Now I have all the information needed. Let me write the article.
IMPORTANT: yes
Why World Models Struggle to Learn Tasks in Sequence
A robot that can stack blocks but then forgets how to reach is not useful. The promise of world models, neural networks that learn to predict how environments respond to actions, is that they give an agent a persistent internal simulation of the world. Unlike model-free reinforcement learning, where each new task starts from scratch, a world model could in principle carry its understanding of physics, geometry, and object dynamics from one task to the next. In practice, this rarely happens cleanly. When a model is fine-tuned on a new task, the parameters that encoded the old one get overwritten. When it is frozen to prevent that forgetting, it cannot adapt to the new task at all.
A new benchmark from researchers at Oxford's Applied AI Lab targets this exact tension. Zhou et al. argue that existing continual learning benchmarks conflate two distinct abilities: learning something entirely new, and reusing what you already know. Every new task in a standard sequence brings both novel content and familiar structure, so a model that scores well on adaptation could be succeeding for the wrong reason, brute-force learning instead of genuinely reusing prior knowledge. The paper proposes isolating reuse by designing task curricula that end with a composition task built entirely from primitives the model has already seen. This makes forward transfer on that final task a direct proxy for reuse, while backward transfer on earlier tasks measures forgetting.
How the Benchmark Decomposes Reuse Along Action and Perception
The benchmark consists of six robot manipulation suites built on Meta-World assets. All share the same observation interface, three camera views plus a 77-dimensional proprioceptive vector, and a four-dimensional continuous action space controlling relative end-effector displacement and gripper state. Each suite presents two or three primitive tasks in sequence, then a final composition task that recombines them.
The key design choice is factoring composition along the two axes a world model actually conditions on: action and perception. This decomposition is not arbitrary. It maps directly to how world models process their inputs. Perception enters through an encoder that compresses visual observations into a latent state. Actions condition the dynamics model that predicts how that latent evolves. By varying one axis while holding the other fixed, the benchmark localizes where reuse succeeds and where it breaks.
Action-composition suites keep the scene and objects constant while combining motion primitives. In the Reach suite, the model learns to reach in the xy plane, then xz, then yz, and finally must reach in all three dimensions simultaneously. The Grasp suite teaches pick and place separately, then tests pick-and-place. Perception-composition suites hold the manipulation approximately fixed while varying visual appearance. BinPnP moves cubes between colored bins in sequence, then tests an unseen color combination. PnPBlock manipulates red and blue blocks, then must stack them. Full-composition suites change both axes at once. DrawerPnP combines pick-and-place with opening a drawer, then tests picking and placing inside an opened drawer. PnPBoxClose combines pick-and-place with closing a box, then tests placing inside a closed box.
The difficulty ordering is not subtle. BWT worsens monotonically from action to perception to full composition on both backbone architectures. DreamerV3 maintains forward transfer within a narrow band across all three axes, while TD-MPC2 loses it entirely once perception changes. The reason maps to the architecture: recomposing actions under a familiar scene shifts only the action distribution over a stable visual latent, whereas a perception shift disturbs the encoder and propagates through the dynamics. Full composition hits both at once.
Separating Task-Agnostic Dynamics from Task-Specific Heads
The paper introduces a structural principle for applying continual learning to world models. Rather than treating the entire model as a monolithic blob, the authors partition it into a task-agnostic backbone and task-specific heads. The backbone, comprising the encoder and dynamics model, captures how the world evolves and is updated continually across the curriculum. The heads, including reward prediction, actor, and critic, consume the latent produced by the backbone and are specific to each task. They are re-initialized for each new task and frozen after training, with earlier tasks evaluated using their stored heads against the current backbone.
This separation is not just an organizational convenience. It determines where continual learning pressure is applied. The backbone must absorb new dynamics without destroying old ones. The heads only need to learn their own task and are never asked to remember anything. This mirrors a natural division: the physics of how objects move is shared across tasks, but what counts as success differs. Algorithm 1 in the paper formalizes this: for each task, fresh heads are initialized, the backbone is updated using the chosen continual learning method, the heads are trained independently, and then both are evaluated across all prior tasks.
The authors evaluate four continual learning paradigms applied to the backbone: naive fine-tuning, experience replay maintaining a buffer of 5 percent of each prior task's transitions, elastic weight consolidation penalizing updates to important parameters, and PackNet which prunes and freezes sub-networks after each task. The results on these monolithic backbones reveal a clear trade-off. Fine-tuning and ER achieve strong forward transfer but suffer heavy forgetting. EWC prevents forgetting but also prevents recombination, scoring worst overall. PackNet comes closest to balancing both, with near-zero forgetting and competitive forward transfer, but its capacity shrinks with each task because it allocates a fixed fraction of parameters per task, limiting scalability.
The Modular Alternative: Mixture-of-Experts Dynamics
The Prismatic World Model (PWM), a mixture-of-experts extension of TD-MPC2, provides the modular baseline. In the standard multi-task formulation, PWM uses multiple dynamics experts whose outputs are combined by a learned router. The authors construct a continual variant where each incoming task introduces three new trainable dynamics experts while all prior experts are frozen. The router can access all experts from all tasks encountered so far, enabling reuse of previous dynamics while adding new capacity.
Comparing PWM to monolithic TD-MPC2 requires careful control. To isolate what the modular dynamics design contributes, the authors run a frozen-encoder diagnostic: pretrain PWM's encoder with a simple autoencoder reconstruction objective on demonstrations from all tasks, then freeze it throughout sequential learning. The same frozen encoder checkpoint is also used on TD-MPC2. Any remaining performance gap is attributable to the modular dynamics architecture itself, not to encoder differences.
The results are striking. Averaged across all six task suites, PWM with the frozen encoder achieves BWT of 3.85 (near zero forgetting) and FWT of 36.18, while TD-MPC2 with the same encoder scores 34.97 and 38.11. PWM essentially eliminates forgetting while matching the monolithic model's forward transfer. The modular design eliminates the capacity-shrinkage problem of PackNet because new experts are added rather than carved from a shrinking pool, making it more scalable across long task sequences.
But the frozen encoder is a privileged condition. It has access to the full task distribution up front, which is not available in a true continual learning setting. Without it, PWM performs on par with TD-MPC2, meaning the encoder's ability to maintain a stable representation remains an open problem.
What the Router Actually Learns
The most revealing analysis is the probing of PWM's routing behavior. The authors examine whether the router's expert allocation reflects the known task composition structure. For several suites, it does. In BinPnP, during the third primitive task, the router increases weight on experts from the overlapping prior task to 1.20x while reducing a non-overlapping expert to 0.57x. In Reach and Grasp, experts associated with recurring components retain more weight while unrelated ones are down-weighted.
The pattern breaks in PnPBlock and both full-composition suites. The authors attribute this to the router being task-conditioned rather than state-conditioned. A fixed expert allocation per task cannot switch between primitives as a task progresses temporally. Stacking two blocks requires grasping one then the other, a temporal sequence that a single fixed routing cannot capture. Full composition suites involve transitions between primitives, not just their simultaneous presence.
Expert ablation confirms that routing patterns correspond to functional contributions. Prior experts alone retain near-baseline performance on BinPnP. They preserve substantial return on Reach and DrawerPnP, though prediction fidelity drops more sharply. New experts alone perform poorly across suites. Uniform routing falls well below the full mixture. The combination of prior experts and learned routing is what makes composition work. Expert-level preferences also persist: the highest-weighted expert from each primitive remains top-ranked after composition learning in 13 of 14 cases, and weight distributes across both prior and new experts rather than concentrating on the newest capacity.
Where Reuse Without Forgetting Remains Unsolved
The benchmark exposes several open problems that the current methods do not address.
The encoder drift problem is the most fundamental. All analyses showing PWM's superiority depend on a frozen encoder pretrained with knowledge of the full task distribution. In a true continual setting, the encoder must itself learn continually without a stability-plasticity mechanism. DreamerV3's generative reconstruction objective helps, encoding more broadly than the current task requires, but it is not enough for full composition suites. A world model that learns representations continuously while preserving their utility for prior tasks remains unsolved.
The temporal routing problem limits modular architectures. Task-level routing produces a fixed expert mixture that cannot vary within an episode. Temporal composition, where different primitives activate at different stages, requires routing that adapts over time. State-conditioned routing, where the expert allocation depends on the current latent rather than the task identity, is the natural next step but is not explored in this work.
The capacity trade-off between PackNet and PWM is instructive. PackNet freezes sub-networks and reuses them at inference by activating all prior allocations, which prevents forgetting but shrinks available capacity. PWM adds experts rather than freezing subsets, avoiding the capacity problem but introducing the risk that new experts simply learn from scratch without reusing old ones. The router is supposed to handle this reuse, but its limitations in temporal and full-composition settings suggest that architectural reuse alone is insufficient without proper routing.
The difficulty ordering across composition axes points to a concrete research direction. Action-composition is relatively easy because the dynamics model sees a shifted action distribution over a stable visual latent. Perception-composition is harder because visual changes propagate through the encoder into the dynamics. Full composition is hardest because both inputs change simultaneously. A model that handles full composition would need both a stable visual representation and a dynamics model capable of recombining action primitives, neither of which current architectures provide simultaneously.
What This Means for Building Better World Models
For practitioners building continual learning systems, the benchmark offers several concrete takeaways.
Experience replay is the simplest approach and gives the best forward transfer, but it does not stop forgetting and the buffer grows without bound. If you can tolerate some performance degradation on earlier tasks, ER is a reasonable default. PackNet achieves near-zero forgetting but its fixed parameter allocation per task limits how long the curriculum can be. If the task sequence is short and bounded, PackNet is the most reliable choice for preventing forgetting entirely.
The modular approach with mixture-of-experts dynamics is the most promising direction for long task sequences. Adding experts per task avoids the capacity shrinkage problem, and the router learns to combine them in ways that reflect task structure. But the encoder must be handled carefully. Training it from scratch in a continual setting negates the modular advantage. Pretraining the encoder on demonstrations from the full task distribution gives a significant boost, but this is a luxury not available in truly lifelong learning scenarios.
The frozen-encoder diagnostic reveals that the dynamics model itself is where modular reuse succeeds. Prior experts, when routed correctly, retain functional value even when the router does not perfectly match the expected composition structure. This suggests that architectural modularity in the dynamics model is on the right track, but the representation layer feeding into it needs its own continual learning mechanism.
For researchers, the benchmark provides a structured evaluation protocol with clear metrics. Backward transfer measures forgetting as a normalized area under the evaluation curve, ranging from 0 to 100. Forward transfer measures how much faster and better the composition task is learned compared to scratch. Both metrics are computed over the full training budget rather than at a fixed threshold, rewarding both speed and quality of learning. The per-suite breakdown by composition axis tells you exactly where a model fails, which is more actionable than a single aggregate number.
The Oxford team's work makes clear that the field is not yet at the point where world models can learn a sequence of related tasks and genuinely reuse prior knowledge. The compositional benchmark isolates the reuse problem, the model separation principle tells you where to apply continual learning pressure, and the modular baseline shows that architectural priors help but are not sufficient alone. The next step is a world model with a continually learned representation that maintains stability without privileged access to future tasks, combined with dynamics routing that adapts within episodes rather than across them.