Towards Zero-Shot Task Transfer with Neurosymbolic World Models introduces a method to enable reinforcement learning agents to adapt to new tasks without further environment interaction. By decoupling the learning of environment dynamics from reward prediction, the authors allow agents to reuse a single world model for multiple objectives, provided those objectives are defined over the same symbolic state properties.
The Problem with Task-Dependent Models
State-of-the-art reinforcement learning models typically learn latent representations that are tightly coupled with the specific reward function used during training. Because these latent states are uninterpretable, they cannot be easily repurposed when a task changes, even if the underlying environment dynamics remain identical. To adapt to a new goal, current methods usually require either retraining or expensive data relabeling. This paper addresses this by proposing a way to separate the "what" of the environment (dynamics) from the "why" of the task (rewards).
How Neurosymbolic World Models Work
The authors propose Neurosymbolic World Models (NeSy-WMs), which build upon the Recurrent State-Space Model (RSSM) architecture. The key innovation is the introduction of a symbolic bottleneck for reward and continuation prediction.
Instead of predicting rewards directly from the entire latent state, the model is trained to predict a set of user-defined symbolic properties—such as an agent's coordinates or object configurations. These symbolic properties then serve as the sole input for the reward function. Because the reward predictor is now a separate, explicit component, a user can swap it out for a new reward function at test time without needing to retrain the underlying dynamics model. This allows for zero-shot adaptation through either pure planning or imagination-based finetuning.
Supervision and Training
The researchers identify that relying solely on reward signals to learn these symbolic states can be insufficient, as the model might learn "shortcuts" that only work for the training task. To ensure the symbolic states are identifiable and aligned with the intended semantics, they evaluate three supervision regimes:
Full supervision: Providing ground-truth symbolic labels for all states during training.
Partial supervision: Providing labels only for a subset of states, which the authors suggest mimics having sensors in specific parts of an environment.
No supervision: Relying on the model to learn the mapping, which may still provide stability for downstream tasks.
Limitations and Considerations
The authors note that the primary limitation of this approach is the alignment between the predicted symbolic distribution and the ground-truth properties. If the model incorrectly maps a state to a symbolic property, the agent will receive incorrect rewards during test-time interventions, leading to failure. Furthermore, the approach assumes that both training and test tasks are defined over the same vocabulary of symbolic properties. If a new task requires understanding properties outside of this predefined set, the latent model may not be able to accommodate the change.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!