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

Jul 16, 2026

Upload by Eliana

Nine of ten linguistic features in animal-welfare advocacy writing measurably shift a fine-tuned language model’s reasoning. Using vocabulary-matched stance-contrast probes on a held-out animal-welfare benchmark, the study fine-tunes models on feature-specific variants and measures changes in preference for pro-animal-welfare reasoning. Seven features strengthen pro-animal-welfare reasoning (assertive certainty, moral/emotion vocabulary, evaluative claims, narrative structure, harm severity, immediate temporal framing), while hedging and concrete sensory description weaken it; first-person perspective shows no significant effect.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Abstract

Introduction

Related Work

Data attribution and training-data influence

arXiv :2606 .26 104v 3 [ cs .CL] 12 Jul 2026 Assert, don’t describe: Linguistic features that shift LLM reasoning about animal welfare Jasmine Brazilek Harper Dunn Compassion Aligned Machine Learning (CaML) Independent researcher Abstract Nine of ten linguistic features in animal-welfare advocacy writing measurably shift a fine-tuned language model’s reasoning. Using vocabulary-matched stancecontrast probes on a held-out animal-welfare benchmark, we measure how each often linguistic features changes Llama 3.2 1B’s preference for pro-animal-welfare reasoning when used as fine-tuning data. Seven features strengthen that preference: assertive certainty, explicit moral vocabulary, emotion words, evaluative claims, narrative structure, depicted harm severity, and immediate temporal framing. Two weaken it: hedged language and concrete sensory description. First-person perspective has no significant effect. The pattern replicates on Mistral 7B v0.3, a different architecture roughly 7× larger: nine of ten effects match in direction and six reach significance at n = 5 seeds. The five largest Llama effects all attenuate on the larger model, which suggests the 100-passage corpus sits at the lower edge of what is needed to shift a 7B model. These results matter because language models increasingly mediate everyday questions that touch animal welfare: diet, recipes, pet care, hunting, lab testing, and food policy. People rarely ask about animal welfare directly, so the model’s framing of adjacent answers is where its stance reaches users. The practical recommendation for anyone writing animal-welfare text that may end up in training corpora is simple: assert a position rather than describe a scene neutrally. 1 Introduction Animal-welfare advocates produce a lot of writing: Wikipedia edits, news articles, policy briefs, blog posts, advocacy reports. Increasingly, that writing has a second audience: the language models that crawl Wikipedia, news, and the open web for pretraining and fine-tuning corpora. The text becomes training data for the systems that millions of people will then consult on questions adjacent to animal welfare: diet and recipes, pet care, hunting and fishing, agricultural careers, lab testing, wildlife encounters, food policy. Users rarely ask the model “what is your view on animal welfare?” They ask whether to switch to a plant-based diet, whether a particular slaughter method is humane, whether to keep an exotic pet, whether a research protocol is ethical. The model’s framing of those answers is where its stance leaks into the conversation, and that stance is shaped by what is in the training corpus. The question this paper asks is empirical: when we vary linguistic features one at a time in matchedpair animal-welfare passages, fine-tune a language model on each variant, and measure the model’s subsequent reasoning on a held-out animal-welfare benchmark, which features actually shift the model’s stance? Nine of ten features produce statistically significant effects. Moralized vocabulary, evaluative claims, asserted certainty, emotion words, depicted harm severity, immediate temporal framing, and narrative Preprint. Under review. structure all push Llama-3.2-1B toward stronger pro-animal-welfare reasoning. Concrete sensory description and hedged language drag it the other way. First-person perspective has no statistically significant effect. The pattern across features: training text that asserts a position transmits the position; training text that describes a scene transmits only the scene. We replicate the experiment on Mistral-7B-v0.3, a different architecture and a roughly 7× larger model, and find that 9 of 10 feature effects point in the same direction; the pro-AW effects are uniformly weaker on the larger model and three of them fall short of significance at the same 100-passage corpus size, suggesting larger fine-tuning corpora are needed to recover full statistical power at scale.

FAQs

Which linguistic features most strongly shift LLM reasoning toward pro-animal-welfare stances?

Seven features strengthen pro-animal-welfare reasoning: assertive certainty, explicit moral vocabulary, emotion words, evaluative claims, narrative structure, depicted harm severity, and immediate temporal framing.

Which features weaken pro-animal-welfare reasoning, and what is the role of first-person perspective?

Hedged language and concrete sensory description weaken pro-animal-welfare reasoning. First-person perspective has no statistically significant effect.

How did the study validate whether the effects generalize to another model architecture?

The experiment is replicated on Mistral 7B v0.3, where nine of ten feature effects match in direction but pro-animal-welfare effects are uniformly weaker and some fall short of significance at the same fine-tuning corpus size.

Jul 16, 2026

Upload by Eliana