Neurosymbolic Learning of Interpretable World Models
CGWM targets the opacity of neural world models by forcing latent states to be explicit symbolic concepts grounded from images by a neural encoder. It replaces learned transition dynamics with hand-specified LTLf rules and performs planning via gradient ascent on fuzzy temporal formulas. The key appeal is interpretability and zero-shot transfer under changing dynamics, though the evaluation is limited to small, fully observable deterministic grids.
CGWM presents a principled neurosymbolic architecture that yields strong concept-grounding and planning results on toy grid-worlds, but its heavy dependence on complete, hand-engineered symbolic knowledge and deterministic fully-observable assumptions severely limits applicability. The advance is less about learning a world model from raw experience than about coupling a neural encoder to a pre-specified logical simulator.
The authors candidly acknowledge these limitations in Section 7, yet the framing and title of the paper invite broader conclusions than the narrow experiments support.
The fuzzy-semantics ablation is the paper's most compelling empirical finding: log-product semantics reliably ground concepts during training, whereas the Gödel Trick proves superior for planning, and mixing the two without retraining is elegant. The symbolic predictor also eliminates compounding prediction errors, yielding near-perfect F1 scores even on the longest imagined rollouts.
Swapping only the transition rules to achieve zero-shot planning on the periodic-door environment, while keeping the encoder frozen, is a clean demonstration of modularity.
The approach is less a learned world model than a learned visual encoder bolted to a hand-coded symbolic simulator. The predictor has no learnable parameters and instead applies exhaustive cell-level LTLf rules that must be authored correctly a priori. Scaling this knowledge engineering to richer domains is neither tested nor obviously feasible.
The planning evaluation also omits the most natural baseline: a classical planner operating over the same symbolic dynamics would solve these small grids optimally and with guarantees, whereas differentiable gradient ascent is neither complete nor optimal here. In addition, the comparison with LeWM is weakened by the admission that the authors modified the baseline in ways that "all favour the baseline," while still denying it the symbolic knowledge given to CGWM.
The internal evidence supports the restricted claims made for the chosen domains: CGWM attains near-perfect F1 and stable long-horizon rollouts, while the purely neural baseline LeWM degrades because prediction errors accumulate. However, the paper does not quantify the human effort required to author the LTLf knowledge bases, nor does it test whether the method tolerates incomplete or approximate rules.
The external citation to Badreddine et al. (2023) is accurate—we verified that the source advocates log-space product semantics to avoid underflow—but the paper's use of this prior work is incremental.
The authors share code via an anonymous repository and include implementation details in Appendix D. The main barrier to broader reproduction is knowledge engineering: applying CGWM to a new domain requires manually writing an exact, complete set of cell-level LTLf constraints and transition rules.
The paper does not include a systematic ablation of how much weak supervision is required relative to the knowledge, nor does it measure sensitivity to slightly incorrect formulas, leaving unanswered how much of the reported performance stems from the logic versus the neural encoder.
World models aim to learn compact representations of an environment from observations, enabling agents to predict future states and plan their behaviour. However, most existing approaches learn black-box representations, with limited ability to incorporate prior knowledge about the world. We introduce Concept-Grounded World Models (CGWMs), a neurosymbolic world model design that instead learns interpretable concept representations. We learn these concepts from observations using explicit and complete knowledge, expressed in temporal logic, that specifies the valid world states and the temporal relations governing their dynamics. Furthermore, we directly use this knowledge by formulating planning as differentiable reasoning, optimising the agent's actions by gradient ascent to satisfy a goal. We experimentally compare CGWM with a purely neural baseline and a neurosymbolic baseline on two game environments. We find that CGWMs ground concepts much more accurately, even over long horizons, resulting in significantly more robust planning both in and out of distribution.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.