Shared Lexical Task Representations Explain Behavioral Variability In LLMs

Researchers show that LLMs exploit shared lexical task representations to explain behavioral variability across different language tasks—revealing how a single model produces inconsistent outputs without explicit task-specific tuning.