- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper introduces PlanJEPA, a latent world model that augments LeWorldMode (LeWM) with two auxiliary heads during training, something that is said to make the latent space easier to plan in. Alongside a standard nonlinear inverse dynamics head, PlanJEPA uses a multi-horizon linear action objective to train a linear “Plan head” that recovers the first action taken after a current latent state, given that latent state and a randomly sampled future timestep. The key idea behind it is “requiring actions to be predicted through a shared linear map of current and future latent states across multiple temporal horizons” to encourage “the encoder to organize control-relevant variations directly in the representation”.
Critical strengths
- I like how clear the paper's main idea is, and how it presents different experiments to try to support the claim that learning a representation that can be linearly combined with a set of weights to predict an action can lead to more effective solution methods. The paper includes results for ablated methods throughout is a plus.
- Some of the visualizations are quite nice and didactic.
- The method uses no additional test-time computation. During pretraining, the method uses a single inverse dynamics head and a linear PLAN head.
Critical weaknesses
- The work contextualizes itself mostly within the JEPA literature and glosses over what seem to me to be important ideas from the reinforcement learning literature. The successor representation (Dayan, 1993) and more modern variants, such as the successor measure (Touati and Ollivier, 2011), are explicitly designed to be a representation that can be linearly combined with a weight vector to recover a value function. The Laplacian representation has similar properties (Mahadevan, 2005; Gomez et al., 2024). For example, Farebrother et al. (2026) and Shehmar et al. (2026) have explicitly used these methods for planning. I would have expected to see a discussion around these ideas, including the notion of putting together in representation space two states that are close in time, while ensuring that they don’t collapse—these ideas are way older than the JEPA line of work. Ideally, some of these methods would also serve as baselines. The Forward-Backward representation line of work is also likely relevant here (Touati et al., 2023).
- It is unclear to me how meaningful the theoretical results in Section 4 are. Assumption 1 assumes that an agent is able to perfectly predict the action taken, and this is not even in expectation. In large environments, this is bound to be infeasible, and I suspect the same is true for stochastic environments.
- Most of the results do not seem statistically significant, undermining many of the paper's claims. When looking at Tables B.1-B.4, PlanJEPA w/o plan head performs quite similarly to PlanJEPA in most environments (the mean + std. deviation overlap). I imagine that statistical tests across four samples, while correcting for multiple comparisons, would find the results indistinguishable. Other baselines also perform quite competitively in many environments. For example, is Delta-JEPA really worse than PlanJEPA in Tables B.3 and B.4? Even in Table B.2, Delta-JEPA seems to perform worse in very few environments.
- The discussion of the different planners and their assumptions is concise and assumes a lot of background knowledge. While this is not grounds for rejecting a paper, as it is natural to expect some familiarity with the topic, it does not necessarily make the paper easy to read. There’s a middle ground where a brief overview is provided in the paper.
- It is unclear whether the paper reports training the frozen encoder once per environment and that the error bars come only from evaluation noise, or if they come from the representation learning process. Captions such as “mean ± std over four training seeds, each evaluated with three evaluation seeds” make it even more confusing, as it suggests the paper is aggregating different sources of variance.
References
Peter Dayan: Improving Generalization for Temporal Difference Learning: The Successor Representation. Neural Comput. 5(4): 613-624 (1993).
Jesse Farebrother, Matteo Pirotta, Andrea Tirinzoni, Marc G. Bellemare, Alessandro Lazaric, Ahmed Touati: Compositional Planning with Jumpy World Models. ICML 2026.
Diego Gomez, Michael Bowling, Marlos C. Machado: Proper Laplacian Representation Learning. ICLR 2024
Sridhar Mahadevan: Proto-value functions: developmental reinforcement learning. ICML 2005.
Dikshant Shehmar, Matthew Schlegel, Matthew E. Taylor, Marlos C. Machado: Laplacian Representations for Decision-Time Planning. ICML 2026.
Ahmed Touati, Yann Ollivier: Learning One Representation to Optimize All Rewards. NeurIPS 2021.
Ahmed Touati, Jérémy Rapin, Yann Ollivier: Does Zero-Shot Reinforcement Learning Exist? ICLR 2023.
Questions for authors
- Did you tune the hyperparameters of PlanJEPA w/o plan head the same way you tuned the hyperparameters of PlanJEPA? Or did you use the best parameters found for PlanJEPA in PlanJEPA w/o plan head?
- Do your theoretical results assume that the same U and V can satisfy the prediction across all horizons? While you have and in Eq. (5), later you equate them simply to and .
- Did you measure how much Assumption 1 (Locally exact one-step inverse relation) holds in practice? Is it fair to make strong claims based on results with a linear controller when only PlanJEPA was designed to lead to a representation that can be used with a linear relationship? Isn’t it a self-fulfilling prophecy?