Research map / Large language models / Reasoning and chain of thought

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

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

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 SMTL, a long-horizon agentic search framework with parallel task decomposition, plan-driven context management, and a graph-based synthetic data pipeline covering deep search and deep research, trained with SFT then RLOO on Qwen3-30B-A3B.

Problem

Deep research agents gain accuracy by scaling sequential reasoning depth and tool calls, which raises inference cost and latency; agents trained for deterministic QA often fail to generalize to open-ended research evaluation.

Approach

Replace sequential reasoning with parallel subtask decomposition and concurrent tool calls, periodic plan refinement with plan-centric context reset, and a high-density subgraph-based data pipeline that synthesizes both deterministic deep search QA and open-ended deep research questions.

Key results

• SMTL-300 reaches 48.6% pass@1 on BrowseComp, and SMTL-100 reaches 43.6% with 60.4 average steps versus 206.0 for MiroThinker-v1.0-30B at 41.2%.
• SMTL-100 achieves 45.9 overall on DeepResearch Bench RACE, above Tongyi-DeepResearch-30B (45.7) and Kimi-Researcher (44.6).
• The paper reports a 78% reduction in reasoning steps and up to 2.6x lower inference latency relative to baselines.

Surprising findings

• Most failed BrowseComp trajectories terminate exactly at the maximum step budget, so failures are mostly budget exhaustion rather than early wrong answers.
• Increasing retrieval top-k from 4 to 8 gives large BrowseComp gains, suggesting search breadth is a more efficient scaling axis than reasoning length.

Limitations

• Relies on LLM-as-judge for both RL rewards and evaluation, with unnamed judge model.
• Results compare against baselines under differing default settings and step budgets, complicating fair comparison.
• Deep research evaluation only uses DeepResearch Bench RACE; the RL stage covers only deterministic correctness reward.

Experimental setup

Datasets: TaskCraft corpus trajectories BrowseComp GAIA XBench-DeepSearch WebWalkerQA FRAMES SEAL-0 DeepResearch Bench

Benchmarks: BrowseComp GAIA XBench-DeepSearch WebWalkerQA FRAMES SEAL-0 DeepResearch Bench RACE

Metrics: pass@1 accuracy RACE overall score comprehensiveness insight/depth instruction following readability steps tool calls per step

Baselines: Tongyi-DeepResearch-30B MiroThinker-v1.0-30B WebSailor-32B WebDancer-QwQ WebShaper-32B AFM-32B-RL DeepMiner-32B-RL GPT-5 Claude-4.5-Sonnet Gemini-2.5-Pro OpenAI DeepResearch Gemini DeepResearch

Builds on: Flash-Searcher TaskCraft LightRAG RLOO DAPO

Closest papers by meaning

• Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.90)
• Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents (ICLR 2026 · similarity 0.89)
• Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks (NeurIPS 2026 · similarity 0.89)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
• HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.89)
• DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.89)