How language models handle conflicting knowledge stored internally

When language models store information during training, they sometimes retain contradictory facts about the same subject—a phenomenon called intra-memory knowledge conflict. A new study takes the first mechanistic look at how these models manage such internal inconsistencies, revealing surprising gaps between how they handle artificial versus real-world conflicts.

The Problem Within#

Most prior research on knowledge conflicts in language models has focused on mismatches between what a model knows internally and information from external sources—a problem researchers typically address through fine-tuning or editing techniques. By contrast, conflicts that arise entirely within a model's learned parameters have received little scrutiny. This gap prompted researchers to design a framework for identifying exactly where these internal conflicts live and how models resolve them.

Testing across models and scenarios#

The team tested their approach on four language models using both synthetic (artificially constructed) and real-world knowledge conflicts. Their findings showed a consistent pattern: models tend to identify and resolve conflicting information in their final layers, regardless of model architecture. However, when researchers attempted targeted interventions, they found the method worked far better on synthetic conflicts than on conflicts drawn from real data—a significant discrepancy that suggests real-world contradiction is more deeply embedded.

Head-level precision beats broader interventions#

Intervening at individual attention heads proved more effective than applying layer-wide interventions. An analysis of which attention heads specialized in particular facts revealed that synthetic conflicts tend to cluster their competing information in heads dedicated to one fact or another—a structural feature that real-world conflicts largely lack. This difference may explain why interventions struggle when moving from controlled to genuine knowledge contradictions.

No single resolution circuit#

Rather than finding evidence of a universal neural mechanism for resolving all knowledge conflicts, the results suggest that competing pieces of information are handled by separate, distinct circuits that can operate independently. This distributed approach may be why conflicts arise in the first place and why they're harder to resolve than models' ability to retrieve single facts.