Research map / Architectures and efficiency / Attention and transformers
Empirical Analysis of Beam Search Performance Degradation in Neural Sequence Models
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 |
|---|---|---|---|
| Increase beam width from 5 to 250 without constraints | beam search decoding of convolutional seq2seq · WMT'14 En-De | BLEU4: 26.11 to 21.38 | Hurt |
| Apply discrepancy gap constraint (M=1.5) to beam search | beam search decoding of convolutional seq2seq · WMT'14 En-De | BLEU4: 21.38 to 26.29 at B=250 | Helped |
| Apply beam candidate rank constraint (N=2) to beam search | beam search decoding of convolutional seq2seq · WMT'14 En-De | BLEU4: 21.38 to 26.10 at B=250 | Helped |
| Increase beam width from 5 to 250 without constraints | beam search decoding of convolutional seq2seq · WMT'14 En-Fr | BLEU4: 40.83 to 35.03 | Hurt |
| Apply discrepancy gap constraint (M=2.0) to beam search | beam search decoding of convolutional seq2seq · WMT'14 En-Fr | BLEU4: 35.03 to 41.06 at B=250 | Helped |
| Increase beam width from 3 to 250 without constraints | beam search decoding of summarization model · Gigaword | R-1 F: 34.22 to 33.23 | Hurt |
| Apply discrepancy gap constraint (M=0.85) to beam search | summarization beam search · Gigaword | R-1 F: 33.23 to 34.32 at B=250 | Helped |
| Increase beam width from 3 to 250 without constraints | beam search decoding of image captioning model · MSCOCO | BLEU4: 32.36 to 29.79 | Hurt |
| Apply discrepancy gap constraint (M=0.45) to beam search | image captioning beam search · MSCOCO | BLEU4: 29.79 to 32.35 at B=250 | Helped |
| Apply beam candidate rank constraint (N=2) to beam search | image captioning beam search · MSCOCO | BLEU4: 29.79 to 30.87 at B=250; gap constraint is better here | Helped |
| Length normalization of hypothesis log-likelihood | WMT'14 En-De beam search · WMT'14 En-De | BLEU4 | No clear effect |
Closest papers by meaning
- A Continuous Relaxation of Beam Search for End-to-End Training of Neural Sequence Models (AAAI 2018 · similarity 0.90)
- Bidirectional Beam Search: Forward-Backward Inference in Neural Sequence Models for Fill-In-The-Blank Image Captioning (CVPR 2017 · similarity 0.90)
- Provable Test-Time Scaling for Beam Search in LLM Reasoning (NeurIPS 2026 · similarity 0.88)
- Conformal Autoregressive Generation: Beam Search with Coverage Guarantees (AAAI 2024 · similarity 0.86)
- Dynamic-Width Speculative Beam Decoding for LLM Inference (AAAI 2025 · similarity 0.86)
- Closing the Curious Case of Neural Text Degeneration (ICLR 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.