Research map / Large language models / Language, safety and interpretability
Adaptive Beam Search Decoding for Discrete Keyphrase Generation
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 |
|---|---|---|---|
| Use AdaGM with reset-state training and adaptive beam decoding instead of the strongest listed baseline | keyphrase generation · present keyphrase prediction across Inspec, Krapivin, NUS, SemEval, and KP20k | F1@5: 0.299 -> 0.305; 0.342 -> 0.363; 0.432 -> 0.442; 0.349 -> 0.343; 0.321 -> 0.388 | Mixed |
| Use AdaGM with reset-state training and adaptive beam decoding instead of the strongest listed baseline | keyphrase generation · absent keyphrase prediction across Inspec, Krapivin, NUS, SemEval, and KP20k | F1@5: 0.015 -> 0.016; 0.054 -> 0.050; 0.049 -> 0.037; 0.027 -> 0.032; 0.041 -> 0.043 | Mixed |
| Add reset-state training rather than omitting all reset operations (ablation) | AdaGM · Inspec, Krapivin, and SemEval present keyphrase prediction | F1@5: 0.287 -> 0.305; 0.315 -> 0.363; 0.328 -> 0.343 | Helped |
| Add the decoder hidden-state reset in addition to resetting the input state (ablation) | AdaGM · Inspec, Krapivin, and SemEval present keyphrase prediction | F1@5: 0.273 -> 0.305; 0.335 -> 0.363; 0.335 -> 0.343 | Mixed |
| Add the decoder hidden-state reset in addition to resetting the input state (ablation) | AdaGM · Inspec, Krapivin, and SemEval present keyphrase prediction | F1@M: 0.311 -> 0.348; 0.248 -> 0.323; 0.309 -> 0.337 | Helped |
| Increase the adaptive beam filter threshold (ablation) | AdaGM · KP20k validation | present keyphrase F1@M: 0.347 at α=0.015 -> 0.363 at α=0.02 -> 0.386 at α=0.03 | Helped |
| Increase the adaptive beam filter threshold above α=0.015 (ablation) | AdaGM · KP20k validation | absent keyphrase F1@M: 0.075 at α=0.015 -> 0.051 at α=0.02 -> 0.041 at α=0.03 | Hurt |
| Use adaptive beam decoding rather than greedy-search baselines | AdaGM · five scientific keyphrase datasets | DupRatio | Helped |
| Use adaptive beam decoding rather than the compared baselines | AdaGM · one document, GTX-1080Ti | inference time: 89ms | Helped |
Closest papers by meaning
- Improving Open-Ended Text Generation via Adaptive Decoding (ICML 2024 · similarity 0.84)
- Model-Based Minimum Bayes Risk Decoding for Text Generation (ICML 2024 · similarity 0.82)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.82)
- Entropy-informed Decoding: Adaptive Information-Driven Branching (ICML 2026 · similarity 0.82)
- Controlled LLM Decoding via Discrete Auto-regressive Biasing (ICLR 2025 · similarity 0.81)
- Retrieval is Accurate Generation (ICLR 2024 · similarity 0.81)
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.