Reproduce “SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Song Duong, Florian Le Bronnec, Alexandre Allauzen, Vincent Guigue, Alberto Lumbreras, Laure Soulier, Patrick Gallinari
SCOPE, a self-supervised framework from Sorbonne Université, Université Paris-Dauphine, AgroParisTech, and Criteo AI Lab, enhances faithfulness in conditional text generation by using a two-stage fine-tuning approach that automatically generates unfaithful samples. The framework achieves statistically significant improvements in faithfulness metrics, with gains up to 8.8 points in AlignScore and 8.2 points in PARENT Recall, demonstrating robust performance across diverse tasks and modern LLMs.
Key insights
- Self-supervised generation of 'challenging' unfaithful samples, by strategically blending context-aware and context-free model outputs, effectively provides a strong learning signal for faithfulness without human annotation.
- The 'noise level' (α) used in generating unfaithful samples is critical; an optimal range (e.g., [0.4, 0.6]) is necessary to create samples that are neither too similar nor too easily distinguishable from faithful outputs.
- Preference-based training can be successfully adapted to improve context faithfulness in modern LLMs by leveraging automatically constructed negative examples, thus offering a scalable alignment method.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.