Source-linked AI summary
Flash-Searcher: Fast and Effective Web Agents via DAG-Based Parallel Execution
Tianrui Qin, Qianben Chen, Sinuo Wang, He Xing, King Zhu, He Zhu, Dingfeng Shi, Xinxin Liu, Ge Zhang, Jiaheng Liu, Yuchen Eleanor Jiang, Xitong Gao, Wangchunshu Zhou
TL;DR
Sequential tool-augmented agents become inefficient on complex tasks that require extensive interaction. Flash-Searcher reformulates task solving as dynamic DAG scheduling with parallel execution and workflow refinement, achieving strong benchmark performance and lower execution cost.
Problem
Existing multi-agent and tool-integrated reasoning frameworks often use sequential processing, producing long reasoning chains, redundant communication, and high computational overhead on complex tasks.
Method
Flash-Searcher decomposes complex tasks into dependency-aware DAGs, executes eligible subtasks in parallel, and dynamically refines the graph using intermediate results.
Results
Across challenging benchmarks, Flash-Searcher reports state-of-the-art performance, including 67.7% on BrowseComp and 83.0% on xbench, while reducing execution steps by 35% compared with OAgents.
Takeaways & Limitations
The results support parallel reasoning as an efficient and scalable paradigm for complex web-agent tasks, with gains also transferring to distilled open-source agent models.
Takeaways & Limitations
A 40-step execution limit prevented complete resolution of some complex queries, including approximately 25% of framework BrowseComp cases, and web-content truncation can cause information loss.
Abstract
from arXiv · showhide
Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks when equipped with external tools. However, current frameworks predominantly rely on sequential processing, leading to inefficient execution particularly for tasks requiring extensive tool interaction. This paper introduces Flash-Searcher, a novel parallel agent reasoning framework that fundamentally reimagines the execution paradigm from sequential chains to directed acyclic graphs (DAGs). Flash-Searcher decomposes complex tasks into subtasks with explicit dependencies, enabling concurrent execution of independent reasoning paths while maintaining logical constraints. Through dynamic workflow optimization, our framework continuously refines the execution graph based on intermediate results, effectively integrating summary module. Comprehensive evaluations across multiple benchmarks demonstrate that Flash-Searcher consistently outperforms existing approaches. Specifically, it achieves 67.7% accuracy on BrowseComp and 83% on xbench-DeepSearch, while reducing agent execution steps by up to 35% compared to current frameworks. Furthermore, when distilling this parallel reasoning pipeline into single models, we observe substantial performance gains across diverse backbone architectures, underscoring the generalizability of our methodology. Our work thus represents a significant advance in agent architecture design, offering a more scalable and efficient paradigm for complex reasoning tasks.
1 Introduction
Flash-Searcher addresses the inefficiency of sequential tool-augmented agents by organizing complex task solving around parallel execution paths. It reports strong benchmark performance, fewer execution steps, and post-training gains for open-source models.
- Existing multi-agent and tool-integrated reasoning systems are described as incurring inefficient tool use, long reasoning chains, redundant communication, and substantial computational overhead.
- FLASH-SEARCHER decomposes tasks into concurrent subtasks, manages dependency-aware reasoning graphs, and adjusts strategies dynamically to reduce redundant interactions.
- 35% fewer agent execution steps, from 11.2 to 7.4, and approximately 65% shorter execution time versus OAgents are reported for FLASH-SEARCHER with GPT-5mini.
- 82.5% average performance on GAIA and scores of 83.0 on xbench, 44.0 on HLE, and 67.7 on BrowseComp are reported for FLASH-SEARCHER.
- Lightweight post-training on curated parallel reasoning trajectories improves open-source models, while the pipeline and datasets are released for further research.
2 Related Work
Related work has advanced complex problem solving through specialized multi-agent collaboration and tool-integrated reasoning. These approaches use role-based agents, planning, tool interaction, and increasingly post-trained reasoning strategies.
- Multi-agent systems assign specialized roles to support planning, multi-turn reasoning, tool use, and interaction with task environments.
- Prior work establishes collaborative agents and tool-integrated reasoning as major approaches for complex task solving.
- Tool-integrated reasoning evolved from prompt-based fixed Thought–Action–Observation loops toward post-training paradigms.
3 Method
FLASH-SEARCHER reformulates complex tool-augmented task solving as structured concurrency over a dynamic DAG, executing eligible subtasks in parallel while periodically refining dependencies and progress. Its design combines dependency-aware planning, aggressive parallelization, state aggregation, and adaptive optimization to reduce sequential bottlenecks while preserving execution coherence.
- 3 Method: FLASH-SEARCHER reformulates linear agent workflows as dynamic DAG plans, combining decomposition, controlled parallelization, and periodic optimization.The framework is presented as a structured-concurrency approach for complex task solving.
- 3 Method: Each DAG node represents a subtask, and directed edges encode prerequisite relations that determine execution dependencies.The decomposition function identifies constituent subtasks and their interdependencies.
- 3 Method: A readiness predicate enables aggressive parallelization when prerequisites are complete or when partial execution supplies auxiliary dependency-verification signals.This hybrid criterion combines dependency satisfaction with heuristic consistency checks.
- 3 Method: Multiple eligible subtasks execute concurrently through tool or agent invocations, and their observations are combined through structured aggregation into the reasoning state.Aggregated observations support subsequent state transitions.
- 3 Method: Every Δ steps, the framework refines the DAG by removing resolved nodes, revalidating unresolved dependencies, and inserting decomposition nodes when needed.Smaller Δ increases adaptation frequency, whereas larger Δ reduces optimization overhead in stable or complex tasks.
4 Experiment
FLASH-SEARCHER is evaluated across four challenging information-retrieval and reasoning benchmarks, with results comparing framework performance, backbone scaling, and distilled agent models. The framework achieves competitive benchmark performance, while lightweight supervised fine-tuning transfers its parallel reasoning capabilities across model scales.
- 4.1.2 Main Results: FLASH-SEARCHER matches or exceeds existing agent frameworks across BrowseComp, xbench-DeepSearch, GAIA, and HLE.The comparison uses Pass@1 results across the four benchmarks.
- 4.1.2 Main Results: 67.7% on BrowseComp, 83% on xbench-DeepSearch, and 44.0% on HLE demonstrate strong performance across challenging benchmarks.The GPT-5 configuration reaches 67.7% on BrowseComp and 83% on xbench-DeepSearch, while HLE performance is reported as state of the art at 44.0%.
- 4.2.2 Agent Model Results: Lightweight supervised fine-tuning transfers parallel reasoning capabilities to standalone models, producing state-of-the-art results across diverse benchmarks and backbone scales.The paper evaluates distilled agent models against existing methods on multiple challenging benchmarks.
- 4.2.2 Agent Model Results: On Qwen-2.5-32B, FLASH-SEARCHER improves over the strongest prior method by 3.3% on BrowseComp, 5.0% on xBench-DeepSearch, and 2.0% on GAIA.The HLE result is 19.4% without code-interpreter tools.
- 4.2.2 Agent Model Results: Scaling to 72B yields gains across all benchmarks, with the largest improvements on complex multi-step tasks such as BrowseComp and xbench-DeepSearch.The cited passage reports 5% gains on those complex tasks.
- 4.2.2 Agent Model Results: The distilled results are achieved through lightweight supervised fine-tuning without reinforcement learning or tool reliance.The authors characterize parallel reasoning as transferable through minimal supervision.
5 Efficiency Analysis
FLASH-SEARCHER improves execution efficiency by parallelizing reasoning and coordinating information needs across branches. It reduces steps while maintaining higher tool utilization and stronger task success, particularly as benchmark complexity increases.
- Tool Calls: 35% fewer agent steps than OAgents and 30% fewer than OWL-Roleplaying are achieved on GAIA with GPT-5-mini.The comparison is reported in Figure 6b.
- Tool Calls: FLASH-SEARCHER outperforms OAgents on all four benchmarks, with the advantage growing as task complexity increases.The reported complexity ordering is BrowseComp > xbench-DeepSearch > HLE > GAIA.
- Tool Calls: 3.00 tool calls per step versus 0.83 for OAgents and 0.85 for OWL-Roleplaying indicate higher per-step tool utilization.The comparison is reported despite FLASH-SEARCHER using fewer total steps.
- Tool Calls: The DAG-based mechanism coordinates parallel branches to eliminate duplicate searches while preserving reasoning diversity.The framework is presented as addressing the sequential bottleneck of traditional methods.
- Tool Calls: Execution duration remains influenced by external factors such as API rate limits, although step reductions are consistent under comparable conditions.The paper reports a 35% reduction in execution steps under those conditions.
6 Conclusion
FLASH-SEARCHER reframes tool-augmented task solving as structured concurrency over dynamic DAGs. Across several benchmarks, it combines strong performance with reduced latency, and distilled variants support the scalability of parallel reasoning.
- 6 Conclusion: FLASH-SEARCHER uses dynamic DAG scheduling to enable fine-grained parallel execution while preserving logical coherence and correctness.The conclusion describes this as overcoming sequential bottlenecks in conventional tool-augmented agents.
- 6 Conclusion: 67.7% on BrowseComp accompanies substantial computational-efficiency gains from reduced latency and improved resource utilization.The conclusion presents this result alongside the framework’s broader benchmark evaluation.
- 6 Conclusion: Distilled agent variants further support parallel reasoning as a foundational paradigm for efficient, scalable, and robust AI systems.This conclusion is tied to results across complex real-world tasks.
7 Contributions
The passage identifies the paper’s listed authors: Xitong Gao and Wangchunshu Zhou.
- The listed authors are Xitong Gao and Wangchunshu Zhou.
A Limitations and Future Work
FLASH-SEARCHER improves agent efficiency but faces resource, tool, and execution-step constraints that limit some tasks. Future work targets richer orchestration, verification, and specialized tools.
- Limitations: Execution-step limits prevented complete resolution of some complex BrowseComp queries, while crawl summarization introduced information loss.The framework used a 40-step limit, and additional evaluations examined extended reasoning budgets.
- Limitations: The absence of code execution tools contributed to suboptimal mathematical reasoning performance on benchmarks such as HLE.Adding computational tools would require architectural trade-offs because parallel code execution increases output and environment-management overhead.
- Future Work: Future work could combine FLASH-SEARCHER with reflection, verification, and specialized code-execution agents when higher precision justifies additional resources.The architecture is described as compatible with supplementary verification mechanisms and specialized agents for mathematical subtasks.
- Future Work: More sophisticated orchestration could dynamically allocate reasoning resources according to task complexity and integrate emerging models or domain knowledge.
- Conclusion: The authors position FLASH-SEARCHER as a foundation for more efficient and capable agent systems despite these limitations.
B Discussion of DAG/Graph Reasoning Methods
FLASH-SEARCHER distinguishes itself from graph and parallel reasoning methods by targeting parallel external-tool execution. Its dynamic DAG planning combines concurrency with cross-validation and ongoing summarization.
- Graph-Based Reasoning: Graph of Thoughts models reasoning as graphs but emphasizes symbolic reasoning, whereas FLASH-SEARCHER focuses on tool execution.
- Parallel Reasoning: Learning Adaptive Parallel Reasoning parallelizes model-internal computation, while FLASH-SEARCHER coordinates external tool calls.
- DAG Planning: FLASH-SEARCHER relaxes strict dependency enforcement to maximize parallel execution while preserving validity through cross-validation.
- Comparison with Existing Methods: FLASH-SEARCHER specializes DAG structures for efficient information retrieval across multiple sources, unlike methods lacking explicit parallel tool execution.
- FLASH-SEARCHER: The framework dynamically optimizes DAG execution trajectories, invokes tools in parallel, cross-validates dependent subtasks, and summarizes intermediate states.
C.1 Benchmarks.
The evaluation uses four challenging benchmarks spanning general assistance, persistent browsing, Chinese deep search, and difficult academic questions. FLASH-SEARCHER employs a constrained tool setup and documented framework and inference configurations.
- Benchmarks: Four benchmarks assess FLASH-SEARCHER across information retrieval and reasoning scenarios: GAIA, BrowseComp, xbench-DeepSearch, and HLE.The benchmark suite is intended to evaluate efficiency and effectiveness across diverse task complexities.
- GAIA: GAIA evaluates real-world assistant capabilities including reasoning, multimodality, web browsing, and tool use, with experiments focused on 103 text-only validation cases.
- BrowseComp: BrowseComp measures persistent browsing through query formulation, search navigation, information extraction, and answer synthesis across 1,266 questions.
- xbench-DeepSearch: xbench-DeepSearch contains 100 expert-written Chinese questions requiring iterative search refinement, cross-source integration, and a planning-to-summarization pipeline.
- HLE: HLE contains 2,500 highly difficult multimodal questions across dozens of subjects to evaluate broad academic capabilities.
- Tool Configuration: The framework limits external tools to search and crawl, using crawl-time summarization to balance retrieval capability with trajectory simplicity and efficiency.
- Models and Settings: Experiments use diverse state-of-the-art language models with consistent comparable-setting hyperparameters, including temperature 1.0.
- Configurations: Framework, model-inference, and hardware settings are documented through Tables 2–4 to support reproducibility and alignment with execution constraints.
C.5 Detailed Results of FLASH-SEARCHER.
The detailed-results section evaluates FLASH-SEARCHER against existing agent frameworks and models on four benchmarks using Pass@1.
- Evaluation Results: Pass@1 measures the proportion of tasks successfully completed in a single attempt, and Table 5 and Figure 7 compare FLASH-SEARCHER with existing systems.
E Model Ablations on Execution Step Constraints
Extending the reasoning-step budget improves FLASH-SEARCHER model performance, indicating that some errors under the 40-step setting reflect insufficient reasoning steps. The section also situates these ablations alongside the framework’s training-data and evaluation setup.
- E Model Ablations on Execution Step Constraints: Increasing the step limit from 40 to 80 produces consistent, measurable gains across FLASH-SEARCHER model variants.The ablation evaluates FLASH-SEARCHER models on BrowseComp-100.
- E Model Ablations on Execution Step Constraints: 7.0 points: FLASH-SEARCHER-72B improves when the maximum reasoning steps increase from 40 to 80.FLASH-SEARCHER-32B improves by 5.0 points under the same extension.
- E Model Ablations on Execution Step Constraints: The reported gains attribute the 40-step bottleneck to insufficient reasoning steps rather than fundamental model limitations.The analysis frames additional computation as a possible trade-off for higher precision in resource-unconstrained deployment.
- Training setup: The training pipeline uses filtered agent trajectories formatted as multi-turn dialogues for supervised fine-tuning.The supplied training passages describe trajectory filtering, dialogue formatting, and SFT implementation details.
F.3 Model training curves
The supplied passages report a post-1977 Malko Competition winner list and investigate whether any recorded nationality corresponds to a former sovereign state. They identify Claus Peter Flor’s 1983 entry as the sole candidate, although the evidence distinguishes Germany from East Germany.
- Historical sovereignty check: A search result identifies the German Democratic Republic as a former sovereign state spanning 1949–1990.The supplied evidence also equates the German Democratic Republic with East Germany or GDR.
- Uniqueness check: The post-1977 sweep reports no other listed nationality among the reviewed recipients as a defunct country.The reviewed countries include Chile, Japan, Brazil, Austria, Venezuela, Taiwan, the United States, Belarus, and South Korea.
- Recipient and nationality analysis: Claus Peter Flor is the sole post-1977 candidate identified with a nationality potentially corresponding to a defunct country.The investigation maps his 1983 entry to Germany and considers East Germany as the relevant historical interpretation.