Research map / Large language models / Language, safety and interpretability
FACE: Evaluating Natural Language Generation with Fourier Analysis of Cross-Entropy
Zuhao Yang, Yingfang Yuan, Yang Xu, SHUO ZHAN, Huajun Bai, Kefan Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use contrastive decoding instead of nucleus sampling | GPT2 conditional generation · Conditional generation | FACE-SO: 0.42 -> 0.44 | Improved |
| Use contrastive decoding instead of greedy decoding | GPT2 conditional generation · Conditional generation | FACE-CORR: 0.56 -> 0.75 | Improved |
| Use contrastive decoding instead of nucleus sampling | GPT2 conditional generation · Conditional generation | Perplexity: 13.13 -> 14.39 | Mixed |
| Add temperature to top-k sampling | GPT2 conditional generation · Conditional generation | FACE-SO: 0.40 -> 0.34 | Hurt |
| Use FACE-SO instead of MAUVE for human-like correlation | FACE-SO · Human judgments | Spearman correlation: 0.952 -> 0.881 | Hurt |
| Restrict to human-vs-model pairs: FACE-SO instead of MAUVE | FACE-SO · Human-like judgments | Spearman correlation: 0.214 -> 0.357 | Improved |
| Use human-human folds vs human-model fold | FACE-SO · Sanity test | FACE-SO: 0.44 -> 0.45 | Improved |
| Use human-model fold instead of human-human fold | FACE-SPEAR · Sanity test | FACE-SPEAR: 0.05 -> 0.17 | Improved |
| Increase GPT2 size from sm to xl | GPT2 · Wiki text | MAUVE: 0.677 -> 0.186 | Hurt |
| Use nucleus sampling instead of ancestral sampling | GPT2-xl · MAUVE original data | FACE-SO: 0.472 -> 0.481 | Improved |
| Change estimator model m_est among GPT2-sm to xl (ablation) | FACE · webtext and GPT2 outputs | spectra shape | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MAUVE — MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers (NeurIPS 2021) · better on 0 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Language Model Decoding as Direct Metrics Optimization (ICLR 2024 · similarity 0.86)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · similarity 0.85)
- On the Entropy Calibration of Language Models (NeurIPS 2025 · similarity 0.85)
- Improving Open-Ended Text Generation via Adaptive Decoding (ICML 2024 · similarity 0.85)
- Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs (ICLR 2025 · similarity 0.84)
- Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
