Example · part of Thinking inside frameworks

Take a test question for an AI language model and change only its layout: a different separator, different spacing, a capital letter. The meaning stays the same. Sclar, Choi, Tsvetkov and Suhr (2024) did this across more than 50 tasks with several widely used open-source models. For one model, LLaMA-2-13B, accuracy differed by up to 76 points between formats. The median spread was 7.5 points.

How well a format worked on one model only weakly predicted how it worked on another, which undermines comparisons made with a single, arbitrarily chosen format. The sensitivity also persisted with larger models, more examples in the prompt, and instruction tuning.

A prompt format is a default nobody thinks of as a choice, and this is framework pricing measured in a machine. The book calls steering a model through its prompts and settings, with the model itself unchanged, deployment capture, one of the many design choices an AI owner holds at once. The authors’ advice is a form of framework rotation: report a range across plausible formats.

The tests were few-shot tasks on open-source models, and 76 points is the largest gap found.

How solid is this? A direct measurement across 50+ tasks and several open-source models in few-shot settings; the 76-point figure is the maximum for one model, against a median spread of 7.5 points.

Read the full argument

How it connects


← Inductive bias · Thinking inside frameworks (21 of 21)

Explore all the ideas · Read the book in order · Open the interactive map