Mechanisms and failure conditions of emotionally framed prompting and activation steering

Determine why emotionally framed prompting and activation steering influence large language model behavior and ascertain the conditions under which these methods fail, providing a concrete account of the mechanisms and failure modes governing such interventions.

Background

The paper opens by noting that prior work shows emotionally framed prompting and activation steering can influence LLM behavior, but the underlying reasons and boundary conditions remain unclear. The authors propose a representation-level account based on a learned valence–arousal (VA) subspace and present evidence for a mechanism they call lexical mediation, where VA steering modulates the likelihood of refusal- or compliance-associated tokens.

Although the work offers a plausible explanation and demonstrates behavioral control across refusal and sycophancy, the authors frame the broader question of why such methods work and when they fail as an unresolved issue motivating their study.