Research map / Multi-agent systems and games / LLM agents
FlowSearcher: Synthesizing Memory-Guided Agentic Workflows for Web Information Seeking
Keyi Xiang, Zeyu Feng, Zhuoyi Lin, Yueming Lyu, Shi Boyuan, Yew-Soon Ong, Ivor Tsang, Haiyan Yin
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 FlowSearcher, a web information seeking framework that synthesizes memory-guided agentic workflows for deep research agents instead of fixed ReAct-style tool chains.
Problem
Existing web search agents follow rigid ReAct-style tool chains with fixed workflow structures, limiting flexibility across diverse query types and tool-use strategies.
Approach
Builds on agentic workflow synthesis guided by memory, so the agent composes workflow structures adaptively per query rather than following a fixed ReAct chain. Full details unavailable beyond the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| expand block library from First-Hit only to First-Hit + General (ablation) | FlowSearcher (GPT-4o-mini) · GAIA | Pass@1 avg: 27.2 -> 35.0 | Improved |
| remove all block restrictions (add in-depth browsing and parallel/ensemble blocks) (ablation) | FlowSearcher (GPT-4o-mini) · GAIA | Pass@1 avg: 35.0 -> 55.3 | Improved |
| add full hierarchical memory (ablation) | FlowSearcher (GPT-4o-mini) · GAIA 103 tasks | successful tasks: 42 -> 57 | Improved |
| store only successful episodes in memory (ablation) | FlowSearcher (GPT-4o-mini) · GAIA tasks 1-40 | successful tasks: 18 -> 20 (at 40 tasks); 42 -> 53 (at 103) | Mixed |
| store only unsuccessful episodes in memory (ablation) | FlowSearcher (GPT-4o-mini) · GAIA 103 tasks | successful tasks: 42 -> 48 | Improved |
| add RL fine-tuning to WebThinker | WebThinker (QwQ-32B) · GAIA | Pass@1 avg: 44.7 -> 48.5 | Improved |
| add RL fine-tuning to WebThinker | WebThinker (QwQ-32B) · BrowseComp | Pass@1 avg: 2.3 -> 2.7 | No clear effect |
| replace ReAct agent with FlowSearcher workflow synthesis | QwQ-32B · BrowseComp | Pass@1 avg: 3.8 -> 8.0 | Improved |
| add search agency (vanilla ReAct) to Qwen-2.5-32B | Qwen-2.5-32B · GAIA | Pass@1 avg: 13.0 -> 31.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WebDancer — WebDancer: Towards Autonomous Information Seeking Agency (NeurIPS 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Baselines: ReAct-style tool chains
Builds on: ReAct
Closest papers by meaning
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
- Agent Workflow Memory (ICML 2025 · similarity 0.89)
- InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.88)
- Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.87)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.87)
- UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking (ICLR 2026 · similarity 0.87)
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.
