Research map / Large language models / Retrieval and knowledge
Ask the Right Questions: Active Question Reformulation with Reinforcement 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 |
|---|---|---|---|
| Use active reformulation and learned CNN answer selection instead of answering the original question directly | BiDAF on SearchQA test · SearchQA test | F1: 34.6 -> 45.6; 11.4 F1 points | Helped |
| Use the reformulator's top hypothesis as a single rewrite (ablation) | BiDAF on SearchQA test · SearchQA test | F1: 34.6 -> 36.8; 2.2 F1 | Helped |
| Use CNN answer selection over multiple rewrites instead of the top hypothesis alone (ablation) | AQA on SearchQA test · SearchQA test | F1: 36.8 -> 45.6 | Helped |
| Train reformulator with multilingual data only, removing Paralex monolingual data (ablation) | Base-NMT · MSCOCO VAL2014 | BLEU: 11.4 -> 5.0 | Hurt |
| Train with multilingual data only, without Paralex (ablation) | Base-NMT on SearchQA · SearchQA | F1: Top Hypothesis scores are at least 5 points lower in all metrics; CNN scores are 2-3 points lower | Hurt |
| Add Quora duplicate-question training (ablation) | Base-NMT on SearchQA · SearchQA | QA performance: maximum delta ... +0.5 points | No clear effect |
| Add Quora training (ablation) | Base-NMT on MSCOCO VAL2014 · MSCOCO VAL2014 | BLEU: 11.4 -> 11.6 | Helped |
| Optimize reformulations for QA reward rather than natural-language paraphrase quality | AQA reformulator · SearchQA | F1 | Helped |
| Use CNN selection with AQA rewrites rather than CNN selection with benchmark-generated rewrites | AQA on SearchQA · SearchQA | F1: about 3% | Helped |
Closest papers by meaning
- Learning to Query, Reason, and Answer Questions On Ambiguous Texts (ICLR 2017 · similarity 0.90)
- A$^2$Search: Ambiguity-Aware Question Answering with Reinforcement Learning (ICLR 2026 · similarity 0.89)
- Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.86)
- Enhancing Question Generation through Diversity-Seeking Reinforcement Learning with Bilevel Policy Decomposition (AAAI 2025 · similarity 0.86)
- QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation (ICLR 2026 · similarity 0.86)
- AGILE: A Novel Reinforcement Learning Framework of LLM Agents (NeurIPS 2024 · similarity 0.86)
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.