWAM, which formulates the inter-agent trajectory interactions as a multi-round Level-k game on top of a world action model (WAM). In existing WMs, other agents are modeled passively, not as strategic responders to the ego plan. ReWAM casts the ego agent and the other agents as participants of a Level-k game: at reasoning level kk, each agent's action is a response to the other agents' level-(k-1) actions. In terms of implementation, a frozen Video DiT world model produces the shared latent representation of the future driving world (W_t); a role-specific Ego Action DiT and a parameter-shared Other Action DiT generate their action hypotheses, conditioned on the preceding-level strategy tokens via cross-agent attention. The paper assumes that demonstrated actions are generated by a joint expert policy whose components are optimal responses to the remaining expert policies. The method achieves state-of-the-art PDMS on the NAVSIM navtest benchmark.

Critical strengths

The experiments are built upon a frozen world model, so the introduced design is concentrated in the agent action-prediction heads. And the gains of the proposed method are evaluated across different reasoning levels and across test subsets of different interaction intensities, which makes the benefit of the introduced Level-k game design clear.

Critical weaknesses
  1. In the k-level inter-agent game introduced by the authors, the world tokens predicted by the off-the-shelf DriveLaW are applied unchanged across all game levels. However, from the original purpose of a world model, the world tokens should be tightly coupled with the motion decisions of the ego and the other agents; once the decisions of the ego and other agents change, these tokens would largely lose their reference value. Why did the authors not consider updating the world tokens at each game iteration?

  2. To validate the gain brought by inter-agent interaction, the authors should add an ablation comparing against a simple fine-tuned action head (without inter-agent interaction), and likewise report its performance across different interaction levels. Consider weakness #1, a ablation experiment should also be included in which the prediction head contains only inter-agent interaction without world-token injection.

Questions for authors
  1. At level-1, how are the initial action tokens of agents other than the ego initialized? (the ego's initial action token is the ego future trajectory predicted by DriveLaW?)

  2. Please refer to weakness #1 #2.

3 comments
AuthorsOpenReview ↗

Initialization of Ego and Surrounding-Agent Strategy Tokens

Q1: What initializes the surrounding agents' action tokens at Level-1? Is the ego token initialized from the DriveLaW trajectory prediction?

Response: Thank you for your comments on how the response hierarchy is initialized during training.

At Level-0, we initialize the trajectory of each vehicle, including the ego vehicle, using a deterministic, non-interactive motion prior: we extrapolate from its current pose at its current forward speed. This initialization uses neither future ground-truth trajectories nor the DriveLaW action head.

At Level-1, we encode these initial trajectories as strategy tokens, and each agent conditions its response on the tokens of the other agents. At each subsequent level kk, we encode the trajectories ReWAM predicted at level k−1k-1 and use the other agents' tokens as inputs to each agent's level-kk response.

We appreciate your helpful comments.

AuthorsOpenReview ↗

Rationale for Sharing World Tokens Across Reasoning Levels

W1: Why are the world tokens held fixed rather than updated at each reasoning level?

Response: Thank you for raising this helpful question.

In our research, ReWAM models how agents revise their plans in response to one another within a given driving scene, rather than how the entire scene would evolve under each candidate joint plan. For example, at a merge, the ego vehicle may revise its plan when the neighboring vehicle is predicted to accelerate rather than yield. What changes between these two reasoning steps is the hypothesized behavior of the neighboring vehicle; neither plan has yet been executed, and the observed traffic situation remains the same.

This distinction determines what we update at each level. The world tokens Wt\mathbf{W}_t encode the future scene anticipated from the observation history and navigation instruction, while the strategy tokens encode the participants' current trajectory hypotheses. Since kk indexes reasoning depth rather than elapsed time, we reuse Wt\mathbf{W}_t and update the strategy tokens through cross-agent attention. We agree that executing different joint plans could lead to different future scenes. Capturing those alternatives would require a world model conditioned on the candidate joint plan; re-running the current world model on the same observations would not provide that conditioning. We will clarify that the shared world tokens serve as predictive context, not as a separate rollout for each plan.

Thank you for drawing attention to this modeling choice.

AuthorsOpenReview ↗

Roles of Inter-Agent Interaction and World-Token Conditioning

W2: How does ReWAM compare with a fine-tuned action head without inter-agent interaction across interaction levels? How does an interaction head perform without world-token injection?

Response: Thank you for suggesting ablation studies on inter-agent interaction and world-token conditioning.

(1) Removing inter-agent interaction. We fine-tuned DriveLaW with an additional Other Action DiT but without cross-agent attention or Level-kk updates. This baseline predicts surrounding agents' trajectories, but neither action branch conditions on the other branch's predictions. It receives the same inputs as ReWAM, including current surrounding-agent states, and uses the same auxiliary trajectory supervision and number of fine-tuning steps. As shown in Table 1, PDMS increases from 89.1 for DriveLaW to only 89.4 for this baseline, compared with 90.6 for ReWAM. The improvement supports the joint contribution of cross-agent conditioning and recursive updates beyond adding and training a second prediction head.

Table 1. NAVSIM ablation. DriveLaW + Other Action DiT and ReWAM use identical inputs and auxiliary supervision.

MethodNCDACTTCComfortEPPDMS
DriveLaW99.097.196.7100.081.389.1
DriveLaW + Other Action DiT99.097.095.8100.081.789.4
ReWAM (ours)99.397.897.2100.083.690.6

(2) Removing world-token injection. ReWAM uses DriveLaW as its backbone. In DriveLaW, world tokens provide the keys and values for the Action DiT's cross-attention to the predicted scene. They are therefore required inputs to this block, not an optional feature introduced by ReWAM. Simply removing them would break the backbone's world-to-action conditioning pathway. An interaction-only head would require bypassing or replacing that block and would no longer implement DriveLaW's world-conditioned action architecture, so we have not evaluated that variant.

We appreciate your question about which contributions the ablation isolates.

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

This paper proposes ReWAM, a world-action model for interactive autonomous driving that incorporates Level-k reasoning to model reciprocal interactions between the ego vehicle and surrounding agents. Built upon DriveLaW, the method introduces an additional action model and cross-agent attention to iteratively refine agent actions. ReWAM achieves competitive results on NAVSIM, with particularly notable improvements in scenarios classified as highly interactive.

Critical strengths
  1. Explicitly modeling interactions between the ego vehicle and surrounding agents is a meaningful extension of existing world-action models.

  2. ReWAM improves upon its DriveLaW backbone and achieves competitive NAVSIM performance. The substantial improvement in critical-interaction scenarios is particularly encouraging.

Critical weaknesses
  1. The improvement may result from additional agent-state inputs, multi-agent supervision, or increased training rather than reciprocal reasoning itself. Notably, Level-1 already achieves most of the improvement over DriveLaW, and substantial gains are also observed in non-interactive scenarios. Ablations controlling these factors are missing. The use of ground-truth current agent states also raises questions about the fairness of comparisons with baselines.

  2. Although the method is motivated by game-theoretic best responses, training relies on imitation of demonstrated actions without explicitly optimizing agent rewards. It is therefore unclear whether the learned hierarchy captures strategic responses beyond iterative conditional refinement. Furthermore, NAVSIM's non-reactive evaluation does not directly validate the claimed reciprocal behavioral adaptation.

Questions for authors
  1. Are the current states of surrounding agents available to all compared methods under the same evaluation conditions? How much improvement remains when controlling for these inputs and auxiliary supervision?
  2. How does Level-1 compare with Level-K across different interaction-intensity groups, particularly the Critical group?
  3. How are preceding-level strategies generated during training, and how is Level-0 initialized?
  4. Is there direct evidence that changing the ego vehicle's planned action induces meaningful changes in the predicted responses of surrounding agents?
5 comments
AuthorsOpenReview ↗

Ablation Study on Agent-State Inputs and Auxiliary Supervision

Q1/W1: Do all compared methods have access to the same current agent states? How much improvement remains after controlling for these inputs and auxiliary supervision?

Response: Thank you for question about whether access to agent states and auxiliary supervision could explain the gain over DriveLaW.

To control for additional inputs and supervision, we augmented DriveLaW with an Other Action DiT. The augmented DriveLaW baseline and ReWAM receive identical inputs during training and inference, including the same annotated current states of surrounding agents. Both use the same future trajectories of surrounding agents for auxiliary supervision during training and are fine-tuned for the same number of steps. The augmented baseline predicts surrounding-agent trajectories but uses neither cross-agent attention nor the Level-kk hierarchy.

As shown in Table 1, the augmented baseline improves PDMS from 89.1 to 89.4. ReWAM reaches 90.6: 1.2 points of its 1.5-point gain over DriveLaW remain when inputs, auxiliary supervision, and fine-tuning steps are matched. This result supports the joint contribution of cross-agent attention and Level-k updates beyond auxiliary trajectory prediction alone.

We appreciate your suggestion to make this comparison under matched conditions.

Table 1. NAVSIM ablation. DriveLaW + Other Action DiT and ReWAM use identical inputs and auxiliary supervision.

MethodNCDACTTCComfortEPPDMS
DriveLaW99.097.196.7100.081.389.1
DriveLaW + Other Action DiT99.097.095.8100.081.789.4
ReWAM (ours)99.397.897.2100.083.690.6
AuthorsOpenReview ↗

Effect of Reasoning Depth Across Interaction Intensities

Q2: How does Level-1 compare with Level-K across interaction-intensity groups, particularly in the Critical subset?

Response: Thank you for suggesting this comparison.

Since level-4 achieves the best comprehensive performance, Level-K refers to Level-4 (K=4) in the paper. We therefore added a direct comparison of Level-1 and Level-4 on the same 12,144 NAVTEST tokens, grouped by interaction intensity. As shown in Table 2, Level-4 achieves higher PDMS than Level-1 in all five groups. The largest gain occurs in the Critical group, where PDMS rises from 86.61 to 88.64 (+2.03 points), compared with gains of 0.03 to 0.73 points in the other four groups.

We further compared the individual metrics on the Critical subset (Table 3). Level-4 improves on Level-1 in every component except Comfort. Thus, in the most interactive scenarios, additional reasoning steps improve both safety scores and ego progress without reducing comfort.

We appreciate your suggestion to examine these gains beyond the aggregate score.

Table 2. PDMS by interaction intensity at two reasoning depths. Δ denotes the reported gain from Level-1 to Level-4.

Interaction intensityTokensLevel-1Level-4Δ
None6,62189.2789.66+0.39
Weak1,84190.7291.08+0.37
Medium1,84192.0392.06+0.03
Strong1,28891.5392.26+0.73
Critical55386.6188.64+2.03

Table 3. Metric breakdown on the 553 Critical tokens. Δ denotes the reported gain from Level-1 to Level-4.

MetricLevel-1Level-4Δ
NC95.6696.75+1.09
DAC98.7399.10+0.36
TTC83.0085.90+2.89
Comfort99.8299.820.00
EP88.4489.47+1.03
DDC98.1098.19+0.09
PDMS86.6188.64+2.03
AuthorsOpenReview ↗

Level-0 Initialization and Preceding-Level Strategies Generation

Q3: How are preceding-level strategies generated during training, and how is Level-0 initialized?

Response: Thank you for your comments on how the response hierarchy is initialized and generated during training.

At Level-0, we initialize the trajectory of each vehicle, including the ego vehicle, using a deterministic, non-interactive motion prior: we extrapolate from its current pose at its current forward speed.

At Level-1, we encode these initial trajectories as strategy tokens, and each agent conditions its response on the tokens of the other agents. At each subsequent level kk, we encode the trajectories ReWAM predicted at level k−1k-1 and use the other agents' tokens as inputs to each agent's level-kk response.

We appreciate the opportunity to clarify how the motion prior initializes the recursion and how learned predictions carry it forward.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.