Understanding the Impact of Linguistic Realization Choices on LLM Stance with Causal Tracing
For NLP researchers studying LLM sensitivity, this work provides a causal analysis of constructional effects on model decisions, though the findings are incremental and domain-specific to political stance.
This paper investigates how linguistic construction choices (beyond lexical changes) shift LLM stance decisions, finding that both meaning-preserving and meaning-inversing rewrites cause stance instability. Using causal tracing, they localize the effect to mid-to-late decoder layers at the final prompt position.
Large language models (LLMs) are known to be sensitive to prompt and input formulations. However, existing studies have focused on lexical realization and largely ignored constructional choice. This paper studies whether linguistic construction can systematically shift LLM decisions and where these shifts can be causally localized inside the model. We use political stance judgment as a meaning-sensitive case study and extend an English political statements dataset, resulting in six controlled linguistic rewrite types that preserve or invert the meaning of a statement. Experiments on four open-weight models show that stance instability affect both meaning-preserving and meaning-inversing rewrites. Because output shifts reveal that rewrites affect stance, but not where in the model, we apply activation patching, where activations from the original statement are substituted into the forward pass for the rewritten statement and measure which components recover the original stance distribution. The results show that mid-to-late decoder layers, especially block outputs at the final prompt position, provide the strongest restoration signal.