Research map / Large language models / Reasoning and chain of thought

WPT: World-to-Policy Transfer via Online World Model Distillation

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes WPT, a training-time world-to-policy distillation paradigm where a trainable world-model-based reward model guides a multi-modal teacher planner and a lightweight single-modal student learns via query and reward distillation, achieving SOTA open-loop (0.11% collision) and closed-loop (79.23 DS on Bench2Drive) driving.

Problem

End-to-end autonomous driving planners either pay heavy inference overhead when coupled to world models at runtime or rely on offline reward signals that hinder end-to-end optimization; the paper asks how a driving policy can use world knowledge without extra computation at deployment.

Approach

Train a teacher planner with a trainable reward model that scores candidate trajectories against predicted future world states (imitation + simulation rewards), then distill its planning queries and final reward into a lightweight single-modal student so the world model is discarded at test time.

Key results

• WPT-Teacher reaches 0.61m avg L2 and 0.11% avg collision on nuScenes, versus 0.85m / 0.29% for Drive-OccWorld.
• WPT-Teacher achieves 79.23 driving score and 54.54% success rate on Bench2Drive, versus 63.46 DS for DriveTransformer.
• WPT-Student reaches 0.66m / 0.24% on nuScenes at 64 ms latency, versus 312 ms for the teacher with reward model (4.9x faster).

Surprising findings

• Most of the planning gains come from training-time reward supervision; inference-time reward scoring adds only a small margin (0.62m/0.14% to 0.61m/0.11%).
• The WPT-Student has lower comfort score (17.80) than the baseline (23.44) on Bench2Drive, a trade-off the authors acknowledge.

Limitations

• Teacher retains world-model and reward-model overhead at inference, so the best numbers require runtime coupling (312 ms latency).
• Student comfort score is lower than baseline (efficiency-comfort trade-off).
• Simulation reward components are binary or heuristic approximations computed on predicted world states rather than real simulator outcomes.

Experimental setup

Datasets: nuScenes Bench2Drive

Benchmarks: nuScenes Bench2Drive

Metrics: L2 displacement error collision rate driving score success rate efficiency comfort latency

Baselines: UniAD VAD Drive-OccWorld Epona SSR OccWorld RenderWorld GaussianAD DriveTransformer DiffAD WoTE DriveDPO

Builds on: Drive-OccWorld VAD WoTE NAVISIM PDM score UniAD

Closest papers by meaning

• Beyond Prediction: Steering VLM Agents with Retrospective World Modeling (NeurIPS 2026 · similarity 0.89)
• PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.88)
• From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.87)
• RLVR-World: Training World Models with Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry (NeurIPS 2026 · similarity 0.87)
• Reflective Prompted Policy Optimization: Trajectory-Grounded Revision and Salience Bias (NeurIPS 2026 · similarity 0.87)