pei2333/World4Scorer

Outcome-Grounded World Modeling for Autonomous Driving

Project page · Code (GitHub) · Film · Paper (arXiv)

Jieyuan Pei, Meiyi Lu, Sining Ang, Yubo Zhao, Zhangyi Hu, Mingwei Xu, Haokai Ding, Wei Li, Zihan You, Jianwei Zheng, Li Yu, Yifeng Pan, Ji Tao, Rongjunchen Zhang, Yan Wang

_{The 95-second overview film, with sound. Also on the project page.}

About

Checkpoints of World4Scorer (paper, code).

World4Scorer makes the candidate scorer of a generate-and-select planner a trajectory-conditioned latent world model. One shared predictor gives every candidate trajectory a predicted state, and score heads read its outcomes (collision, drivable area, progress, time to collision, comfort) from that state. Simulator outcomes supervise every candidate; the one future the driving log actually recorded anchors the shared predictor during training. No future is needed at test time.

Results

Benchmark	Metric	World4Scorer	Best other method in the paper
NAVSIM-v2 navtest	EPDMS	93.0 (91.4 without inertial re-ranking)	Drive-JEPA 90.8
NAVSIM-v1 navtest	PDMS	94.0	DrivoR 93.7 · Drive-JEPA 93.7
Bench2Drive-220, closed loop	Driving Score	73.27	ReCogDrive 71.36
OGBench-Cube, LeWM planner	Success rate	73.64%	LeWM 68.91% (same world model and budget)

The Bench2Drive system is adapted (official 1000-clip training subset, a route point and two deployment rules). Cube numbers are our runs over 11 seeds × 50 episodes.

Demos

NAVSIM, Las Vegas. 64 candidates every 0.5 s, coloured by predicted score.	OGBench-Cube. Left: LeWM alone fails. Right: with the outcome score it succeeds.
Bench2Drive. A taxi cuts into the lane (closed loop, 1.5×).	Bench2Drive. A pedestrian steps into the road at night (closed loop, 1.5×).

Checkpoints

File	Model	Result	SHA-256
weights/w4s_navsim_ep23.ckpt	NAVSIM planner, seed 3, epoch 23, 35.5M parameters	91.4 EPDMS · 93.0 with re-ranking · 94.0 PDMS	22604f83dda8a2bd2dd05356fca8457f0c7dc78439f47db396e5c85d74971a44
weights/w4s_b2d_route_tp_ep14.ckpt	Bench2Drive planner, route point, epoch 14	73.27 Driving Score	ed3f61a9ac850627c07bbf1d37cc0a9e123d5b41a9c1cecefe820644e14aff5b
weights/w4s_b2d_route_blind_ep14.ckpt	Bench2Drive route-blind model, epoch 14	initialization of the route-point model	640dd42934b57d9305c336db7234747533977ad916c18723a9bb97b385407d31
ogbench_cube/heads/n10_r1/head.pt	OGBench-Cube outcome head n10_r1	73.64% success	fce54aa53576587f002a8a515db393bc346b1daee2716630ecf14352fb08bf0e
weights/realized_future_emb_navtrain_f7_camf0.pt	NAVSIM future targets (navtrain)	used only in training	2d64716d8f07a0dbc94e3f240108167acf9c9ef5093635eacb857725a13f0d82

From the root of the code repository, this puts every file where the scripts expect it:

hf download pei2333/World4Scorer --include "weights/*" --include "ogbench_cube/*" --local-dir .

The SHA-256 values identify the exact files behind the reported numbers; check a download with python tools/verify_release.py --artifact KIND=PATH in the code repository.

Citation

@article{pei2026world4scorer,
  title   = {World4Scorer: Outcome-Grounded World Modeling for Autonomous Driving},
  author  = {Pei, Jieyuan and Lu, Meiyi and Ang, Sining and Zhao, Yubo and Hu, Zhangyi and Xu, Mingwei and Ding, Haokai and Li, Wei and You, Zihan and Zheng, Jianwei and Yu, Li and Pan, Yifeng and Tao, Ji and Zhang, Rongjunchen and Wang, Yan},
  journal = {arXiv preprint arXiv:2609.36438},
  year    = {2026}
}