Back to feed
arXiv cs.CL
arXiv cs.CL
6/23/2026
Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures

Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures

Short summary

Researchers quantified how individual adjectives steer LLM outputs across five models using Shapley values on MMLU. They found a 'family effect' where model lineage determines sensitivity patterns, and that adjectives' effects vary by syntactic position and scale non-additively in larger models. These findings challenge universal prompting strategies and highlight the need for model-specific steering techniques.

  • •Shapley value analysis reveals certain adjectives act as powerful steering 'levers' with model-specific effects
  • •Models from the same family (e.g., OpenAI GPT-4o variants) show correlated sensitivities, while different architectures respond unpredictably
  • •Larger models exhibit complex non-additive interactions between adjectives; smaller models respond more literally

Generated with AI, which can make mistakes.

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

Is this a good recommendation for you?

Explore more

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
arXiv cs.CL
Assert, don't describe: Linguistic features that shift LLM reasoning about animal welfare
arXiv cs.CL
Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior
arXiv cs.CL
Adaptive scheduling steers diffusion L
arXiv cs.LG