Characterizing additional non-lexical pathways by which VA steering modulates behavior

Characterize the additional mechanisms, beyond lexical mediation, through which steering along valence and arousal subspace affects large language model outputs; in particular, identify whether and how higher-level planning processes and attention patterns mediate VA-induced changes in token probabilities and downstream behaviors.

Background

The authors present evidence that lexical mediation—changes in the probabilities of refusal- or compliance-associated tokens—constitutes a mechanism linking VA steering to behavior. They also observe distributed MLP neuron effects and suggest that additional mechanisms may contribute.

They note that VA steering could affect higher-level planning or attention patterns, which were not measured in this work, and explicitly flag the task of identifying these pathways as an open problem, suggesting future causal-tracing or circuit-level analyses to locate specific attention heads and MLP sublayers.