ReWAM: Reciprocal World Action Models for Interactive Autonomous Driving

• End-to-end policies map observations directly to actions, but action supervision alone does not ensure explicit environmental understanding or out-of-distribution generalization.Source: §1 · p. 1
• Language-centered methods strengthen semantic reasoning, yet do not explicitly model continuous physical evolution under different actions.Source: §1 · pp. 1–2
• Existing WAMs condition future worlds on ego actions, while other agents are generally predicted rather than modeled as strategic responders.Source: §1 · p. 2
• Interaction-aware planning models reciprocal responses, but typically separates interaction modules from generative world models and uses explicit states, maneuvers, or actions.Source: §1 · p. 2
• ReWAM embeds bounded strategic reasoning in WAMs, with mutually responsive ego and other-agent action branches grounded in shared future-world representations.Source: §1 · pp. 2–3
• On NAVSIM, ReWAM achieves state-of-the-art performance, with particularly pronounced improvements in dense-interaction scenarios.Source: §1 · p. 3

2.1 Level-k Game for Solving Interactive Policy

• Level-k reasoning replaces simultaneous multi-agent coupling with finite strategic responses, preserving reciprocity without online joint-equilibrium computation.Source: §2.1 · p. 3

2.2 World Representation Conditioned Level-k Reasoning

• A shared future-world representation grounds separate ego and parameter-shared other-agent Action DiTs, whose responses condition on other participants’ preceding-level strategies.Source: Figure 2 · p. 4, §2.2 · p. 4

2.3 Learning Best Responses from Demonstrations

• Expert joint demonstrations are treated as reward-maximizing responses, and conditional response models learn a stochastic best-response surrogate over demonstrated strategies.Source: §2.3 · pp. 4–5

2.4 Level-k Interaction Block

• Each Action DiT combines temporal self-attention and world-feature cross-attention with cross-agent attention to other valid participants’ preceding-level action hypotheses.Source: §2.4 · p. 5

2.5 Joint Learning and Inference

• Conditional flow matching jointly trains both response roles across supervised reasoning levels, using normalized depth weights and role-balancing coefficients.Source: §2.5 · pp. 5–6

3.1 Experimental Setup

• Four front-camera frames sampled at 2 Hz condition predictions of eight poses at 0.5-second intervals for ego and up to ten other agents within 80 m.Training and evaluation use NAVSIM’s navtrain and navtest splits; navtest contains 136 logs and 12,146 scenarios.Source: §3.1 · p. 7

3.2 Main Results

• PDMS 90.6 gives ReWAM the highest score, 1.5 points above DriveLaW and 0.2 above ReWorld.ReWAM improves NC and EP over ReWorld, which has slightly higher DAC and TTC.Source: §3.2 · p. 7

3.3 Analysis of Level-k Reasoning

• PDMS rises from 90.03 at Level-1 to 90.64 at Level-4, with the largest single-level gain at Level-2 and diminishing returns at greater depth.Level-5 adds no PDMS gain and slightly reduces DAC.Source: §3.3 · p. 7, Table 2 · p. 8
• ReWAM outperforms both baselines in all five interaction-intensity groups, with its DriveLaW advantage increasing from 1.52 PDMS in None to 4.59 in Critical.In Critical, ReWAM also surpasses DiffusionDrive by 2.54 PDMS.Source: §3.3 · p. 7, Table 3 · p. 8
• In Critical scenarios, ReWAM improves all non-saturated metrics over both baselines while retaining perfect Comfort, with safety gains alongside higher compliance and progress.The reported PDMS gain reflects improvements across safety, traffic-rule compliance, and driving progress rather than a single metric.Source: §3.3 · p. 7, Table 4 · p. 8

3.4 Qualitative Results

• Across three interactive cases, ReWAM completes a narrow passage, follows a left turn, and avoids a lead-vehicle conflict without unnecessary progress loss.The baselines respectively fail the narrow traversal, miss the turn, or exhibit excessive deceleration or insufficient response.Source: Figure 4 · p. 8, §3.4 · p. 9

4 Conclusion

• ReWAM achieves the best NAVSIM performance with PDMS 90.6, and its largest advantage over DriveLaW occurs in the Critical interaction group.Recursive reasoning gains saturate at deeper levels; stratified results show advantages over DriveLaW and DiffusionDrive across all interaction groups.Source: §4 · p. 9
• The framework uses a finite hierarchy of bounded responses, shared future-world grounding, cross-agent attention, and conditional flow matching trained from demonstrations.It avoids online equilibrium computation while learning response distributions across the hierarchy.Source: §4 · p. 9
• The authors present bounded reciprocal-response reasoning as a practical foundation for interactive world-action modeling, with gains spanning safety, traffic-rule compliance, and ego progress.Qualitative examples are also cited as evidence that gains are not simply from conservative planning.Source: §4 · p. 9