Research map / Multi-agent systems and games / LLM agents

Flash-Searcher: Fast and Effective Web Agents via DAG-Based Parallel Execution

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

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 Flash-Searcher, a DAG-based parallel tool-using agent framework that cuts execution steps and latency versus sequential agents, and distills its trajectories into Qwen-2.5 32B/72B models via SFT.

Problem

Tool-augmented LLM agents and multi-agent web search systems execute sequentially, requiring many interaction steps and long latency, creating a quality-efficiency trade-off on complex information-seeking tasks.

Approach

Decompose a task into a DAG of subtasks and aggressively schedule ready subtasks in parallel (including ones partially verifiable via cross-validation), periodically refining the DAG and summarizing progress, so that tool calls run concurrently and fewer steps are needed.

Key results

• Flash-Searcher with GPT-5 reaches 67.7% on BrowseComp, 83 on xbench-DeepSearch, 82.5% on GAIA text-only and 44.0 on HLE.
• Flash-Searcher (GPT-5-mini) reduces average steps from 11.2 to 7.4 (about 35%) versus OAgents and cuts execution time by about 65%.
• SFT-distilled Flash-Searcher-72B on Qwen-2.5-72B reports 18.9 BrowseComp, 68.0 xbench-DeepSearch, 61.2 GAIA and 20.2 HLE.

Surprising findings

• Extending the step limit from 40 to 80 improved performance by 5.0 (32B) and 7.0 (72B) points on BrowseComp-100, suggesting the 40-step cap is a bottleneck.
• Disabling code execution tools is a deliberate choice that leaves math-heavy HLE performance suboptimal.

Limitations

• 40-step budget truncates many BrowseComp tasks, and crawl truncation to 60k characters loses information.
• No code execution tool, causing weak math reasoning on HLE.
• Judge-based Pass@1 evaluation with GPT-4.1-mini; execution time depends on external API rate limits.

Experimental setup

Datasets: AFM ASearcher WebShaper WebWalkerQA GAIA BrowseComp xbench-DeepSearch HLE

Benchmarks: GAIA BrowseComp xbench-DeepSearch HLE

Metrics: Pass@1 execution steps tool calls per step execution time

Baselines: OAgents OWL-Roleplaying WebDancer WebSailor AFM-RL WebShaper BrowseMaster MiroFlow OpenAI DeepResearch Alita

Builds on: ReAct OAgents Search-o1 ParallelSearch Graph of Thoughts

Closest papers by meaning

• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
• Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents (ICLR 2026 · similarity 0.88)
• ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents (ICLR 2025 · similarity 0.88)
• Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.87)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.87)
• A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.87)