acceptodds
Withdrawn from ICLR 2027

Rethinking the Predictive Loss in Latent World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Latent world models rely on predictive penalties to learn representations, meaning the choice of loss function dictates how prediction errors shape the latent space. The derivative of standard squared-error loss grows linearly with the error, amplifying the influence of large prediction errors. To address this, we investigate the element-wise log-cosh loss, which behaves quadratically for small errors but smoothly bounds the derivative for large ones. Through residual-space analysis on models with explicit representation regularization, we show that this attenuation improves local gradient interaction where it reduces gradient conflict more than cooperation. We applied this training-time loss replacement across multiple latent world model families without changing their underlying architectures or inference methods. On navigation, manipulation, and continuous control benchmarks, log-cosh consistently improved planning success over squared error and outperformed a locally curvature-matched quadratic control. On real-world driving data NAVSIM, log-cosh also improved perception and planning. We further study the learned representations through physical probing and action-space analyses. We observe that log-cosh makes task-relevant state information more accessible, and action-space analyses reveal closer alignment with the true-system cost landscape. These findings establish the predictive penalty as a fundamental component of representation design.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Latent world models, Predictive Loss

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CD7iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how the choice of predictive loss shapes the representations learned by latent world models. The method is a training-time swap: squared error is replaced by an element-wise log-cosh penalty, which is quadratic near zero and smoothly saturates for large residuals. No architecture or inference procedure changes, and the planner keeps its original squared latent goal cost. The theory is a residual-space decomposition (Prop. 1) showing that, relative to a curvature-matched 1/2 r^2 control, log-cosh improves the local gradient interaction with an explicit representation objective if and only if the attenuated coordinates remove more conflicting interaction than cooperative interaction. The evaluation covers the TS family (4 tasks, 2 encoders, 2 latent shapes, open-loop and MPC), LeWM (4 tasks, 2 execution modes), longer-horizon planning, and real-world driving on NAVSIM with perception MAE and planning ADE/FDE, plus physical-state probes and geometric visualizations. Log-cosh consistently outperforms both r^2 and the matched quadratic across all reported settings.

Critical strengths
  1. The curvature-matched quadratic control is the strongest part of the paper. Matching the second derivative of log-cosh at the origin while keeping quadratic growth rules out the trivial explanation that the gains come from uniformly shrinking the predictive term. Across Tables 2-5 the matched control tracks r^2 closely while log-cosh consistently improves on both. This isolates the residual-dependent, nonlinear attenuation as the active ingredient and gives the empirical claim real weight.
  2. The breadth and internal consistency of the evaluation. Two model families with different representation objectives and planners, two encoders, two latent shapes, open-loop and MPC, longer horizons, real driving data, probes, and geometric analyses. I verified the ordering in every cell: log-cosh improves on r^2 in all 32 TS comparisons and on both quadratics in all 8 LeWM comparisons, and it ties the matched quadratic exactly once (Wall under MPC with ResNet spatial latents, both at 100%). The appendix arithmetic all reconcile
Critical weaknesses
  1. The central mechanistic claim is never tested, and the theory behind it is tautological. Prop. 1 is an algebraic identity that restates the definition (A_lc exceeds A_q if and only if removed conflict exceeds lost cooperation). It gives no characterization of when that condition holds. More importantly, the paper never measures residual magnitudes, coordinate-level interaction terms alpha_i, or gradient angles during training, so we do not know whether the mechanism is even active, or how often the condition holds. The evidence supports the statement that log-cosh helps, but it does not support the causal story in the conclusion that this happens through nonlinear residual attenuation reducing gradient conflict. This is the main factor in my rating.
  2. Missing comparisons within the robust loss family put the novelty at risk. The paper itself cites Huber, Charbonnier, Barron, and the known stability benefits of L1-type penalties in V-JEPA, yet only three points are compared: r^2, 1/2 r^2, and log-cosh. If a pseudo-Huber loss performs equally well, the contribution reduces to applying a known robust loss, which would not clear the bar. Relatedly, the analysis is stated for raw gradients, and the only optimizer the paper specifies is the driving recipe (AdamW with global gradient clipping at 1.0); the optimizer settings for TS and LeWM are inherited from the prior protocols and not restated. Under Adam-style normalization, halving the loss changes the effective update scale very little, so the matched control plausibly entangles an effective learning-rate or gradient-scale effect that the paper does not fully rule out.
  3. Data provenance and completeness concerns. The AI use statement acknowledges that earlier drafts contained AI-generated hypothetical numerical examples and that some provenance is unresolved. Appendix C.2 reports curvature ablation gains (56 and 78 percentage points) while explicitly stating that absolute scores and uncertainty are not reconstructed. The caption of Figure 11 is an unrevised placeholder note ("This panel is reserved for a comparison of r^2 and log-cosh on the same scenes") that does not describe the figure. The appendix section headings are misplaced: D.1 (Self-Supervised Data and Architecture) is empty, and all of its content appears under D.2, which carries a copied title (Physical Structure of the Latent Space) that describes Appendix E instead. Until these are clarified, the credibility of parts of the empirical record is uncertain.
  4. The headline driving experiment is not reproducible as described and carries no uncertainty quantification. The planner used for ADE and FDE (cost function, horizon, search method) is never described. The protocol for the five perception MAE probes is not described. The separate pretraining data is not identified. All driving results are reported as single scalars with no error bars and no seed count. With 3 seeds and 50 episodes per seed, several gains elsewhere are 2-4 percentage points, within the range of the reported seed-level standard deviations. For a 400 GPU-hour experiment, the absence of any variance estimate makes the strongest real-world claim impossible to assess for significance from the information given.
Questions for authors
  1. Does every number in the submitted version have a corresponding experiment log? What are the absolute success rates behind the 56 and 78 percentage point gains in Appendix C.2, and which specific values does the unresolved provenance refer to? This directly determines my confidence in the empirical record.
  2. What do residual magnitude distributions look like during training for each model, and can you report the fraction of heavily attenuated coordinates (large delta_i) whose interaction alpha_i is negative, so that Prop. 1's condition is shown to actually hold? Also, what is the reason for preferring log-cosh over Huber or Charbonnier, and why is that comparison absent?
  3. How are the driving ADE and FDE produced, meaning which planner, cost, and horizon are used? How many seeds per configuration and what is the variance? If these are single runs, the 9.6 to 13.9 percent relative reductions cannot be judged for significance.
Reviewer_98sGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies replacing squared prediction loss with log-cosh in latent world models. It argues that attenuating large residuals can improve the interaction between prediction and representation regularization. It reports gains across TS, LeWM, and a driving setup, along with physical-state probes and planning-cost visualizations.

Critical strengths

I like the question. When the encoder and predictor are trained jointly, changing the prediction loss can affect what information the representation keeps. This seems worth studying.

The reported gains are fairly consistent across TS and LeWM, and some are substantial. Keeping the planner and inference-time goal cost fixed is a useful control. The physical-state probes also give some evidence beyond planning success alone.

Critical weaknesses
  1. The most obvious missing comparison is with other robust losses. L1 and Huber/SmoothL1 seem like natural baselines, especially since the paper itself notes that V-JEPA already uses L1. With only two quadratic baselines, it is hard to tell what is specific to log-cosh.

  2. The half-scaled quadratic control does not fully settle the loss-weighting question. In TS and LeWM, log-cosh changes the effective balance between prediction and regularization, so a coefficient sweep for the quadratic baseline would help. The driving comparison has an additional issue: both branches share the same penalty, so using half-scaled squared loss simply halves the entire stated objective. Ignoring epsilon and clipping, this global scale cancels in AdamW’s moment normalization. The similar driving results therefore add little evidence for the claimed mechanism.

  3. I am also not quite convinced by the mechanism. Proposition 1 gives a correct decomposition into removed conflict and lost cooperation, but I could not find measurements showing which term dominates during training. Even the residual distribution is missing. For example, at a residual magnitude of 0.5, the attenuation weight is still about 0.92. It would help to know where the actual residuals lie and which magnitude ranges contribute most to the gradients. The conclusion says the controls and visualizations “confirm” the mechanism, which seems stronger than the evidence supports. Section 4.6 itself acknowledges the limits of the visualizations.

  4. Some experimental details remain unclear. The driving section needs a proper downstream training and evaluation protocol, including pretraining data, prediction heads, evaluation split, metric definitions, and planning horizon. The longer-horizon experiments should also state the actual horizons. For uncertainty, the probe error bars only reflect probe initialization. The TS results use “data-sampling seeds,” but it is unclear whether each seed includes a separately trained world model.

Questions for authors
  1. Could you compare against L1 and Huber/SmoothL1, and check whether tuning the quadratic-loss coefficient closes the gap?

  2. What do the residual distributions look like after the native normalization? At shared checkpoints and minibatches, does log-cosh actually remove more conflicting interaction than cooperative interaction? Gradient norms and cosine similarities would also help interpret the unnormalized interaction measure.

  3. Does each TS data-sampling seed involve retraining the world model? Which sources of randomness are included in the reported standard deviations?

  4. Could you clarify the driving downstream protocol and give the actual horizons for both the standard and longer-horizon planning experiments?

  5. The AI use statement on page 10 mentions hypothetical numerical examples in earlier drafts and unresolved provenance. Are all empirical results in this submission verified measurements from actual runs? Are any displayed values illustrative or still awaiting verification?

Reviewer_Sen4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose using a different loss function when training world models. They propose replacing the L2 part of the standard next-step prediction objective in latent space (L2 distance between predicted and target latent) with an element-wise log-cosh loss. Their motivation is that the standard L2 objective penalizes larger prediction errors more, which is undesirable according to the authors. They test this on two existing ways of training world models:

  • Training with stop-gradient to prevent representation collapse, and using temporal straightening as a regularization (TS).
  • Training using SIGReg to prevent collapse (LeWM). They do experiments on the Wall, PointMaze-UMaze, PointMaze-Medium for TS, and on TwoRoom, PushT, Reacher and Cube for LeWM. They also do experiments on real-world driving data (NAVSIM). They compare planning metrics and perception metrics across models trained with their log-cosh prediction term vs models trained with the standard L2 prediction term.
Critical strengths
  • The proposed solution is simple (using a different loss function) and can be applied to any method for training a world model.
Critical weaknesses
  • The writing is poor and not rigorous. Functions and scalars are mixed (log-cosh is a function and r^2 is a scalar value). Some terms used are not standard nor precise (readout, cooperative updates, conflicting updates, “when it removes more conflict than the cooperation it sacrifices.”).
  • We do not know what the training objectives are, nor what is being trained. The encoders and predictors are not defined.
  • The planning protocol is not explained. Wall (TS experiments) and TwoRoom (LeWM experiments) are two names for the same environment.
  • Regarding the motivation of the paper, I’m not sure I understand why penalizing larger prediction errors more than smaller prediction errors would be an issue in the first place.
Questions for authors
  • “Abstract: “Through residual-space analysis on models with explicit representation regularization, we show that this attenuation improves local gradient interaction where it reduces gradient conflict more than cooperation.” What does this mean?”
  • What does “outperformed a locally curvature-matched quadratic control” mean? What is “control” here?
  • What is vec in l154?
  • Objective (1): What is the embedding? What is the predictor? No notation is introduced at all, so we don’t understand what the objective is. What is trainable?
  • Why would r² and ½ r² have different effects during training?
  • For the long-horizon tasks, is the planning horizon also increased?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.