Research map / Multi-agent systems and games / Cooperative multi-agent RL
DECOR: Learning to Decompose and Collaborate in Deep Search via Multi-Agent Reinforcement Learning
Ruiqing Chen, Zekun Zhang, Gong-Duo Zhang, Lihong Gu, Lin Zhou
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 |
|---|---|---|---|
| use hybrid reward (team F1 + judge + role-specific) instead of F1-only, EM-only or LLM-Judge-only (ablation) | DECOR · 2WikiMQA | LLM-Judge accuracy | Improved |
| vary reward weight alpha between team and role rewards (ablation) | DECOR · 2WikiMQA | LLM-Judge accuracy: peak at alpha = 0.6 | Mixed |
| increase retrieved documents top-k from 5 to 10 (ablation) | Search-R1 · HotpotQA | LLM-Judge accuracy: 59.1 -> 56.4 | Hurt |
| increase retrieved documents top-k from 5 to 10 (ablation) | DECOR · HotpotQA | LLM-Judge accuracy: 59.8 -> 60.1 | Improved |
| increase retrieved documents top-k from 5 to 10 (ablation) | DECOR · 2WikiMQA | LLM-Judge accuracy: 62.0 -> 62.2 | No clear effect |
| add Qwen3-Reranker (ablation) | DECOR · 2WikiMQA | LLM-Judge accuracy: 62.0 -> 65.8 | Improved |
| add Qwen3-Reranker (ablation) | DECOR · HotpotQA | LLM-Judge accuracy: 59.8 -> 60.5 | Improved |
| replace static Wikipedia corpus with live Google Search API (ablation) | DECOR · 2WikiMQA | LLM-Judge accuracy: 62.0 -> 68.4 | Improved |
| replace static corpus with live Google Search API (ablation) | DECOR · HotpotQA | LLM-Judge accuracy: 59.8 -> 62.6 | Improved |
| use multi-agent DECOR instead of monolithic Search-R1 | deep search agents · 2WikiMQA, Bamboogle | F1: 2WikiMQA 61.8 (R1-Searcher) vs 61.2; Bamboogle 52.5 (Search-R1) vs 50.3 | Mixed |
| use DECOR on MuSiQue | deep search agents · MuSiQue | F1: 24.9 (SimpleDeepSearcher) -> 23.8 | No clear effect |
Closest papers by meaning
- Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.86)
- Search-R2: Enhancing Search-Integrated Reasoning via Actor-Refiner Collaboration (ICML 2026 · similarity 0.86)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.86)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.85)
- Search Self-Play: Pushing the Frontier of Agent Capability without Supervision (ICLR 2026 · similarity 0.85)
- Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (ICML 2026 · 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.
