Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning through Dialogue Interactions by Asking Questions
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 |
|---|---|---|---|
| Allow question asking during both training and evaluation rather than training and evaluating with no questions | Cont-MemN2N · Mechanical Turk Task 4: Knowledge Verification | accuracy: 0.712 -> 0.774 | Helped |
| Allow question asking during both training and evaluation rather than training and evaluating with no questions | Cont-MemN2N · Mechanical Turk Task 8: Triple | accuracy: 0.308 -> 0.797 | Helped |
| Use contextual word embeddings instead of vanilla word embeddings (ablation) | MemN2N · Mechanical Turk Task 4: Knowledge Verification (TestAQ) | accuracy: 0.375 -> 0.774 | Helped |
| Use contextual word embeddings instead of vanilla word embeddings (ablation) | MemN2N · Mechanical Turk Task 8: Triple (TestAQ) | accuracy: 0.422 -> 0.797 | Helped |
| Train on human-collected dialogue instead of simulator-only dialogue (ablation) | Cont-MemN2N, TrainAQ/TestAQ · Mechanical Turk Task 4: Knowledge Verification | accuracy: 0.714 -> 0.774 | Helped |
| Train on human-collected dialogue instead of simulator-only dialogue (ablation) | Cont-MemN2N, TrainAQ/TestAQ · Mechanical Turk Task 8: Triple | accuracy: 0.788 -> 0.797 | Helped |
| Train with question asking but evaluate without it, rather than using no questions in either phase | Cont-MemN2N · Question Clarification and Knowledge Operation | accuracy | Hurt |
| Use a mixed QA/AQ training set instead of AQ-only training (ablation) | Cont-MemN2N · Question Clarification and Knowledge Operation | accuracy | Hurt |
| Ask a question at test time rather than answering immediately | Cont-MemN2N · Missing Question Entity, Missing Answer Entity, and Missing Everything | accuracy: <0.01 -> 0.639; <0.01 -> 0.885; <0.01 -> 0.908 | Helped |
| Use a model to select the question rather than asking the correct question directly (ablation) | Cont-MemN2N · TestModelAQ verification tasks | accuracy | No clear effect |
| Combine simulated and human-collected training data instead of using human-collected data alone (ablation) | Cont-MemN2N, TrainAQ/TestAQ · Mechanical Turk Tasks 4 and 8 | accuracy: 0.774 -> 0.792; 0.797 -> 0.825 | Helped |
| Use one answer-prediction model for QA and AQ rather than separate models (ablation) | Online RL answer prediction · Online reinforcement learning | cumulative reward | Hurt |
Closest papers by meaning
- Dialogue Learning With Human-in-the-Loop (ICLR 2017 · similarity 0.91)
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward (NeurIPS 2025 · similarity 0.86)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.85)
- Decoupling Conversational Dynamics in Full-Duplex Spoken Models through Reinforcement Learning (NeurIPS 2026 · similarity 0.84)
- Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.84)
- Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue (AAAI 2025 · similarity 0.84)
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.