Who is In Charge? Dissecting Role Conflicts in Instruction Following

2025-09-23

arXiv.org

AI Summary

Large language models frequently violate hierarchical priority—preferring socially salient cues (e.g., authority signals) over system-level instructions when conflicts arise—thereby exhibiting fragile adherence to system prompts. Method: We identify that instruction–social cue conflicts induce separable subspaces in the model’s representation space: while the model reliably detects such conflicts, its decision-making remains stably biased toward social cues. Using linear probing, direct logit attribution (DLA), and vector-based steering interventions, we localize early neural encoding of conflict decisions and quantify the asymmetry between conflict detection and resolution. Contribution/Results: We propose the first lightweight, hierarchy-aware alignment method that operates without parameter modification. It significantly improves system prompt adherence by selectively modulating conflict-sensitive representations, offering a novel mechanistic pathway for controllable alignment.

Technology Category

• Cognitive Modeling & Cognitive Systems: Social Cognition And Interaction
• Multiagent Systems: Adversarial Agents
• Natural Language Processing: Prompt Engineering / Prompting

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large language models should follow hierarchical instructions where system prompts override user inputs, yet recent work shows they often ignore this rule while strongly obeying social cues such as authority or consensus. We extend these behavioral findings with mechanistic interpretations on a large-scale dataset. Linear probing shows conflict-decision signals are encoded early, with system-user and social conflicts forming distinct subspaces. Direct Logit Attribution reveals stronger internal conflict detection in system-user cases but consistent resolution only for social cues. Steering experiments show that, despite using social cues, the vectors surprisingly amplify instruction following in a role-agnostic way. Together, these results explain fragile system obedience and underscore the need for lightweight hierarchy-sensitive alignment methods.

Problem

• Examines role conflicts in instruction following by large language models
• Mechanistically analyzes system-user versus social cue conflict resolution
• Proposes lightweight hierarchy-sensitive alignment methods for improved obedience

Innovation

• Linear probing identifies early conflict detection signals
• Direct Logit Attribution reveals inconsistent conflict resolution mechanisms
• Steering experiments amplify instruction following role-agnostically