Research map / Multi-agent systems and games / Human-AI and goal-conditioned
WebSeer: Training Deeper Search Agents through Reinforcement Learning with Self-Reflection
Guanzhong He, Zhen Yang, Jinxin Liu, Bin Xu, Lei Hou, Juanzi Li
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 self-reflective multi-answer submission during RL (vs single answer only) (ablation) | WebSeer · MHQA benchmarks | accuracy | Improved |
| add cold-start SFT initialization before RL (ablation) | WebSeer · MHQA benchmarks | accuracy | Improved |
| apply SFT to 3B model | Qwen2.5-3B · HotpotQA | accuracy: drops 3.52 points | Hurt |
| apply SFT to 3B model | Qwen2.5-3B · SimpleQA | accuracy: gains 8.06 | Improved |
| apply SFT to 7B model | Qwen2.5-7B · HotpotQA and SimpleQA | accuracy | Hurt |
| scale to 14B model for SFT then RL | WebSeer · HotpotQA and SimpleQA | accuracy | Improved |
| increase share of long multi-refinement trajectories in SFT data (ablation) | WebSeer SFT · HotpotQA | tool calls and accuracy | Mixed |
| use web search at inference instead of local RAG | WebSeer · 7 MHQA benchmarks in-domain avg | accuracy: 78.9 -> 82.6 | Improved |
| WebSeer vs Search-r1 in-domain | WebSeer · In-domain avg | accuracy: exceeding Search-r1 by 12.5 points | Improved |
| WebSeer vs prior best on Bamboogle (local RAG) | WebSeer · Bamboogle | accuracy: 81.6, 12.8-point improvement | Improved |
| add RL after SFT for 3B and 7B models | small models · HotpotQA | accuracy | No clear effect |
Closest papers by meaning
- DeepDiver: Adaptive Web-Search Intensity Scaling via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards (ICML 2026 · similarity 0.89)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
- Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.89)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.89)
- Search Self-Play: Pushing the Frontier of Agent Capability without Supervision (ICLR 2026 · 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.
