VIScore: Diagnosing Planning-Relevant Quality in Latent World Models
Haiyu Wu, Randall Balestriero, Morgan Levine
cs.RO
2026-08-12
VIScore rates latent world models on rollout fidelity, action influence, and search sobriety, topping diagnostics on correlation and calibration across four settings in ~7 seconds.
JEPA-style latent world models such as LeWorldModel (LeWM) learn a compact latent space, predict the future in it, then plan with a search-based controller (model-predictive control with CEM). During training a regularizer called SIGReg pulls the latent toward an isotropic Gaussian to stop the representation collapsing. The persistent gap is that how well regularized the latent is says little about how often planning succeeds. Existing diagnostics make this worse, not better: straightness, physical-state probing, and empowerment all inspect only the encoded latent vector and ignore the predictor and the planner that actually carry out the plan.
The authors first confirm the gap is real with a controlled swap, replacing SIGReg with VISReg, a regularizer that targets the same Gaussian but factorizes into separately weighted center, scale, and shape terms. The properties that help self-supervised learning do not transfer to planning, while finer distribution matching helps out-of-domain planning. That motivates the real question: which factors actually track planning success?
VIScore = Veracity x Influence x Sobriety, one factor per component.
Multiplying the three means a failure in any one component sinks the score. Together the probes take about 7 seconds per checkpoint on one H100, against at least 20x the wall-clock and two orders of magnitude more rollout steps for a single 50-episode planning evaluation. The design choice is that each factor isolates one component (the encoder's contribution enters through Influence), so a low score localizes the bottleneck. This is a diagnostic dashboard, not another benchmark champion.
The yardsticks are Spearman correlation (does the metric rank checkpoints in the same order as success?) and calibration error (are the metric's numbers comparable across tasks?), across four settings, against straightness, physical-state probing, and empowerment.
| Setting | VIScore pooled corr. | Next best | Calibration note |
| Development pool | +0.88 | probing +0.77, empowerment +0.40 | VIScore 10.2 < constant 24.4 |
| Held-out checkpoints | +0.91 | empowerment +0.49 | VIScore 7.0 < 18.2 |
| Held-out methods (4 models) | +0.75 | empowerment +0.12 | largest gap |
| Held-out dataset (MAZE) | +0.87 | empowerment +0.87 (tie) | empowerment calibrates better at 9.2 |
The headline the paper repeats: VIScore is the only metric whose calibration error beats a constant baseline in all four settings. Empowerment also beats the constant in the first three but tips over it on held-out methods (12.0 against an 11.3 baseline), which is exactly where VIScore's lead is widest (+0.75 versus +0.12). MAZE is the one setting VIScore does not sweep: it ties empowerment on correlation and loses on calibration, because MAZE operates at a median of 20.3 nats, far below the 82 cap, so Influence never saturates and adds little discrimination.
The factor ablation shows no single factor dominates: Sobriety leads on the development pool (+0.89), Veracity and Sobriety carry the held-out methods, and Influence only becomes decisive in MAZE where capacity is far from saturated. Transferring VIScore across six planners (CEM, MPPI, iCEM, predictive sampling, multi-start and single-start gradient descent) keeps pooled correlation positive throughout (CEM/MPPI +0.81, iCEM +0.87), so it is not bound to one search family.
The SIGReg-to-VISReg swap also surfaces an uncomfortable detail: the OOD gain is entangled with batch size. At matched batch 128, VIS-WM (77.4) is marginally below SIGReg's LeWM (79.1); VISReg's edge only appears at batch 512 (82.0). The appendix explains why: VISReg's sorted-quantile shape term has a finite-sample floor of order 1/log B, so realizing the gain needs the larger batch.
For world-model practitioners this is a cheap ranking and attribution tool. When selecting checkpoints or tuning regularization, a 7-second VIScore pass can pre-screen before committing to dozens of planning episodes, and it says whether the bottleneck is the predictor, action authority, or planner exploitation. The methodological warnings it carries are just as useful: single-seed evaluation inflates apparent gains, and SSL-friendly regularization does not automatically carry over to planning. Both will undercut a lot of "I swapped the regularizer and gained a point" claims.
What it is not: a replacement for real planning evaluation, or a predictor of absolute success rate. A high or low VIScore does not guarantee success or failure. It is a diagnostic, not a regression.
The authors list several. First, it applies only to search-based planners that contain a predictor; amortized planners, which emit actions directly with no explicit predictor, fall outside its scope, and no diagnostic exists for them yet. Second, Influence assumes the predictor's error directions are roughly stable between one step and H steps; if they rotate, the H-step noise floor is misallocated across eigendirections. Third, on OGBench-Cube gripping success is a discrete event VIScore cannot measure, and memp sits at 236 nats against the 82 cap, so Influence equals 1 on every Cube checkpoint and contributes nothing. Fourth, multi-task planning and training without action labels are untouched.
Two further points seem under-examined. First, VISReg's OOD advantage is essentially bought with batch size: at matched batch VISReg is marginally worse, so "switch to VISReg for better OOD" needs the "and scale the batch" caveat to be true. Second, the all-settings calibration victory rests on a thin margin: empowerment only fails the constant baseline by 0.8 on held-out methods, and on MAZE it actually beats VIScore on calibration, so the headline is not as clean as "VIScore wins everywhere".