Names Don't Matter: Symbol-Invariant Transformer for Open-Vocabulary Learning

TL;DR

This paper introduces a symbol-invariant Transformer mechanism that effectively handles interchangeable tokens, enabling better generalization to unseen symbols in open-vocabulary tasks through parallel embedding streams and structured attention.

Contribution

It presents a novel Transformer architecture that is provably invariant to token renaming, improving open-vocabulary learning and generalization to new symbols.

Findings

The method achieves theoretical invariance guarantees.

Substantial performance improvements on open-vocabulary tasks.

Effective generalization to unseen symbols.

Abstract

Current neural architectures lack a principled way to handle interchangeable tokens, i.e., symbols that are semantically equivalent yet distinguishable, such as bound variables. As a result, models trained on fixed vocabularies often struggle to generalize to unseen symbols, even when the underlying semantics remain unchanged. We propose a novel Transformer-based mechanism that is provably invariant to the renaming of interchangeable tokens. Our approach employs parallel embedding streams to isolate the contribution of each interchangeable token in the input, combined with an aggregated attention mechanism that enables structured information sharing across streams. Experimental results confirm the theoretical guarantees of our method and demonstrate substantial performance gains on open-vocabulary tasks that require generalization to novel symbols.