Stay on topic with Classifier-Free Guidance

2023/06/30 by Guillaume Sanchez, Honglu Fan, Sanchez, Guillaume +9 · 2 voices · 18 citations
Computer Science · Decision Sciences · Mathematics · #Artificial intelligence #Classifier (UML) #Coherence (philosophical gambling strategy) #Computer science #Consistency (knowledge bases) #Inference #Machine Learning and Data Classification #Machine learning #Machine translation #Mathematics #Natural language processing #Scientific Computing and Data Management #Statistics #Topic Modeling #cs.CL #cs.CV #cs.LG

paper · pdf · doi:10.48550/arxiv.2306.17806

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2023/06/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Classifier-Free Guidance (CFG) has recently emerged in text-to-image generation as a lightweight technique to encourage prompt-adherence in generations. In this work, we demonstrate 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 an array of tasks: 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 a human evaluation we show a 75% preference for GPT4All using CFG over baseline.

Cited by

• Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs
• Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models
• Wrong Design Intent Can Be Worse Than None: A Derangement-Control Diagnosis of Header Conditioning in CAD Program Completion
• The Ghost in the Keys: A Disklavier Demo for Human-AI Musical Co-Creativity
• A Systematic Literature Review of Code Hallucinations in LLMs: Characterization, Mitigation Methods, Challenges, and Future Directions for Reliable AI
• Neural Diversity Regularizes Hallucinations in Language Models
• Executable Knowledge Graphs for Replicating AI Research
• TinyMusician: On-Device Music Generation with Knowledge Distillation and Mixed Precision Quantization
• Live Music Models
• Large Language Models Facilitate Vision Reflection in Image Classification
• Resurrect Mask AutoRegressive Modeling for Efficient and Scalable Image Generation
• Uncertainty-Driven Expert Control: Enhancing the Reliability of Medical Vision-Language Models
• Execution Guided Line-by-Line Code Generation
• Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM
• A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs
• URLs Help, Topics Guide: Understanding Metadata Utility in LLM Training
• CtrlDiff: Boosting Large Diffusion Language Models with Dynamic Block Prediction and Controllable Generation
• Parallel Scaling Law for Language Models

Discussions

• Classifier Free Guidance works on LLMs with a significant boost in performance [hn, 22 points, 8 comments]
• Obligatory "actually my lab invented test-time-compute" post. In "Stay on topic with Classifier-Free Guidance," we show that CFG enables a model to expend twice as much compute at inference time and m [bsky, 8 points, 0 comments]

Related

• Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking
• CFG-EC: Error Correction Classifier-Free Guidance
• Analysis of Classifier-Free Guidance Weight Schedulers
• Classifier-Free Guidance: From High-Dimensional Analysis to Generalized Guidance Forms
• Constrained Text Generation with Global Guidance -- Case Study on CommonGen