On Limitations of the Transformer Architecture
This reveals foundational limitations of Transformers for compositional reasoning tasks, which is critical for improving reliability in AI systems.
The paper identifies root causes of hallucinations in large language models by proving that Transformer layers are fundamentally incapable of composing functions (e.g., identifying grandparents in genealogies) when domains are sufficiently large, with empirical evidence showing this limitation occurs even with small domains.
What are the root causes of hallucinations in large language models (LLMs)? We use Communication Complexity to prove that the Transformer layer is incapable of composing functions (e.g., identify a grandparent of a person in a genealogy) if the domains of the functions are large enough; we show through examples that this inability is already empirically present when the domains are quite small. We also point out that several mathematical tasks that are at the core of the so-called compositional tasks thought to be hard for LLMs are unlikely to be solvable by Transformers, for large enough instances and assuming that certain well accepted conjectures in the field of Computational Complexity are true.