Language models are expected to answer from a user-supplied document even when that document contradicts what the model has learned in training. The new arXiv paper "Shaping the Document, Not the Model: Optimized Suffixes for Context Following" takes on this problem. According to the abstract, prior work has focused on making the model rely on context by steering the model itself.
The paper proposes an alternative: shape the document, not the model. The idea is to append an optimized suffix to the input document so that the model follows the supplied content over its parametric knowledge. This turns context following into a text-level intervention rather than a weight-level one.
The abstract gives only a high-level description; the exact method for generating the suffix is not detailed in the announcement. Still, the framing is notable for shifting attention from model training or prompting to the structure of the provided document itself.