Tag
This paper presents a cross-model observational study on fixed-point structures to arbitrate between data-side and weights-side explanations of neural text degeneration, finding that the structural class is not determined by training data and varies across model architectures.
The paper investigates using cross-model disagreement in LLMs to target expert effort in revising codebooks for large-scale text annotation, showing that rationale labeling outperforms other feedback methods.