Research map / Large language models / Agents, code and math
Thinker: Training LLMs in Hierarchical Thinking for Deep Search via Multi-Turn Interaction
Jun Xu, Xinkai Du, Yu Ao, Peilong Zhao, Yang Li, Ling Zhong, Lin Yuan, Zhongpu Bo, Xiaorui Wang, Mengshu Sun, Zhengke Gui, Dalong Zhang and 8 more
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 depth solving (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.415 -> 0.452 | Improved |
| add knowledge boundary determination (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.447 -> 0.452 | No clear effect |
| add focusing and reasoning module (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.445 -> 0.452 | Improved |
| add logical function (Action) representation (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.449 -> 0.452 | No clear effect |
| add knowledge boundary determination (ablation) | Thinker · TriviaQA and Bamboogle | search queries: 16.0% and 17.8% reduction in search queries | Improved |
| increase training data from 1% to 100% (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.406 -> 0.452 | Improved |
| raise max depth searches from 1 to 5 (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.415 -> 0.452 | Improved |
| raise max depth searches from 2 to 5 (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.442 -> 0.452 | No clear effect |
| retrieve 3 documents instead of 1 (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.405 -> 0.452 | Improved |
| retrieve 5 documents instead of 3 (ablation) | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.452 -> 0.447 | No clear effect |
| use KAG hybrid graph retriever with native Math and Deduce | Thinker (BGE-M3) · Self-built corpus (3 multi-hop datasets) | EM: 0.490 -> 0.520 | Improved |
| add GRPO reinforcement learning after SFT | Thinker (Qwen2.5-7B) · Seven QA datasets average | EM: 0.452 -> 0.479 | Improved |
Closest papers by meaning
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.90)
- Thinkless: LLM Learns When to Think (NeurIPS 2025 · similarity 0.90)
- DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.89)
- DeepCompress: A Dual Reward Strategy for Dynamically Exploring and Compressing Reasoning Chains (ICLR 2026 · similarity 0.89)
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.89)
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.
