[Preprint 2026] Search More, Think Less (SMTL): Rethinking Agentic Efficiency via Parallelism

TL;DR

The "Deep Research" trend in AI agents is currently obsessed with "Thinking Harder"—scaling test-time compute through deeper sequential reasoning. OPPO's SMTL turns this on its head by arguing we should "Search More, Think Less." By replacing linear reasoning chains with parallel subtask execution and a graph-grounded data pipeline, SMTL smashes SOTA benchmarks while cutting interaction steps by over 70%.

The "Linear Reasoning" Bottleneck

Modern research agents (like OpenAI Deep Research or MiroThinker) operate on a sequential loop: Reason -> Act -> Observe -> Repeat. While effective, this creates a massive latency bottleneck. In search-intensive tasks, if an agent spends 20 turns linearly clicking through URLs to find one fact, the inference cost and context window bloat become unsustainable.

The core insight of SMTL is that information acquisition is often parallelizable. Most complex research queries can be decomposed into independent sub-queries that don't need to wait for each other.

Methodology: High-Concurrency Agentic Workflow

SMTL introduces a radical shift in how agents interact with the web. Instead of one tool call per turn, it optimizes for information density.

1. Parallel Agentic Workflow

The system initializes by decomposing a composite task into a Directed Acyclic Graph (DAG) of subtasks. At each timestep, the agent selects all "ready" subtasks and executes them concurrently using tools like web_search and crawl_page.

2. Plan-Centric Context Management

To handle the massive amount of data returned by parallel searches, SMTL uses a "Periodic Plan Refinement" strategy. Every 5 steps, it summarizes progress and refreshes the plan. If the 128K context window is hit, it performs a forced plan-centric reset, dropping old history but keeping the current refined execution state—effectively allowing "infinite" horizons within a fixed window.

3. Unified Data Synthesis

The authors didn't just build a better engine; they built a better fuel. They used LightRAG to build massive knowledge graphs from diverse corpora and then performed "Controlled Random Walks" to extract subgraphs. These subgraphs were then turned into:

• Deep Search Data: Deterministic QA requiring multi-hop verification.
• Deep Research Data: Open-ended questions requiring synthesis of the entire subgraph.

Experimental Results: The Pareto Frontier of Efficiency

SMTL doesn't just win on accuracy; it wins on efficiency.

• BrowseComp SOTA: Attained 48.6% accuracy (SMTL-300), significantly outperforming open-source weights like MiroThinker.
• Step Reduction: SMTL-100 achieves higher accuracy than MiroThinker while requiring only 60.4 steps vs. 206 steps.
• Generalization: Unlike specialized agents, SMTL performs equally well on GAIA (fact-finding) and DeepResearch Bench (report writing).

Figure: SMTL consistently stays on the Pareto frontier, providing more "bang for the buck" per interaction step.

Deep Insight: Breadth vs. Depth

A fascinating ablation study in the paper reveals that Retrieval Top-K (Search Breadth) is a more powerful scaling dimension than Reasoning Horizon (Depth). By increasing the number of URLs returned per search, SMTL packs more evidence into each step, allowing the agent to "see more" and "guess less."

Critical Analysis & Conclusion

SMTL proves that the next frontier for AI agents isn't just "more layers" or "more parameters," but better orchestration. By treating research as a parallel data-gathering mission rather than a singular philosophical journey, SMTL sets a new standard for practical, high-speed research agents.

Limitations: The reliance on an LLM-as-a-Judge for open-ended tasks remains a potential source of bias, though the authors' two-stage filtering process (hard rules + semantic LLM evaluation) is a robust step toward quality control.

Takeaway: If you want your agent to solve 100-step problems, don't teach it to think 100 times—teach it to search 10 places at once.