Stay on Topic with Classifier-Free Guidance

Guillaume Sanchez, Alexander Spangher, Honglu Fan, Elad Levi, Stella Biderman
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:43197-43234, 2024.

Abstract

Classifier-Free Guidance (CFG) has recently emerged in as a lightweight technique to encourage prompt-adherence in generations, yet has not yet been successfully applied to language modeling. In this work, we demonstrate across a wide array of benchmarks that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across: Q&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in human evaluations we show a 75% preference for using CFG over baseline.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-sanchez24a, title = {Stay on Topic with Classifier-Free Guidance}, author = {Sanchez, Guillaume and Spangher, Alexander and Fan, Honglu and Levi, Elad and Biderman, Stella}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {43197--43234}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/sanchez24a/sanchez24a.pdf}, url = {https://proceedings.mlr.press/v235/sanchez24a.html}, abstract = {Classifier-Free Guidance (CFG) has recently emerged in as a lightweight technique to encourage prompt-adherence in generations, yet has not yet been successfully applied to language modeling. In this work, we demonstrate across a wide array of benchmarks that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across: Q&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in human evaluations we show a 75% preference for using CFG over baseline.} }
Endnote
%0 Conference Paper %T Stay on Topic with Classifier-Free Guidance %A Guillaume Sanchez %A Alexander Spangher %A Honglu Fan %A Elad Levi %A Stella Biderman %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-sanchez24a %I PMLR %P 43197--43234 %U https://proceedings.mlr.press/v235/sanchez24a.html %V 235 %X Classifier-Free Guidance (CFG) has recently emerged in as a lightweight technique to encourage prompt-adherence in generations, yet has not yet been successfully applied to language modeling. In this work, we demonstrate across a wide array of benchmarks that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across: Q&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in human evaluations we show a 75% preference for using CFG over baseline.
APA
Sanchez, G., Spangher, A., Fan, H., Levi, E. & Biderman, S.. (2024). Stay on Topic with Classifier-Free Guidance. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:43197-43234 Available from https://proceedings.mlr.press/v235/sanchez24a.html.

Related Material