As agentic foundation models continue to evolve, how to further improve their performance in vertical domains has become an important challenge. To this end, building upon Tongyi DeepResearch, a powerful agentic foundation model, we focus on the Chinese medical deep search scenario and propose QuarkMedSearch, systematically exploring a full-pipeline approach spanning medical multi-hop data construction, training strategies, and evaluation benchmarks to further push and assess its performance upper bound in vertical domains. Specifically, for data synthesis, to address the scarcity of deep search training data in the medical domain, we combine a large-scale medical knowledge graph with real-time online exploration to construct long-horizon medical deep search training data; for post-training, we adopt a two-stage SFT and RL training strategy that progressively enhances the model's planning, tool invocation, and reflection capabilities required for deep search, while maintaining search efficiency; for evaluation, we collaborate with medical experts to construct the QuarkMedSearch Benchmark through rigorous manual verification. Experimental results demonstrate that QuarkMedSearch achieves state-of-the-art performance among open-source models of comparable scale on the QuarkMedSearch Benchmark, while also maintaining strong competitiveness on general benchmarks.
Figures & tables
Figure 1: Medical Deep Search Task Synthesis Pipeline.
Figure 2: Long-Horizon Deep Search QA Augmentation in Online Environments.
Figure 3: Comparison of the distribution of tool-call counts between QuarkMedSearch and BrowseComp.
Figure 4: Post-training stages for QuarkMedSearch.
Dimension
Scoring Criteria
Intent Planning
Accuracy of problem understanding and solution path planning; requires clear decomposition of complex problems.
Tool Invocation
Parameter accuracy, query construction quality, and fault tolerance; requires precise and efficient invocation without redundancy.
Information Integration
Extraction of key information, strategy adjustment upon failed retrieval, and (for medical QA) terminology standardization, source authority and timeliness.
Factual Rigor
Hard constraint: any factual error results in a score of 0.
Table 1: Multi-Dimensional Rubrics for Trajectory Evaluation
Figure 5: Distribution of Sample Categories in the QuarkMedSearch Benchmark.
Category
Key Entity Coverage
Biomedical Fundamentals
Anatomical structures, biochemical molecules, pathophysiological mechanisms, pathogenic microorganisms, etc.
Drugs and Medical Products
Chemical drugs, active ingredients of herbal medicine, biologics, vaccines, medical devices, etc.
Medical Research and Knowledge
Clinical trial findings, evidence-based medicine, medical statistics, diagnostic and treatment technical standards, etc.
Diseases and Clinical Manifestations
Disease diagnosis and classification, clinical syndromes, symptoms and signs, key differential diagnosis points, etc.
Clinical Procedures
Clinical laboratory tests, surgical procedures, treatment protocols, medication strategies and interventions, etc.
Medical Institutions
Hospitals, pharmaceutical companies, medical device companies, medical academic associations, health regulatory agencies, etc.
Table 2: Category Taxonomy of the QuarkMedSearch Benchmark
Backbone
Size
QuarkMedSearch Benchmark
Large-scale Models
Seed1.8
-
55.71
Kimi-K2.5
1T-A32B
56.74
GLM-5
744B-A40B
58.57
DeepSeek-V3.2
671B-A37B
57.14
Qwen3.5-Plus
397B-A17B
51.42
Table 3: Results on the QuarkMedSearch Benchmark.
Backbone
Size
BrowseComp-EN
BrowseComp-ZH
Xbench DeepSearch
Commercial Closed-source Models
Seed1.8
-
67.6
81.3
-
Gemini-3-Pro
-
59.2
66.8
-
Claude Opus 4.5
-
57.8
62.3
-
GPT-5.2 (xhigh)
-
65.8
76.1
-
Open-source Large-scale Models (w/o ctm)
Table 4: Evaluation results on general deep search benchmarks. For commercial closed-source models, we directly adopt the best results from official reports. For open-source models, we report only the results without context management (w/o ctm) to ensure fair comparison, and include only those models that explicitly provide such results in the comparison among models of similar parameter scale. † denotes results obtained under our unified evaluation environment. Bold indicates the best result among models of similar parameter scale.
Stage
QuarkMedSearch Benchmark
BrowseComp-EN
BrowseComp-ZH
Base(Tongyi DeepResearch)
40.71
42.67
45.40
+ 32K Stage I
49.99
43.23
49.40
+ 128K Stage II
52.21
45.47
55.36
Table 5: Ablation study results of the two-stage SFT training strategy.
Figure 6: RL training dynamics: response length and reward over training steps.
Metric
BrowseComp-ZH
BrowseComp-EN
w/o RL
w/ RL
w/o RL
w/ RL
Termination Rate ↑
0.676
0.725
0.580
0.635
Avg. Tool Calls (Correct) ↓
28.09
24.01
40.03
37.12
Table 6: Comparison of termination rate and average number of tool call rounds required to correctly answer questions before and after RL training. w/o RL and w/ RL correspond to the model states at the end of two-stage SFT and after RL convergence, respectively. ↑ means higher is better, ↓ means lower is better.
Figure 7: Training dynamics of correct score and format score under two reward strategies: correct score + format score (blue) and strict correct score (red).
Method
BrowseComp-EN
BrowseComp-ZH
Large-scale Open-source Models (w/ ctm)
Kimi-K2.5
74.9
-
GLM-5
75.9
72.7
Qwen3.5-Plus
78.6
70.3
DeepSeek-V3.2
67.6
65.0
Open-source Models of Comparable Scale (w/ ctm)
Table 7: Results with ctm on BrowseComp.
Figure 8: Win/Tie/Lose comparison between the fused model and the non-fused baseline, as evaluated by professional medical experts on medical Deep Research report generation.
Recent advances in large language models (LLMs) have enabled search agents to autonomously tackle complex tasks across extended search and reasoning horizons. However, training effective search agents remains challenging due to the lack of scalable and long-horizon tasks, and the difficulty of evaluating and correcting intermediate reasoning and tool-use behaviors. We introduce SearchArt, a scalable framework for training long-horizon search agents through verification-driven task synthesis and a multi-stage post-training pipeline. SearchArt constructs large-scale datasets for complex search-, research- and user-oriented tasks by synthesizing diverse information-seeking QA pairs and corresponding search trajectories from web documents and automatically generated evidence graphs. To ensure the reliability of the synthesized data, we design a verification pipeline that jointly evaluates QA consistency, trajectory quality, and the relevance of retrieved evidence. The verified trajectories are subsequently used in a multi-stage training process comprising supervised fine-tuning and reinforcement learning-based policy optimization. Search agents trained with SearchArt exhibit adaptive search planning, iterative evidence aggregation, and complex reasoning over extended interaction horizons. Experimental results demonstrate that, with only (Qwen3.5-) 27B parameters, SearchArt scores 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on Deepresearch-bench, matching or surpassing frontier closed-source agents on both deepsearch and deepresearch benchmarks.
Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark built on UK Biobank, comprising 50,401 clinical questions derived from real-world ICD-10-coded diagnoses of 4,739 participants. Each question links to a patient-specific environment containing clinical context and multi-sequence MRI from baseline and follow-up visits, where agents autonomously select which visits, organs, modalities, slices, and specialist tools to inspect and compare. Supervised fine-tuning transfers evidence-seeking workflows from 147,434 frontier-model interaction trajectories, followed by agentic reinforcement learning on the learner's own environment interactions. Privileged on-policy self-distillation and rubric-based LLM feedback refine evidence-to-conclusion reasoning without prescribing tool sequences. Experiments show that CASE moves beyond question-answer imitation toward transferable investigation policies, strengthening evidence-grounded longitudinal reasoning. Under matched evaluation conditions, our Qwen3-VL-8B-Instruct based agent achieves over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8. Project page available at: https://vinyehshaw.github.io/case-project/.
Minye Shao, Chaohui Yu, Yixuan Wu +6
Durham University · DAMO Academy, Alibaba Group · Hupan Laboratory +1
Deep search agents tackle complex knowledge tasks through iterative retrieval, multi-hop reasoning, and evidence synthesis across multiple sources. Existing approaches typically assume relatively stable retrieval systems and operate over short-horizon tool interaction. However, when retrieval is sensitive to query formulation, even a semantically appropriate query may fail to surface critical evidence because of mismatched entity names, aliases, or keyword combinations. Recovering from such failures requires repeated query reformulation and longer interaction trajectories. This setting poses a distinct training challenge, as the policy must sustain long-horizon query exploration while managing an expanding volume of retrieved content. We propose LexiHorizon, a framework for training search agents over long horizons that expands the trajectory context budget, manages accumulated retrieval content using a window over recent tool observations while preserving the reasoning history, and introduces an outcome-gated search-effort reward that provides a bounded bonus for tool invocations to trajectories with nonzero answer reward. Experiments on XBench, WebWalkerQA, and BrowseComp-ZH show that the resulting 9B model consistently outperforms both its base model and MiroThinker-1.7-mini, with maximum absolute gains of 8.7 and 23.8 percentage points, respectively. These results suggest that combining an extended context budget with reasoning-preserving context management benefits long-horizon deep search agents.