Reproduce “Stay on topic with Classifier-Free Guidance” 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

Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, Stella Biderman

A practical solution adapts Classifier-Free Guidance (CFG) from text-to-image generation to autoregressive language models, improving prompt adherence and generation quality through an inference-time modification without requiring model retraining. This enables smaller models to achieve performance comparable to models with twice their parameter count across various benchmarks.

Key insights

  • CFG significantly reduces the entropy of logit distributions during generation, leading to more focused and consistent token selection.
  • The technique operates distinctly from instruction tuning, having largely non-overlapping vocabulary distribution patterns, yet both reduce entropy.
  • CFG enables smaller language models to achieve performance comparable to models with twice their parameter count, democratizing advanced LLM capabilities despite doubling inference FLOPs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.