Vernier: Probing Representational Misalignment Behind Lexical Gaps in Causal Reasoning

Zhenyu Yu

Instruction-tuned language models can answer the same causal-reasoning question differently after its English variable names are replaced by type-preserving placeholders, although…