Why do language models ignore information in their context?

The REMEDI paper names a specific failure mode: "failure of context integration." The example: an LM is prompted with a context establishing that Anita works in a law office, but when generating a continuation, the LM describes Anita as a nurse — overriding the contextual information with a prior association (names like Anita may statistically co-occur with certain occupations in training data).

This is a named, empirically documented failure mode, not a hypothetical. The failure occurs because the LM's parametric knowledge (compressed into weights from training) and its in-context information (the prompt) are not cleanly integrated. When they conflict, the parametric association can win.

The implication is important for how we think about context windows and RAG-style augmentation. Just providing information in context does not guarantee that a model will use it. If the information conflicts with strong prior associations, the prior may dominate — not because the model misread the context, but because context integration is not a lossless operation. The provided information gets processed through the same mechanisms that already have strong priors.

Fixing this requires causal intervention, not just better prompting: you need to modify the representations that carry the prior association, not just add more context on top of them. This is what REMEDI demonstrates — that adding a learned vector directly to entity representations can override the prior in a way that textual prompting cannot.