Frozen World Models Plan Better When Aimed at Nearby Goals, Not Far Ones

Robots that plan with visual world models typically imagine possible futures, then pick the action whose predicted outcome looks closest to an image of the final goal. A new paper argues that this seemingly natural strategy has a hidden flaw: sometimes reaching a goal requires actions that initially move away from it, and scoring purely against the final target can cause a planner to miss this entirely.

The paper, titled "Aim Short to Reach Far: Your Frozen World Model Can Plan Better Than You Think," was posted to arXiv (arXiv:2609.30036) by author Xvyuan Liu. It shows that even with exact dynamics and globally optimal short-horizon search, planning toward a single distant goal image can limit control performance. The issue isn't the world model itself, the authors argue — it's what the planner is told to aim for.

Using frozen LeWM models — pretrained world models that are not further fine-tuned — across four benchmark tasks (Cube, PushT, Reacher, and TwoRoom), the researchers tested what happens when planners are given intermediate targets instead of always scoring against the final goal. The gains were substantial: intermediate targets meaningfully improved both action synthesis and the ranking of recorded action sequences. This held true whether the intermediate targets were learned or simply pulled from observed past experience.

Building on this finding, the team introduces a technique called Anchored Planning. Rather than retraining anything, it retrieves a recorded segment from past experience whose start and end observations resemble the robot's current state and its ultimate goal. Instead of aiming at the distant goal, the planner targets an observation captured shortly after the start of that retrieved segment — a nearer, more immediately achievable waypoint. The same frozen world model then scores candidate actions against this closer target.

Crucially, this requires no additional training of either the world model or the planner. Yet in long-range evaluations, planning toward these observed intermediate targets outperformed the previously released LeWM planner on every one of the four tasks tested, according to the paper. Simply adding more search effort toward the final goal did not produce the same gains — the improvement came specifically from changing what the planner aims at, not from searching harder.

The study also surfaces a finding worth noting for the field: lower prediction error in a world model does not necessarily translate into better control outcomes. The success of Anchored Planning depended on two further factors — how far ahead the intermediate target was placed, and progressively shrinking the span of retrieved segments as execution advanced toward the goal.

The central message, as reported in the paper, is that the underlying frozen model and planner can remain exactly the same; only the target needs to change. That single adjustment let the system reach goals that conventional final-goal scoring missed entirely.

Because the available material is the paper's abstract, some details — including the precise architecture behind "LeWM" and specific quantitative success-rate figures — are not spelled out here and would require the full paper for confirmation.