Research map / Large language models / Reasoning and chain of thought
CP-Search: A Chain Progressive Search Training Framework Incentivizing the Cognitive Behaviors for Searching in LLMs
Zehua Wang, Shipeng Li, Buzhou Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CP-Search, a chain-progressive search training framework that uses reinforcement learning to incentivize cognitive search behaviors in LLMs for multi-hop retrieval-augmented reasoning.
Problem
Existing RAG methods model reasoning paths poorly for complex multi-hop reasoning, and RL-based reasoning training has been applied to search but the abstract is truncated before details of its limitations are stated.
Approach
A chain-progressive search training framework that incentivizes cognitive search behaviors in LLMs during RL training, aimed at improving multi-hop reasoning with retrieval; specific mechanism not stated in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cognitive-behavior SFT stage (ablation) | CP-Search · 2WikiMultiHopQA | ACCR: 56.0 -> 72.4 | Improved |
| add cognitive-behavior SFT stage (ablation) | CP-Search · MuSiQue | ACCR: 21.6 -> 32.2 | Improved |
| add procedural-supervision RL stage after SFT (ablation) | CP-Search · 2WikiMultiHopQA | ACCR: 64.2 -> 72.4 | Improved |
| add procedural-supervision RL stage after SFT (ablation) | CP-Search · MuSiQue | F1: 26.5 -> 36.8 | Improved |
| replace GRPO with CPGRPO procedural reward (ablation) | CP-Search · 2WikiMultiHopQA | ACCR: 68.8 -> 72.4 | Improved |
| replace GRPO with CPGRPO procedural reward (ablation) | CP-Search · MuSiQue | F1: 35.2 -> 36.8 | Improved |
| CP-Search vs R1-Searcher on HotpotQA | Qwen2.5-7B-Instruct · HotpotQA | AVG: 70.2 -> 69.2 | Hurt |
| CP-Search vs R1-Searcher on 2Wiki | Qwen2.5-7B-Instruct · 2WikiMultiHopQA | AVG: 64.3 -> 72.2 | Improved |
| RL stage after SFT: effect on capability (ablation) | CP-Search · 2Wiki, MuSiQue | pass@k | Mixed |
Closest papers by meaning
- DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.89)
- Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search (ICML 2025 · similarity 0.89)
- Scent of Knowledge: Optimizing Search-Enhanced Reasoning with Information Foraging (NeurIPS 2025 · similarity 0.88)
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning (NeurIPS 2025 · similarity 0.88)
- Erase to Improve: Erasable Reinforcement Learning for Search-Augmented LLMs (ICLR 2026 · similarity 0.88)
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.
