Language models silently pick a side in conflicting answers
Hidden Commitment: When Language Models Silently Pick a Side and How Steering Can Surface It
Computer Vision and Pattern Recognition
Summary
Sometimes when language models answer questions using two conflicting documents, they choose one side without saying so aloud. The authors studied this silent choice in models and found patterns inside the model’s layers that reveal which document it preferred, even when asked to show the conflict. They also developed ways to influence the model to reveal or change that hidden choice. This helps in making AI answers more transparent when sources disagree.
What this means in practice
- •For ai system builders: Detect silent document preference in language model answers to improve transparency in retrieval-based systems.
- •For customer support teams: Surface hidden conflicts in AI-generated answers when multiple sources disagree to provide clearer explanations to users.
Authors
Samuel Dawit Assefa, Jae Won Cho
Abstract
Language-model answers are often treated as a visible record of the internal decision that produced them. We test this assumption in conflicting-context question answering, where two supplied documents support incompatible answers. A model can process the disagreement and still return a one-sided answer, even under prompts that ask it to disclose conflict. In Qwen 0.8B, hidden source commitment is captured by a residual source-choice direction whose sign tracks which document the answer uses. Sparse middle-layer MLP writers feed the same coordinate, and steering the coordinate changes the chosen source. A transfer run on Gemma 3 1B preserves the source-choice readout and sparse-write bridge under LLM-audited labels. Conflict disclosure follows a different pattern, appearing weaker, later-layer, and only partially steerable. These results support a practical monitoring strategy for retrieval systems that detects silent source commitment with residual readouts and pairs conflict-surfacing prompts with disclosure steering when the model has selected a side under disagreement.