Selected work KV / Research 01 Read paper
LeWMRO · ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation · Oral

When a useful world model looks broken.

A planner can optimize a future it never executes. And a latent state can look close to the goal without offering a controllable route there. We isolated both interface failures and showed what changes when the evaluation matches the controller.

Role Project lead · co-first author Setting Offline model, online replanning Tools LeWM · MPC · CEM

Stand inside the planning loop.

A staged world for both failure modes. Watch a fifteen-step plan grow and get thrown away three chunks at a time, then steer an agent through the detour that a scalar objective cannot see. Every number in its console is quoted from the paper. The motion is staged, not replayed.

Still image · click to load the world Open on its own page ↗

A world-model benchmark does not measure the model alone. It measures a chain: representation → objective → proposal mechanism → actuator. A mismatch anywhere in that interface can turn a usable model into an apparent failure.

The planner imagines fifteen steps. The controller takes three.

Receding-horizon MPC plans to horizon H, executes only K action chunks, observes the new state, and replans. Terminal@H scores a state the controller will not directly reach before feedback changes the plan.

Protocol explainer · not a live rerun

What exactly gets scored?

score ẑH
H = 15
nowimagined future
executed prefix

The objective rewards the imagined endpoint at H=15, although feedback arrives after K=3. The unexecuted suffix can dominate which prefix is selected.

Measured on PushT · H=15 · K=3
17.0%terminal@H
94.5%prefix@K
86.5%running

The same check changes TwoRoom from 52.0% to 83.5%; a running cost changes Reacher from 20.5% to 100%.

Move the feedback boundary without changing the model.

Across the measured replanning intervals K ∈ {1, 3, 5, 15}, terminal scoring is fragile under partial execution while running cost stays strong with frequent feedback. The intermediate PushT values are not monotonic. The finding is alignment, not a smooth performance law.

Measured PushT and TwoRoom terminal and running-cost success across replanning intervals
Success at H=15. Each plotted value comes from the paper’s measured sweep; no values are interpolated.

Closer in latent space does not always mean reachable.

TwoRoom Far Door forces the agent to move away from the final goal to reach a distant doorway. Even aligned scalar objectives remain weak because they reward similarity, not locally executable progress through the bottleneck.

Best scalar/value objective26.5%
Waypoint + nearest retrieval92.7%
Matched-configuration reproduction · held-out fold 0 · episode 04 / 12

Watch the interface change the first move.

H 15 planK 3 execute replan
Best scalar/value formepisode failed
Prefix-value objective, α = 0.50; no route is supplied.Closest approach 39.70 · success requires distance <16.
Nearest waypoint + local actuatorepisode succeeded
first post-step hit held
The ring is the solver’s recorded waypoint, updated every K=3 chunks.Goal distance 12.85 < 16 after transition 49; the post-step state is held.

Matched-configuration reproduction on held-out fold 0. The agent starts left of the far doorway; the waypoint controller first redirects it through that opening. TwoRoom counts a hit when agent-goal distance is below 16; success is reach-once within the 75-step budget. The waypoint library and local actuator exclude that evaluation fold. The overlay is the solver's recorded selection, not an inferred path. The 92.7% headline is the paper's five-fold mean, not the outcome of this clip.

The useful first move is away from the goal.

The stored expert trajectories descend to the low far doorway, cross the wall, then climb toward the high goal. The waypoint interface turns that detour into a sequence of locally executable edges.

Actual TwoRoom Far Door expert dataset trajectories and representative start, middle, and final frames
Quicklook generated from 400 successful expert episodes used to construct and audit the deceptive task.

I led the experiments that separated the two failure modes.

01

Separated H from K

Turned an implicit evaluation choice into the central controlled variable.

02

Built diagnostic controls

Compared terminal, prefix-terminal, running-cost, value, retrieval, and waypoint interfaces.

03

Designed deceptive tasks

Created detour settings that distinguish latent similarity from controllable progress.

04

Made the claim narrower

Reported strong retrieval controls and reset-state artifacts instead of hiding inconvenient evidence.

A diagnostic, not a universal planner.

Hidden Failure Modes in Latent World-Model Planning from Offline Data

The evaluation interface is part of the method.

Open the 8-page paper ↗ OpenReview ↗ Code ↗ Back to selected work