How Language Models Process Negation
For researchers and practitioners in NLP, this work deepens mechanistic understanding of how LLMs handle negation, revealing competing internal mechanisms and a path to improve accuracy.
LLMs possess internal components that correctly process negation, but late-layer attention shortcuts cause poor accuracy; ablating these modules improves accuracy. The models use both a suppressive and a constructive mechanism for negation, with the constructive mechanism being more prominent.
We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and suppress related concepts, or they could directly construct a representation of the entire negative phrase (e.g., representing "not gas" as a vector that promotes liquids and solids). We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the "constructive" mechanism being more prominent. Combined, our work deepens the understanding of LLMs' internals, highlighting construction-dominant computations and the coexistence of competing mechanisms within LLMs.