Adjectives steer LLMs unpredictably across model families, new study finds

A new paper by Lars Malmqvist, accepted for TMLR, tackles the challenge of predictable LLM steering by rigorously quantifying the effect of individual adjectives. Using Shapley values for attribution, the study tested 100 adjectives on five diverse models—o3, GPT-4o-mini, Phi-3, Llama 3 70B, and DeepSeek-R1—using the MMLU benchmark. Two key findings emerge: a small subset of adjectives act as disproportionately powerful 'levers,' but their influence is not universal. A strong 'family effect' means models from the same architecture family (e.g., Llama-derived) show correlated sensitivity, while those from different families are largely uncorrelated. This challenges the assumption of a one-size-fits-all prompting strategy.

Further analysis reveals that the steering direction of powerful adjectives is highly contingent on syntactic role and position within the prompt. In larger models like GPT-4o-mini, the study provides the first quantitative evidence of strong, non-additive interaction effects: adjectives can synergistically amplify, antagonistically dampen, or even reverse each other's impact. In contrast, smaller models like Phi-3 exhibit more literal and less compositional responses. These results suggest that as models scale, their interpretation of prompts becomes more sophisticated but also less predictable, posing significant challenges for robust AI alignment and underscoring the need for model-specific, compositional steering techniques.