Papers › Calibrating Sequence likelihood Improves Conditional Language Generation
Calibrating Sequence likelihood Improves Conditional Language Generation
Yao Zhao, Misha Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, Peter J. Liu
Conditional language models are predominantly trained with maximum likelihood estimation (MLE), giving probability mass to sparsely observed target sequences. While MLE trained models assign high probability to plausible sequences given the context, the model probabilities often do not accurately rank-order generated sequences by quality. This has been empirically observed in beam search decoding as output quality degrading with large beam sizes, and decoding strategies benefiting from heuristics such as length normalization and repetition-blocking. In this work, we introduce sequence likelihood calibration (SLiC) where the likelihood of model generated sequences are calibrated to better align with reference sequences in the model's latent space. With SLiC, decoding heuristics become unnecessary and decoding candidates' quality significantly improves regardless of the decoding method. Furthermore, SLiC shows no sign of diminishing returns with model scale, and presents alternative ways to improve quality with limited training and inference budgets. With SLiC, we exceed or match SOTA results on a wide range of generation tasks spanning abstractive summarization, question generation, abstractive question answering and data-to-text generation, even with modest-sized models.
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
1 archive task tag without a task page not shown.
Results from the paper archive 2025-07-28
| Task | Dataset | Model | Metric | Value | Rank at snapshot | Leaderboard | Report |
|---|---|---|---|---|---|---|---|
| Abstractive Text Summarization | CNN / Daily Mail | Pegasus | ROUGE-1 | 47.36 | #3 of 53 | Archive leaderboard | report |
| Abstractive Text Summarization | CNN / Daily Mail | Pegasus | ROUGE-2 | 24.02 | #3 of 53 | Archive leaderboard | report |
| Abstractive Text Summarization | CNN / Daily Mail | Pegasus | ROUGE-L | 44.45 | #3 of 53 | Archive leaderboard | report |
| Text Summarization | Reddit TIFU | PEGASUS 2B + SLiC | ROUGE-1 | 32.03 | #1 of 5 | Archive leaderboard | report |
| Text Summarization | Reddit TIFU | PEGASUS 2B + SLiC | ROUGE-2 | 11.13 | #1 of 5 | Archive leaderboard | report |
| Text Summarization | Reddit TIFU | PEGASUS 2B + SLiC | ROUGE-L | 25.51 | #1 of 5 | Archive leaderboard | report |
| Text Summarization | SAMSum | PEGASUS 2B + SliC | ROUGE-1 | 54.37 | #4 of 12 | Archive leaderboard | report |
| Text Summarization | SAMSum | PEGASUS 2B + SliC | ROUGE-2 | 29.88 | #4 of 12 | Archive leaderboard | report |
| Text Summarization | SAMSum | PEGASUS 2B + SliC | ROUGE-L | 45.89 | #4 of 12 | 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
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