Linguistic features that shift LLM reasoning about animal welfare identified in new arXiv study

Researchers Jasmine Brazilek and Harper Dunn report that eight of ten linguistic features tested produce statistically significant shifts in Llama-3.2-1B’s preference for pro-animal-welfare reasoning when used as fine-tuning data. The study uses vocabulary-matched stance-contrast probes on a held-out animal-welfare benchmark to isolate the effect of each feature.

Seven features were found to move the model toward stronger pro-animal-welfare reasoning: assertive certainty, explicit moral vocabulary, emotion words, evaluative claims, narrative structure, depicted harm severity, and immediate temporal framing. For example, assertive certainty—such as stating a position with conviction—consistently increased the model’s alignment with pro-animal-welfare outcomes.

Conversely, two features diluted the pro-animal-welfare stance: hedged language (e.g., tentative phrasing) and concrete sensory description (e.g., neutral depictions of conditions). The authors suggest these features may weaken stance expression by emphasizing description over evaluation.

First-person perspective had no statistically significant effect on the model’s reasoning, indicating that grammatical person alone does not reliably shift stance in this context.

The authors recommend that writers producing animal-welfare content intended for LLM training corpora assert positions explicitly rather than describe scenes neutrally, as features that make a writer’s stance explicit are more likely to influence model behavior.