What has been overturned: raw latent distance is not a reliable proxy for decision-relevant reachability
The old assumption was that if a latent world model encodes the state well, then ranking candidate plans by Euclidean distance in latent space to the goal would work. That assumption is now directly falsified: in a hard TwoRoom benchmark, a latent world model (LeWorldModel) with raw latent planning reached only 7.0% success, even though the XY position was linearly decodable from the latent space with R²=0.998 [1]. In plain words: the model 'knew' where the agent was, yet the raw distance metric still misranked candidate plans, because the dimensions that matter for reachability carried less than 1% of the terminal-goal latent MSE [1].
The fix is to replace or augment raw latent distance with a trajectory reachability metric (TRM) trained on logged trajectory structure. With that metric, success jumped to 97.0% on the same task, and a control with shuffled temporal labels stayed at 0.0%, proving the improvement came from the reachability signal, not from overfitting [1]. This overturns the idea that a good latent representation automatically yields a good decision metric—the metric itself must be aligned to the planning objective.
What a fair evaluation must measure: horizon-matched reachability, planner-facing outcomes, and closed-loop impact
First, measure whether the metric is trained on the same horizon as the planning problem. In the TRM study, a short-horizon variant reached only 35.0% success with the same 100,000-pair budget, while the full-horizon version reached 97.0% [1]. So a fair evaluation must include a horizon-matched condition—otherwise you might conclude a metric is bad when it was just trained on the wrong temporal scale.
Second, measure the metric's effect on the planner's actual decisions, not just on prediction error. The TRM study used SCSA audits (a way to inspect the ordering and selected endpoint) and showed that TRM improved both the ranking and the endpoint the planner chose [1]. In continuous manipulation tasks (PushT), TRM-style task-state metrics improved ranking and selected final distance more cleanly than closed-loop success, suggesting that a fair evaluation should include auxiliary hybrid costs and not rely solely on final task success [1].
Third, measure closed-loop performance and the open-loop/closed-loop mismatch. A unified taxonomy for automated driving explicitly calls for a closed-loop metric suite and a resource-aware deliberation cost to reduce the gap between open-loop prediction and closed-loop decision-making [3]. This means a fair evaluation should test the metric in the loop where planning actually happens, not just on static prediction benchmarks.
A fair evaluation must also account for when more compute helps or hurts—depth is not always better
Another layer of alignment is whether adding compute (e.g., deeper rollouts) actually improves decision-relevant predictions. A pre-registered study on nine DeepMind Control tasks found three regimes: depth helped on 6/9 tasks (up to 8x better rollout error), hurt on 2/9 (down to 0.87x), and barely mattered on the rest [5]. The key finding: whether depth helps is not a fixed property of the task but of the operating configuration—training data, metric space, horizon, encoder, and backbone all shift the outcome [5]. So a fair evaluation of decision-metric alignment must include a sweep over these configuration factors, not just a single default setting.
This connects to the broader point that alignment is not a one-time property. The same study showed that a simple dimensionality-only classifier could predict the regime out-of-sample, meaning the effect is stable and predictable once you know the configuration [5]. For a fair evaluation, you should therefore report the configuration and test whether the metric's alignment holds across it—otherwise you might mistakenly attribute a failure to the metric when it was actually a depth/compute issue.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2024 to 2026, 5 from 2024 or later — selected as the most relevant from 5 studies that passed quality screening, drawn from 35 papers retrieved from a database of over 500 million.
Sources used in this answer
Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics
In a hard TwoRoom benchmark, raw latent Euclidean planning achieved only 7.0% success despite near-perfect XY decodability (R²=0.998), while a horizon-matched trajectory reachability metric (TRM) reached 97.0%; a short-horizon TRM variant reached only 35.0% with the same pair budget, and shuffled temporal labels gave 0.0%.
Inference-time Physics Alignment of Video Generative Models with Latent World Models
Using a latent world model (VJEPA-2) as a reward to steer video generation at inference time substantially improved physics plausibility across multiple settings, winning the ICCV 2025 PhysicsIQ Challenge with a score of 62.64%, outperforming the previous state of the art by 7.42%.
Latent World Models for Automated Driving: A Unified Taxonomy, Evaluation Framework, and Open Challenges
A unifying taxonomy for latent world models in automated driving identifies five internal mechanics (including structural isomorphism and value-aligned objectives) and proposes a closed-loop metric suite and resource-aware deliberation cost to reduce the open-loop/closed-loop mismatch.
UniZero: Generalized and Efficient Planning with Scalable Latent World Models
UniZero, a modular transformer-based world model that jointly predicts latent dynamics and decision-oriented quantities, outperformed baselines in long-term memory benchmarks and showed superior scalability in multitask Atari experiments, matching or surpassing state-of-the-art in single-task settings.
Adaptive Compute in Latent World Models: When Depth Helps, Hurts, or Doesn't Matter
A pre-registered study across nine DeepMind Control tasks found three regimes for adaptive compute depth: helped on 6/9 tasks (up to 8x better rollout error), hurt on 2/9 (down to 0.87x), and flat on the rest; the regime was predictable by a frozen dimensionality-only classifier and was configuration-dependent, not a fixed task property.