A sentence has structure before anyone calculates with it. In “the clever doctor helped a lawyer,” the adjective modifies the doctor. Move the adjective, and the sentence becomes a different claim about the lawyer. People can describe this relationship using the explicit language of grammar: roles, words and rules.[1]
A Yale-led preprint proposes a mathematical bridge between that explicit structure and the numerical vectors manipulated by neural networks. R. Thomas McCoy and collaborators report that role–filler representations can approximate the internal states of tested networks while preserving much of their behaviour. The preprint was submitted on 30 August 2026 and explained in a Yale News account on 8 October.[1][2]
Separating a value from the role it plays
Imagine separating what a symbol is from the job it does. A word such as “doctor” is a filler; being the subject of a clause is a role. A number such as five is also a filler, while being the numerator of a fraction is a role. In tensor-product representations, one can mathematically bind fillers to roles, producing a distributed representation of the structure.[1]
The researchers examined small networks trained on list-manipulation tasks and seven large language models across language, arithmetic, logic and coding. According to the paper and Yale's account, substituting role–filler approximations for the representation-generating process left much of the tested behaviour intact. This asks a more specific question than whether a model can produce a correct answer: can a structured mathematical account reproduce how it represents the task?[1][2]
The distinction at the centre of the research
| Representation | Basic form | Role in the argument |
|---|---|---|
| Explicit symbols | Named elements plus grammatical or logical roles | Human-readable account of a structureSource 1 |
| Neural vectors | Many continuous numerical values | How network layers encode and transform informationSource 1 |
| Tensor-product role–filler representation | A value bound to a structured position | Proposed mathematical bridge between the twoSource 1 |
| Targeted internal intervention | Change a particular role–filler relation | Test whether the predicted relation influences behaviourSource 1 |
That is stronger than noticing that a model can answer a grammar quiz. Good quiz performance might come from correlations in training data. An internal representation analysis asks whether the hypothesised structure has a measurable causal role in the network's operation. The team examined this more directly through interventions: changing a vector component associated with a grammatical relation and observing the resulting change in the model's response.[1]
A specific intervention is more informative than an analogy
Yale describes an intervention in the sentence “the clever doctor helped a lawyer.” The researchers changed the vector component linking “clever” to the subject so that the model behaved as if the adjective modified the object: “the doctor helped a clever lawyer.” This supports a causal role for the identified representation in that task, rather than merely a correlation between the representation and an output.[1][2]
The result remains a preprint finding across the tested networks and tasks. A useful symbolic approximation of those computations does not establish that all model reasoning follows the same structure, or that the model reasons in the same way as a person.[1]
Why a bridge from symbols to vectors would matter
Representations that can be identified and edited give interpretability researchers a way to test explicit hypotheses. They can ask which information is bound to a grammatical or logical role, change that relationship and examine whether the answer changes as predicted.[1][2]
The adjective intervention illustrates the value of that approach. The proposed structure did more than describe a sentence after the model processed it: changing the identified relationship changed the model's response in the expected direction. That connects the mathematical explanation to behaviour inside the tested network.[1][2]
Sources & further reading
- The Emergent Symbolic Structure of Artificial Neural Networks arXiv research preprint · 30 Aug 2026 · Accessed 27 Oct 2026
- How do AI chatbots reason like humans? A Yale-led study offers clues Yale News · 8 Oct 2026 · Accessed 27 Oct 2026