Research map / Multi-agent systems and games / LLM agents
MindSearch: Mimicking Human Minds Elicits Deep AI Searcher
Read the paperCheck your idea against 108,967 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 MindSearch, an LLM multi-agent framework where a WebPlanner builds a query DAG via code and parallel WebSearchers perform hierarchical web search, improving depth and breadth of AI search answers.
Problem
LLM-plus-search-engine systems perform poorly on complex information-seeking queries: single retrieval cannot cover decomposed sub-questions, integrating information spread over many noisy pages is hard, and long web content exceeds LLM context length.
Approach
Model multi-step search as a dynamically constructed DAG of atomic sub-questions that the planner LLM builds through Python code calls; each node is handled in parallel by a WebSearcher agent that does multi-query search, page selection and summarization, which keeps each agent's context short.
Key results
- On closed-set QA with GPT-4o, MindSearch averages 59.8 vs 55.1 for ReAct Search and 53.5 without search.
- With InternLM2.5-7B-Chat on HotpotQA, MindSearch reaches 64.0 average performance versus 58.0 for ReAct and 61.3 for CodeAct.
- Human experts preferred MindSearch (InternLM2.5-7B) responses over ChatGPT-Web and Perplexity.ai Pro on depth and breadth, though factuality was not clearly better (70% vs 83% breadth comparison noted).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MindSearch instead of ReAct Search | GPT-4o · Closed-set QA (AVG) | accuracy: 55.1 -> 59.8 | Helped |
| use MindSearch instead of ReAct Search | InternLM2.5-7B-Chat · Closed-set QA (AVG) | accuracy: 42.9 -> 49.2 | Helped |
| replace ReAct planning with graph-based code planning (ablation) | MindSearch WebPlanner · HotpotQA | performance: 58.0 -> 64.0 | Helped |
| replace CodeAct function calling with graph construction planning (ablation) | MindSearch WebPlanner · HotpotQA | performance: 61.3 -> 64.0 | Helped |
| add web search (ablation) | MindSearch · HotpotQA | performance: 37.6 -> 64.0 | Helped |
| add multi-query generation and aggregation (ablation) | MindSearch WebSearcher · HotpotQA | performance: 60.6 -> 64.0 | Helped |
| add prefix previous context (ablation) | MindSearch WebSearcher · HotpotQA | performance: 63.3 -> 64.0 | Helped |
| add page selection (ablation) | MindSearch WebSearcher · HotpotQA | performance: 58.0 -> 64.0 | Helped |
Surprising findings
- MindSearch generated slightly fewer search queries than ReAct (3.2 vs 3.5 on average), suggesting better decomposition reduces redundant searches.
- Despite improved depth and breadth, MindSearch did not clearly improve factuality over baselines in human evaluation, which the authors attribute to long-context distraction.
Limitations
- Citation quality of web references was not comprehensively evaluated.
- Does not support visual inputs or interaction with web pages.
- Factuality/hallucination during web browsing remains a problem; open-set evaluation uses only 100 queries and 5 experts.
Experimental setup
Datasets: Bamboogle Musique HotpotQA
Benchmarks: Bamboogle Musique HotpotQA open-set human-curated queries (100)
Metrics: accuracy human preference (depth, breadth, factuality) inference time
Baselines: ReAct Search w/o Search Engine CodeAct self-ask Searchain Perplexity.ai Pro ChatGPT-Web
Builds on: ReAct CodeAct Searchain self-ask
Closest papers by meaning
- Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.91)
- InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.91)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
- WideSearch: Benchmarking Agentic Broad Info-Seeking (ICLR 2026 · similarity 0.88)
- UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking (ICLR 2026 · similarity 0.88)
- GISA: A Benchmark for General Information-Seeking Assistant (NeurIPS 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.