Advancing the Success–Latency Frontier
for Latent World Models
1Wuhan University 2Zhongguancun Academy 3Harbin Institute of Technology
4Huazhong University of Science and Technology 5Nanjing University 6East China Normal University
7Zhongguancun Institute of Artificial Intelligence 8DeepCybo
† Corresponding authors · Work done during internship at DeepCybo
DynaGoal-Shift
Original models, without fine-tuning. Watch the highlighted goal switches, drift reversal, and doorway motion. Event indicators follow the recorded simulation time.
Cube
Goal switchGoal A → B at 1.25 s
TwoRoom
Moving doorwayDoor shifts down, then returns
PushT
Drift reversalExternal drift reverses at 5.10 s
Reacher
Goal switchGoal A → B at 3.00 s
DynaGoal-Delay
An imposed 800 ms response delay makes the observation-to-action gap visible. The environment continues evolving while a new plan is pending. Cube, Reacher, and PushT use the same initial cases as above; TwoRoom uses a separate recorded case.
Cube
800 ms delayInitial action held; environment keeps evolving.
TwoRoom
800 ms delayInitial action held; environment keeps evolving.
PushT
800 ms delayInitial action held; environment keeps evolving.
Reacher
800 ms delayInitial action held; environment keeps evolving.
