Research map / Large language models / Language, safety and interpretability
Calibrating Sequence likelihood Improves Conditional Language Generation
Yao Zhao, Mikhail Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, Peter J Liu
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 |
|---|---|---|---|
| Apply sequence likelihood calibration after fine-tuning | PEGASUS-XXL · CNN/DailyMail test | ROUGE-1: 44.31 -> 47.97 | Improved |
| Apply sequence likelihood calibration after fine-tuning | PEGASUS-XXL · RedditTIFU-long test | ROUGE-1: 28.73 -> 32.03 | Improved |
| Replace token embeddings with contextualized decoder representations for candidate-target similarity (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.51 -> 46.55 | Improved |
| Use decoder representations rather than directly optimizing ROUGE as the similarity function (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.47 -> 46.55 | No clear effect |
| Use rank loss instead of margin, list-rank, or reward loss (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.73 vs 46.11; 46.62; 46.49 | Mixed |
| Add token-level regularization rather than removing it (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.54 -> 46.73 | Improved |
| Use KL divergence rather than cross-entropy regularization (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.73 -> 46.80 | No clear effect |
| Use beam search rather than diverse beam or nucleus sampling for calibration candidates (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.31 -> 46.50; 46.45 -> 46.50 | Mixed |
| Select the fine-tuning checkpoint by perplexity instead of ROUGE (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.66 -> 47.36 | Improved |
| Omit length normalization for calibrated decoding (ablation) | SLiC · CNN/DailyMail validation | ROUGE-1: 46.44 -> 46.49 | No clear effect |
| Increase evaluation beam candidates for calibrated models | SLiC · CNN/DailyMail validation | ROUGE-1: 45.55 -> 46.63 | Improved |
| Increase evaluation beam candidates for fine-tuned-only models | PEGASUS-LARGE fine-tuned · CNN/DailyMail validation | ROUGE-1: 45.34 -> 44.59 | Hurt |
Closest papers by meaning
- Language Model Decoding as Direct Metrics Optimization (ICLR 2024 · similarity 0.86)
- Language Generation with Strictly Proper Scoring Rules (ICML 2024 · similarity 0.86)
- Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo (ICLR 2025 · similarity 0.86)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.86)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.85)
- Large language model validity via enhanced conformal prediction methods (NeurIPS 2024 · similarity 0.85)
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.
