Shaping the Document, Not the Model: Optimized Suffixes for Context Following

Key points

• Prior work has focused on making the model rely on context by steering activations, contrasting logits during decoding, or instructing it in plain language.1
• We optimize a document suffix by discrete token search, starting from the answer to a single question and extending the objective to a set of passages, to multi-hop questions that follow from the planted fact, and to later turns of a conversation.1
• On counterfactual passages that contradict parametric knowledge of the model, the suffix flips nearly all items the model resists across three instruction-tuned model families, in most settings outperforms the decoding baseline and the strongest instruction we could write, and matches activation steering.1
• Controls show that the suffix largely acts through the passage rather than injecting the answer, and a single suffix per model transfers to unseen passages.1

Sources (1)

• [1]Shaping the Document, Not the Model: Optimized Suffixes for Context Following
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 12:01 PM

  When a user supplies a document and wants the answer drawn from it, a language model is expected to follow it even when it contradicts its parametric knowledge.

  Prior work has focused on making the model rely on context by steering activations, contrasting logits during decoding, or instructing it in plain language.

Extractive summary: sentences quoted from the sources.