Research map / Large language models / AI and society
Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration
Ante Wang, Yujie Lin, Jingyao Liu, Suhang Wu, Hao Liu, Xinyan Xiao, Jinsong Su
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add sampling-based data filtering (ablation) | SFT, Qwen3-1.7B (no thinking) · GSM-MC | accuracy: 34.71 -> 38.60 | Improved |
| add sampling-based data filtering (ablation) | SFT, Qwen3-1.7B (thinking) · GSM-MC | accuracy: 27.20 -> 44.15 | Improved |
| add sampling-based data filtering (ablation) | RL, Qwen3-1.7B (no thinking) · GSM-MC | accuracy: 55.19 -> 60.82 | Improved |
| add sampling-based data filtering (ablation) | RL, Qwen3-1.7B (thinking) · GSM-MC | accuracy: 57.09 -> 62.13 | Improved |
| enable thinking mode | vanilla Qwen3-8B · GSM-MC | accuracy: 55.59 -> 44.92 | Hurt |
| enable thinking mode | RL-trained Qwen3-8B · GSM-MC | accuracy: 81.73 -> 85.53 | Improved |
| increase interaction from two to four turns | RL Qwen3-8B (thinking) · GSM-MC | accuracy: 85.53 -> 88.67 | Improved |
| add SFT before RL | RL Qwen3-8B (no thinking) · GSM-MC | accuracy: 81.73 -> 79.24 | Hurt |
| add SFT before RL | RL Qwen3-1.7B (no thinking) · GSM-MC | accuracy: 60.82 -> 73.68 | Improved |
| apply SFT training | vanilla Qwen3-1.7B (no thinking) · GSM-MC | accuracy: 0.95 -> 38.60 | Improved |
| apply SFT+RL training | Llama-3.2-3B-Instruct · GSM8K | accuracy: 64.06 -> 75.74 | Improved |
| apply SFT training | Llama-3.2-3B-Instruct · GSM8K | accuracy: 64.06 -> 48.60 | Hurt |
Closest papers by meaning
- Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.86)
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.86)
- Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL (NeurIPS 2025 · similarity 0.86)
- SophiaVL-R1: Reinforcing MLLMs Reasoning with Thinking Reward (ICLR 2026 · similarity 0.86)
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025 · 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.
