Skip to content

MindSearch: Mimicking Human Minds Elicits Deep AI Searcher

🕒 Published (v1): 2025-01-01 · Source: ICLR · Venue: ICLR 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MindSearch is a multi-agent LLM framework for deep web information retrieval that models query decomposition as a dynamically constructed directed acyclic graph (DAG), coordinating a WebPlanner and parallel WebSearchers. It processes 300+ web pages in under 3 minutes and, using InternLM2.5-7B, outperforms ChatGPT-Web (GPT-4o) and Perplexity.ai Pro in human preference evaluations.

Problem

Single-shot RAG-style search engines fail on complex queries for three reasons: (1) complex questions require multi-step decomposition before retrieval; (2) massive web content volume introduces noise that degrades LLM integration; (3) retrieved content rapidly exceeds LLM context windows, limiting coherent synthesis.

Method

MindSearch consists of two agent roles:

WebPlanner — models query resolution as iterative DAG construction \(G(Q) = \langle V, E \rangle\), where nodes \(V\) are atomic sub-questions and edges \(E\) encode reasoning dependencies. The planner emits Python code at each turn (using predefined graph-manipulation functions) to add nodes/edges, which are executed by a Python interpreter. Newly added sibling nodes (no inter-dependencies) are dispatched to WebSearchers in parallel.

WebSearcher — a hierarchical RAG agent that (1) rewrites the sub-question into multiple similar queries, (2) calls search APIs (Google/Bing/DuckDuckGo) and deduplicates results by URL, (3) prompts the LLM to select pages for full-text reading, and (4) produces a summary answer. Context isolation is maintained by prefixing each WebSearcher's input with its parent node response and the root question, preventing local-receptive-field information loss without overwhelming any single agent's context.

Key Contributions

  • DAG-based "planning-as-coding" paradigm for multi-step web search, enabling sequential and parallel sub-query execution
  • Hierarchical coarse-to-fine retrieval within WebSearcher (multi-query generation → URL dedup → page selection → summarization)
  • Explicit multi-agent context partitioning that scales to 300+ pages within a 3-minute wall-clock budget
  • Demonstration that an open-source 7B model (InternLM2.5-7B) with MindSearch beats proprietary search products in human preference

Results

  • Open-set QA (human preference, 100 real queries): MindSearch (InternLM2.5-7B) preferred over ChatGPT-Web and Perplexity.ai Pro on depth (83%) and breadth; facticity improvement smaller (70%)
  • Closed-set QA (Table 1):
  • GPT-4o backend: +4.7% average over ReAct Search (59.8 vs 55.1)
  • InternLM2.5-7B backend: +6.3% average over ReAct Search (49.2 vs 42.9)
  • Largest gains on hard HotpotQA: 78% vs 70% (GPT-4o), 49% vs 42% (InternLM)
  • Ablations (Table 2, HotpotQA, InternLM2.5-7B): Graph construction (64.0) > CodeAct (61.3) > ReAct (58.0) > raw LLM (37.6)
  • WebSearcher ablations (Table 3): Removing page selection costs 6 pts; removing multi-query generation costs 3.4 pts; removing prefix-previous-context costs 0.7 pts
  • Inference scaling: MindSearch achieves better performance per inference step than ReAct multi-turn search, showing more efficient test-time compute scaling

Limitations

  • Facticity gains are weaker than depth/breadth gains; the authors attribute this to long-context hallucination when fine-grained results distract from the original query
  • Maximum interaction turns capped at 10; performance plateaus beyond this, limiting adaptability for extremely deep queries
  • Context management heuristic (prefix parent + root) is empirical; no principled ablation of alternative context propagation strategies
  • Evaluation of open-set QA relies on human preference voting (5 experts, majority vote) rather than automated metrics, limiting reproducibility at scale

Relevance to Harnesses / Meta-Harnesses

MindSearch is a concrete meta-harness design: a coordinating planner agent that dynamically generates a DAG of sub-tasks and dispatches them to specialized executor agents, directly instantiating the orchestrator-worker pattern central to meta-harness research. Its "planning-as-coding" mechanism — where the harness structure itself is emitted as executable Python at runtime — is a strong example of self-describing, dynamically reconfigured pipelines. The explicit context partitioning strategy (parent-prefix propagation through DAG edges) addresses a core meta-harness engineering problem: how to pass just enough context between agents without blowing up any single agent's window. The inference-time scaling analysis (single-turn vs. ReAct vs. DAG multi-agent) provides empirical evidence for why hierarchical harness topology outperforms flat sequential ones on complex tasks.