A new arXiv paper takes on a core question for large language models: when internal parametric knowledge and external contextual knowledge disagree, which one should the model trust? The authors note that such conflicts are frequent, yet the mechanisms governing the model's choice remain unclear.

The abstract states that the work uncovers a generalizable steering direction across knowledge conflicts, suggesting a way to influence which knowledge source the model favors. However, the provided text cuts off after that claim, so the concrete method, experimental results, and scope of the steering direction are not available from this source alone.

Readers should treat the headline result as preliminary and look to the full paper for details on how the steering direction is derived and evaluated.