Back to feed
arXiv cs.CL
arXiv cs.CL
6/26/2026
Assert, don't describe: Linguistic features that shift LLM reasoning about animal welfare

Assert, don't describe: Linguistic features that shift LLM reasoning about animal welfare

Short summary

Research on Llama-3.2-1B shows that assertive language, explicit moral vocabulary, emotion words, and evaluative claims significantly shift LLMs toward pro-animal-welfare reasoning during training, while hedged language and neutral sensory description dilute this effect. For writers: assert positions rather than describe scenes neutrally to influence model behavior.

  • •Eight of ten linguistic features produced statistically significant shifts in model pro-animal-welfare reasoning
  • •Seven features strengthen pro-welfare stance: assertive certainty, moral vocabulary, emotion, evaluative claims, narrative, depicted harm, and temporal framing
  • •Two features dilute it: hedged language and concrete sensory description—explicit position matters more than neutral description

Generated with AI, which can make mistakes.

#research-breakthrough#ai-agents
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Explore more

Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training
arXiv cs.CL
Small edits, large models: How Wikipedia advocacy shapes LLM values
arXiv cs.CL
Reducing Conversational Escalation in Large Language Model Dialogue with Nonviolent Communication Constraints
arXiv cs.CL
Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures
arXiv cs.CL