LLMs Learn Constructions That Humans Do Not Know
This reveals a potential flaw in construction probing methods for LLMs, which could lead to incorrect syntactic knowledge being validated, impacting linguistic research and model evaluation.
The paper investigates false positive constructions in LLMs, where models hallucinate grammatical structures not recognized by humans, and finds that probing methods would confirm these false hypotheses with high accuracy, highlighting a confirmation bias.
This paper investigates false positive constructions: grammatical structures which an LLM hallucinates as distinct constructions but which human introspection does not support. Both a behavioural probing task using contextual embeddings and a meta-linguistic probing task using prompts are included, allowing us to distinguish between implicit and explicit linguistic knowledge. Both methods reveal that models do indeed hallucinate constructions. We then simulate hypothesis testing to determine what would have happened if a linguist had falsely hypothesized that these hallucinated constructions do exist. The high accuracy obtained shows that such false hypotheses would have been overwhelmingly confirmed. This suggests that construction probing methods suffer from a confirmation bias and raises the issue of what unknown and incorrect syntactic knowledge these models also possess.