Research map / Architectures and efficiency / RNNs and normalization
A Continuous Relaxation of Beam Search for End-to-End Training of 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 |
|---|---|---|---|
| Train with continuous beam direct loss and anneal α | Seq2seq · CCG Supertagging test | tag-level accuracy: 80.35 -> 85.82 | Helped |
| Train with continuous beam direct loss and anneal α | Seq2seq · NER test | macro F1: 54.92 -> 56.38 | Helped |
| Use hard beam search instead of greedy decoding with a CE-trained model | Seq2seq · CCG Supertagging test | tag-level accuracy: 80.35 -> 82.42 | Helped |
| Use hard beam search instead of greedy decoding with a CE-trained model | Seq2seq · NER test | macro F1: 54.92 -> 51.34 | Hurt |
| Use annealed α instead of constant α=1.0 for direct-loss training (ablation) | Continuous beam direct loss · CCG Supertagging test | tag-level accuracy: 82.65 -> 85.82 | Helped |
| Use annealed α instead of constant α=1.0 for direct-loss training (ablation) | Continuous beam direct loss · NER test | macro F1: 54.08 -> 56.15 | Helped |
| Use annealed α instead of constant α=1.0 for approximate hinge training (ablation) | Approximate hinge · CCG Supertagging test | tag-level accuracy: 83.36 -> 83.54 | Helped |
| Use annealed α instead of constant α=1.0 for approximate hinge training (ablation) | Approximate hinge · NER test | macro F1: 44.67 -> 45.98 | Helped |
| Train with approximate hinge loss rather than continuous beam direct loss | Seq2seq · CCG Supertagging test | tag-level accuracy: 85.82 -> 83.54 | Hurt |
| Train with approximate hinge loss rather than continuous beam direct loss | Seq2seq · NER test | macro F1: 56.15 -> 45.98 | Hurt |
| Use soft beam decoding instead of hard beam decoding after direct-loss training | Continuous beam direct loss with annealed α · CCG Supertagging test | tag-level accuracy: 85.82 -> 85.78 | No clear effect |
| Use soft beam decoding instead of hard beam decoding after direct-loss training | Continuous beam direct loss with annealed α · NER test | macro F1: 56.15 -> 56.38 | No clear effect |
Closest papers by meaning
- Empirical Analysis of Beam Search Performance Degradation in Neural Sequence Models (ICML 2019 · similarity 0.90)
- Provable Test-Time Scaling for Beam Search in LLM Reasoning (NeurIPS 2026 · similarity 0.84)
- Entropy-informed Decoding: Adaptive Information-Driven Branching (ICML 2026 · similarity 0.84)
- Conformal Autoregressive Generation: Beam Search with Coverage Guarantees (AAAI 2024 · similarity 0.84)
- Dynamic-Width Speculative Beam Decoding for LLM Inference (AAAI 2025 · similarity 0.83)
- Cut Your Losses in Large-Vocabulary Language Models (ICLR 2025 · similarity 0.82)
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.