SearchArt: Training Long-Horizon Search Agent with Scalable Synthetic and Verified Task

TL;DR

SearchArt is a long-horizon search-agent training framework from Renmin University + Alibaba that pairs verification-driven task synthesis with a multi-stage SFT + RL pipeline on a Qwen3.5-27B backbone. Complex information-seeking QA pairs and matching search trajectories are synthesized from web documents and auto-generated evidence graphs, then filtered by a verifier that jointly evaluates QA consistency, trajectory quality, and evidence relevance. The resulting 27B agent reports 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on Deepresearch-bench — claimed to match or surpass frontier closed-source deep-research agents while sitting between the 30B-A3B and 72B open recipes already on the wiki.

Key claims

Task synthesis constructs large-scale datasets for complex search-, research-, and user-oriented tasks by generating diverse information-seeking QA pairs and matching search trajectories from web documents and automatically generated evidence graphs [Abstract].
Data reliability is enforced by a verification pipeline that jointly evaluates three axes: QA consistency, trajectory quality, and relevance of the retrieved evidence [Abstract].
Post-training is a multi-stage pipeline of supervised fine-tuning on verified trajectories followed by reinforcement-learning-based policy optimization [Abstract].
Trained agents exhibit adaptive search planning, iterative evidence aggregation, and complex reasoning over extended interaction horizons [Abstract].
With only a Qwen3.5-27B backbone, SearchArt scores 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on Deepresearch-bench [Abstract].
Claimed to match or surpass frontier closed-source agents on both deepsearch and deepresearch benchmarks at 27B active parameters [Abstract].

Method

The framework has three components. The task-synthesis stage seeds tasks from web documents and automatically constructed evidence graphs to produce diverse information-seeking QA pairs paired with candidate search trajectories — this is the “scalable” leg of the title. The verification pipeline then filters both the QA pairs and their trajectories along three dimensions simultaneously: whether the answer is consistent with the question given the graph, whether the trajectory is a plausible way to arrive at the answer, and whether the retrieved evidence is relevant. Only jointly-verified data flows into training. The training pipeline is SFT on the verified trajectories followed by RL-based policy optimization, targeted at long-horizon adaptive planning, iterative evidence aggregation, and complex multi-step reasoning. Base model is Qwen3.5-27B.

Results

Headline numbers (Qwen3.5-27B, single-model): 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, 52.55 on Deepresearch-bench [Abstract]. For rough context on the wiki, REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents at 30B-A3B reports BrowseComp 42.1 / 57.4-with-CM and BrowseComp-ZH 49.8 / 58.2-with-CM, Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design at 8B and OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis at 30B-A3B sit lower, and MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling at 72B reports 47.1 on BrowseComp — so SearchArt’s abstract BrowseComp number is substantially above the filed open-agent frontier at any active-parameter size. No per-component ablations, wall-clock or interaction-depth numbers are in the abstract.

Why it’s interesting

SearchArt is the sixth filed deep-research agent recipe alongside MiroThinker, Marco-DeepResearch, SMTL, Tongyi, OpenResearcher, and REDSearcher, and it stakes out a distinct point in the design space: a dense Qwen3.5-27B rather than a 30B-A3B MoE, which changes the parameter-per-active-parameter trade the 30B-A3B papers (REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents, Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization, OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis) all sit at. Its three-axis jointly-verified filter (QA consistency × trajectory quality × evidence relevance) is directly comparable to REDSearcher’s six-step cascaded verifier and Marco-DeepResearch’s verification-at-every-stage — the Synthetic Training Data “overfit-to-your-own-synthesizer” open question is the load-bearing concern all three are pulling on, and SearchArt claims strong out-of-distribution BrowseComp numbers as evidence its verifier survives that failure mode. If the 70.06 BrowseComp number holds up under community re-evaluation, it also intensifies the Tool-Use Agents “interaction depth vs. base capability” question — a 27B dense model apparently beats 30B-A3B recipes and the 72B MiroThinker, suggesting verified-synthesis + RL on a stronger base dominates the interaction-depth scaling axis for this class of benchmarks.