Back to AI Research

AI Research

Grounding vs. Compositionality: On the Non-Compleme... | AI Research

Key Takeaways

  • Compositionality: On the Non-Complementarity of Reasoning in Neuro-Symbolic Systems This research investigates a long-standing assumption in ar...
  • Compositional generalization remains a foundational weakness of modern neural networks, limiting their robustness and applicability in domains requiring out-of-distribution reasoning.
  • A central, yet unverified, assumption in neuro-symbolic AI is that compositional reasoning will emerge as a byproduct of successful symbol grounding.
  • This work presents the first systematic empirical analysis to challenge this assumption by disentangling the contributions of grounding and reasoning.
  • To operationalize this investigation, we introduce the Iterative Logic Tensor Network ($i$LTN), a fully differentiable architecture designed for multi-step deduction.
Paper AbstractExpand

Compositional generalization remains a foundational weakness of modern neural networks, limiting their robustness and applicability in domains requiring out-of-distribution reasoning. A central, yet unverified, assumption in neuro-symbolic AI is that compositional reasoning will emerge as a byproduct of successful symbol grounding. This work presents the first systematic empirical analysis to challenge this assumption by disentangling the contributions of grounding and reasoning. To operationalize this investigation, we introduce the Iterative Logic Tensor Network ($i$LTN), a fully differentiable architecture designed for multi-step deduction. Using a formal taxonomy of generalization -- probing for novel entities, unseen relations, and complex rule compositions -- we demonstrate that a model trained solely on a grounding objective fails to generalize. In contrast, our full $i$LTN, trained jointly on perceptual grounding and multi-step reasoning, achieves high zero-shot accuracy across all tasks. Our findings provide conclusive evidence that symbol grounding, while necessary, is insufficient for generalization, establishing that reasoning is not an emergent property but a distinct capability that requires an explicit learning objective.

Grounding vs. Compositionality: On the Non-Complementarity of Reasoning in Neuro-Symbolic Systems
This research investigates a long-standing assumption in artificial intelligence: that if a model can successfully learn to map raw sensory data to symbols—a process known as "symbol grounding"—it will automatically gain the ability to reason logically and generalize to new, unseen scenarios. By testing this hypothesis through a series of controlled experiments, the authors demonstrate that grounding and reasoning are actually distinct capabilities. They conclude that while grounding is a necessary foundation, it is not enough on its own to achieve robust compositional generalization.

The Iterative Logic Tensor Network

To test their hypothesis, the authors introduced the Iterative Logic Tensor Network ($i$LTN). Unlike standard models that attempt to solve problems in a single pass, the $i$LTN is designed to perform multi-step deduction. It treats reasoning as an iterative refinement process, where the model updates its "belief state" about a puzzle over several steps. By using a differentiable framework, the model can learn to apply logical rules—such as those found in Sudoku—to refine its predictions, allowing it to handle increasingly complex chains of logic.

Testing Generalization

The researchers evaluated their model using a formal taxonomy of generalization, focusing on three specific challenges:

  • Entity Composition: Can the model handle symbols (like digits) that it did not see during training?

  • Relational Composition: Can the model adapt to new logical rules, such as adding arithmetic constraints to a standard puzzle?

  • Rule Composition: Can the model solve complex, multi-step puzzles when it was only trained on simpler, shorter reasoning chains?

Key Findings

The study reveals that models trained solely on a grounding objective—mapping images to symbols—consistently fail when faced with these compositional shifts. In contrast, the full $i$LTN model, which was trained with both a grounding objective and an explicit reasoning objective, achieved high accuracy across all tasks. The results provide strong evidence that reasoning is not an emergent property that "just happens" once a model understands symbols. Instead, it is a separate skill that must be explicitly taught through a dedicated learning objective.

Implications for AI Design

The findings suggest that the current trend of focusing heavily on perceptual grounding may be insufficient for building truly robust AI systems. For models to perform well in out-of-distribution environments—where they encounter novel combinations of familiar concepts—developers must incorporate explicit architectural support for reasoning. By disentangling these two processes, the research highlights that future neuro-symbolic systems should treat logical deduction as a primary, intentional component of the learning process rather than a secondary byproduct of perception.

Comments (0)

No comments yet

Be the first to share your thoughts!