Neural Nets May Be Hiding Symbolic Logic Under Their Vectors

A new arXiv paper tackles a long-standing puzzle in AI: why do neural networks, which encode everything as continuous vectors, succeed at tasks like language, logic, and arithmetic that were traditionally thought to require structured symbol manipulation? The authors argue the two views aren’t in conflict. Their claim is that a network’s internal vector representations implicitly encode symbolic structure — the symbols are there, just embedded in continuous space rather than stated explicitly.

To test this, they replace a network’s entire representation-generating machinery with a closed-form equation that instantiates an explicit symbolic structure, and find the model’s behavior stays largely intact. The result holds across scales, from small networks trained to manipulate lists up to large language models working in four domains central to symbolic AI: arithmetic, logic, code, and natural language. Crucially, the symbolic approximation is not just descriptive — the researchers use it to steer an LLM’s outputs through targeted edits to its internal representations, evidence that the identified structures actually drive behavior rather than merely correlating with it.

If the findings generalize, this offers a bridge between the decades-old symbolic tradition of intelligence and the vector-based reality of modern deep learning. Practically, symbolic approximations that permit precise interventions point toward more interpretable and controllable models — a meaningful direction for anyone concerned with auditing, debugging, or constraining LLM behavior.