ABSTRACT
This paper proposes SymbDisentangle, a novel framework that disentangles perception from reasoning through learned symbolic representations. It introduces a differentiable perception-to-symbol interface and a compositional reasoning module, achieving state-of-the-art performance on visual reasoning benchmarks with strong systematic generalization.
PAPER · PDF
Loading PDF...
Key findings
SymbDisentangle achieves state-of-the-art performance on compositional visual reasoning benchmarks.
The framework exhibits strong systematic generalization to novel compositional concepts.
Explicit disentanglement improves data efficiency, robustness, and interpretability compared to end-to-end neural approaches.
Limitations & open questions
Further research is needed to scale the approach to more complex real-world scenarios.