Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization

Abstract: Recent deep research agents primarily improve performance by scaling reasoning depth, but this leads to high inference cost and latency in search-intensive scenarios. Moreover, generalization across heterogeneous research settings remains challenging. In this work, we propose \emph{Search More, Think Less} (SMTL), a framework for long-horizon agentic search that targets both efficiency and generalization. SMTL replaces sequential reasoning with parallel evidence acquisition, enabling efficient context management under constrained context budgets. To support generalization across task types, we further introduce a unified data synthesis pipeline that constructs search tasks spanning both deterministic question answering and open-ended research scenarios with task appropriate evaluation metrics. We train an end-to-end agent using supervised fine-tuning and reinforcement learning, achieving strong and often state of the art performance across benchmarks including BrowseComp (48.6\%), GAIA (75.7\%), Xbench (82.0\%), and DeepResearch Bench (45.9\%). Compared to Mirothinker-v1.0, SMTL with maximum 100 interaction steps reduces the average number of reasoning steps on BrowseComp by 70.7\%, while improving accuracy.

• The paper presents a parallel search paradigm (SMTL) that reduces reasoning steps by up to 70.7% while enhancing accuracy for long-horizon tasks.
• It introduces structured context resets and a unified data synthesis pipeline to generalize across both deterministic and open-ended research problems.
• Empirical results show SMTL outperforming mainstream models on benchmarks like BrowseComp and DeepResearch, achieving state-of-the-art performance.

Rethinking Long-Horizon Agentic Search: The Search More, Think Less (SMTL) Paradigm

Introduction and Motivation

Long-horizon, tool-augmented language agents frequently improve accuracy by deepening sequential reasoning and tool call chains. However, such approaches incur prohibitive inference cost and latency, especially in web-based or open-ended research tasks requiring extensive environmental interaction. Existing agents also demonstrate limited generalization across heterogeneous objectives—deterministic question answering (QA) versus open-ended synthesis—partly because datasets and supervision pipelines remain highly specialized and context-inefficient. The "Search More, Think Less" (SMTL) framework (2602.22675) advances the agentic search paradigm by advocating efficiency-centric, parallelized evidence-seeking, with structured context management and unified data synthesis for versatile generalization.

Parallel Agentic Workflow for Efficient Long-Horizon Search

SMTL departs from the conventional, strictly sequential agent design by explicitly decomposing complex tasks into multiple, parallelizable subtasks. The agent constructs an initial plan, executes concurrent tool-based actions across independent subgoals, and synchronizes intermediate results via periodic plan refinements. This model prioritizes information density per interaction step and minimizes redundant reformulation cycles characteristic of single-token or single-step tool invocation.

Figure 1: Overview of the parallel agentic workflow enabling concurrent subtask execution and periodic plan refinement.

Parallel execution is realized through batching web search and focused page crawling, coupled with plan-driven context resets that efficiently operate under constrained context budgets (e.g., 128K token windows). The agent periodically refines its plan and, if the context budget overflows, prunes irrelevant context to preserve critical subgoal structures. Such structured overflow handling markedly extends effective trajectory length—crucial for long-horizon environments like BrowseComp—without sacrificing performance consistency.

Data Construction: Generalization Across Search and Research Tasks

SMTL introduces an automated synthesis pipeline generating both deterministic search and open-ended research problems grounded in subgraph structures constructed from real-world web trajectories. The process involves:

• Corpus collection spanning diverse domains, with inter-document relations extracted for multi-hop, fact-dense task support.
• Construction of knowledge graphs using LLM-driven entity/attribute extraction, followed by controlled subgraph sampling for variable task difficulty.
• Hierarchical question synthesis for deterministic QA and open-ended synthesis, with strict LLM-based verification and obfuscation to avoid premature answer leakage.

  Figure 2: Schematic of the data construction pipeline integrating subgraph extraction, hierarchical question synthesis, and multi-type supervision.

A two-stage quality filter—rule-based rejection and LLM-as-judge semantic check—enforces both structural validity and semantic rigor for all supervised trajectories. This unified data approach allows a single agent to generalize efficiently across disparate evaluation regimes.

Training Protocol: SFT and Reinforcement Learning with RLOO

The SMTL agent is post-trained via supervised fine-tuning (SFT) on carefully curated search and research datasets, applying rigorous context and trajectory efficiency constraints. Reinforcement learning refines the agent's behavior using a modified REINFORCE Leave-One-Out (RLOO) method with token-level loss, rollout correction, and trajectory filtering targeting stability and outcome-aligned tool use.

Evaluation employs LLM-as-judge scoring for both deterministic answers and report-style research synthesis, ensuring nuanced and task-relevant agent assessment.

Empirical Results: Efficiency, Accuracy, and Generalization

On major benchmarks including BrowseComp, GAIA, XBench, and DeepResearch Bench, SMTL establishes leading or state-of-the-art performance relative to foundation models, proprietary research systems, and top open-source agents. Notable findings include:

• BrowseComp: SMTL-300 achieves 48.6% accuracy, surpassing all prior 30B-scale open-source models, and outperforms MiroThinker-v1.0-30B with 70.7% fewer reasoning steps at higher accuracy.
• SMTL consistently lies on the Pareto frontier of accuracy vs. interaction steps—no mainstream baseline achieves comparable accuracy at lower or equal interaction cost.

Figure 3: SMTL lies on the Pareto frontier in accuracy vs. interaction steps on BrowseComp.

• DeepResearch Bench RACE: SMTL-100 delivers 45.9% overall (with high balance across coverage, depth, compliance, and readability), slightly outperforming Tongyi-DeepResearch-30B while exceeding all other open 30B-class baselines.

Cross-benchmark analysis demonstrates that gains are robust to variations in task type and interaction budget—SMTL's efficiency advantage is especially apparent for deep or hard tasks, where traditional sequential agents exhaust their step budget without locating optimal evidence.

Analytical Case Study

A detailed case study juxtaposes SMTL-30B and MiroThinker-v1.0-30B on a representative BrowseComp instance. SMTL rapidly partitions the task into parallel evidence streams, localizing the needed entity in just 8 assistant turns and producing the final answer with minimal redundant work. Conversely, MiroThinker-v1.0-30B executes narrow, sequential steps, requiring 16 turns merely to accumulate equivalent evidence, and ultimately 150 turns for final verification.

Ablations and Scaling Properties

Ablation studies confirm that increasing the interaction budget primarily benefits unsuccessful cases, as success trajectory lengths plateau well below the maximum allowed steps; most failures transition to longer but more focused exploratory behaviors. Increasing web search top-$k$ also provides monotonic but diminishing returns, evidencing the advantage of broad search breadth over deeper sequential reasoning in long-horizon agentic contexts.

Figure 4: BC score as a function of the maximum allowed steps, illustrating rapid performance gains with extended interaction horizons before saturating.

Practical and Theoretical Implications

SMTL’s parallel, search-first approach marks a decisive shift from computation-heavy, depth-centric agent scaling strategies. This reframing suggests that future research agents should prioritize concurrent evidence acquisition and plan-driven information density, especially as context windows remain bottlenecked. By demonstrating strong cross-task transfer with a unified workflow and data synthesis regime, SMTL paves the way for generalist agents that scale more by search width and data/topology diversity than by sheer reasoning depth.

Practically, the SMTL framework is well-suited for real-world deployments constrained by inference latency and API cost, as its interaction patterns yield both improved generalization and sharply reduced wall-clock time per task. The architectural principles underpinning SMTL—parallel execution, structured context management, and overflow-centric plan updates—are readily extensible to multi-agent, multimodal, or cross-environment scenarios.

Conclusion

SMTL advances long-horizon agentic search by structurally decoupling evidence acquisition from depth-limited, sequential reasoning. Its parallel workflow, unified data pipeline, and efficient context management yield strong empirical results, robust task generalization, and efficient scaling under interaction or context constraints. This framework advocates a transition toward search-density–maximizing design in the next generation of research agents, with important implications for how AI systems pursue knowledge synthesis in increasingly open-ended and resource-limited environments.