LLMs Reveal Internal Conflict Resolution in System vs. User Instructions

This research delves into the internal mechanisms by which instruction-tuned Large Language Models (LLMs) handle conflicting instructions from system prompts versus user inputs. A new benchmark with paired constraints and deterministic verifiers was used to evaluate eight models, categorizing their behavior into three types: hierarchy-respecting, anti-hierarchy (preferring user instructions), or showing no channel sensitivity. Notably, Llama-3.1-8B emerged as a strong anti-hierarchy model, frequently disregarding system instructions. Despite this behavioral preference, the study revealed that even in anti-hierarchy models like Llama-3.1-8B, an internal conflict-resolution signal is highly decodable from residual-stream activations. This signal can be leveraged for steering, significantly improving system compliance. The findings suggest that user-preferring arbitration does not imply an absence of internal conflict awareness, but rather a specific internal weighting, and that successful intervention depends on understanding the geometry of these internal representations.