Papers › Stay on topic with Classifier-Free Guidance

Stay on topic with Classifier-Free Guidance

30 Jun 2023arXiv:2306.17806archive 2025-07-28

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

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.

PaperPDF

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

No code repository is listed for this paper in the archive or in Syntology's graph.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Code GenerationCommon Sense ReasoningImage GenerationLAMBADALanguage ModelingLanguage ModellingMachine TranslationSentence CompletionText GenerationText to Image GenerationText-to-Image GenerationZero-Shot Learning

Datasets

Introduced by this paper, per the archive.

SUDOER

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Common Sense Reasoning ARC (Easy) LLaMA 65B + CFG (0-shot) Accuracy 84.2 #8 of 47 Archive leaderboard report
Common Sense Reasoning ARC (Easy) LLaMA 30B + CFG (0-shot) Accuracy 83.2 #11 of 47 Archive leaderboard report
Common Sense Reasoning ARC (Easy) LLaMA 13B + CFG (0-shot) Accuracy 79.1 #18 of 47 Archive leaderboard report
Common Sense Reasoning ARC (Easy) LLaMA 7B + CFG (0-shot) Accuracy 58.9 #42 of 47 Archive leaderboard report
Language Modelling LAMBADA LLaMA-65B+CFG (Zero-Shot) Accuracy 84.0 #4 of 37 Archive leaderboard report
Language Modelling LAMBADA LLaMA-30B+CFG (zero-shot) Accuracy 83.9 #5 of 37 Archive leaderboard report
Language Modelling LAMBADA LLaMA-13B+CFG (zero-shot) Accuracy 82.2 #8 of 37 Archive leaderboard report
Sentence Completion HellaSwag LLaMA 65B + CFG (0-shot) Accuracy 86.3 #20 of 89 Archive leaderboard report
Sentence Completion HellaSwag LLaMA 30B + CFG (0-shot) Accuracy 85.3 #25 of 89 Archive leaderboard report
Sentence Completion HellaSwag LLaMA 13B + CFG (0-shot) Accuracy 82.1 #39 of 89 Archive leaderboard report
Text Generation SciQ LLaMA-65B+CFG (zero-shot) Accuracy 96.6 #1 of 3 Archive leaderboard report
Text Generation SciQ LLaMA-30B+CFG (zero-shot) Accuracy 96.4 #2 of 3 Archive leaderboard report
Text Generation SciQ LLaMA-13B+CFG (zero-shot) Accuracy 95.1 #3 of 3 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

AdamAttentionAttention DropoutBPECosine AnnealingDense ConnectionsDiscriminative Fine-TuningDropoutGPT-2Layer NormalizationLinear LayerLinear Warmup With Cosine AnnealingMulti-Head AttentionPythiaResidual ConnectionSoftmaxWeight Decay

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections