Research map / Learning theory and algorithms / Combinatorial and search optimization
Optimal Completion Distillation for Sequence Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace MLE with OCD | seq2seq · WSJ eval92 | WER (%): 10.6 -> 9.3 | Helped |
| replace MLE with OCD | seq2seq · WSJ eval92 | CER (%): 3.6 -> 3.1 | Helped |
| replace scheduled sampling (0.0 -> 0.55) with OCD | seq2seq · WSJ eval92 | WER (%): 10.2 -> 9.3 | Helped |
| sample 100% from model with ground-truth log-likelihood (SS 1.0->1.0) instead of OCD (ablation) | seq2seq · WSJ eval92 | CER (%): 3.1 -> 12.1 | Hurt |
| fine-tune pretrained SS model with 100% sampling | seq2seq SS · WSJ eval92 | CER (%): 3.6 -> 3.8 | Hurt |
| use a single optimal target (shortest) instead of all optimal targets (ablation) | OCD · WSJ eval92 | WER (%): 9.3 -> 12.7 | Hurt |
| use a single optimal target (same #words) instead of all optimal targets (ablation) | OCD · WSJ eval92 | WER (%): 9.3 -> 10.2 | Hurt |
| use a single optimal target (same #words) instead of all optimal targets (ablation) | OCD · WSJ eval92 | CER (%): 3.1 -> 3.3 | Hurt |
| replace MLE with OCD | seq2seq · LibriSpeech test-clean | WER (%): 5.7 -> 4.5 | Helped |
| replace MLE with OCD | seq2seq · LibriSpeech test-other | WER (%): 15.4 -> 13.3 | Helped |
| replace MLE with OCD | seq2seq · LibriSpeech test-clean | CER (%): 2.9 -> 1.7 | Helped |
| replace MLE with OCD | seq2seq · LibriSpeech test-other | CER (%): 8.4 -> 6.4 | Helped |
Closest papers by meaning
- Sequence to Multi-Sequence Learning via Conditional Chain Mapping for Mixture Signals (NeurIPS 2020 · similarity 0.87)
- Imputer: Sequence Modelling via Imputation and Dynamic Programming (ICML 2020 · similarity 0.87)
- Latent Sequence Decompositions (ICLR 2017 · similarity 0.87)
- AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness (NeurIPS 2025 · similarity 0.85)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.84)
- Heuristic-free Knowledge Distillation for Streaming ASR via Multi-modal Training (AAAI 2025 · similarity 0.83)
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.