REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents

Our Pipeline

Figure 2. QA Generation. We construct directed acyclic graphs (DAGs) from Knowledge-Graph entities and Web-Walk hyperlinks, enabling explicit difficulty control via topological complexity. Each node is enriched with multi-source evidence, then sampled to form reasoning paths with Query Fuzzing (entity/attribute anonymization) to increase search challenge. Verifier Pipeline. A cascaded filter progressively validates quality: LLM difficulty check, QA-graph alignment, Google retrieval verification, hallucination detection, agent rollout confirmation, and answer uniqueness validation—producing 20K+ verified trajectories at >85% accuracy.

Performance

Comparison between REDSearcher and closed / open agentic models

Model	BrowseComp	BrowseComp-zh	GAIA	HLE	Overall
Proprietary Deep Research Agents
Seed–1.8	67.6	81.3	87.4	40.9	69.3
Gemini–2.5–pro–DR	7.6	27.3	-	-	-
Gemini–3–Pro	37.8	51.6	74.8	45.8	52.5
Claude–4.5–sonnet	24.1	42.4	66.0	32.0	41.1
OpenAI–o3	49.7	58.1	70.5	20.2	49.6
GPT–5–Thinking–high	54.9	63.0	76.7	41.7	59.1
GPT–5.2–Thinking–xhigh	65.8	76.1	-	-	-
Open-source Deep Research Agents
Kimi–K2.5–Agent	60.6 / 74.9*	-	-	50.2	-
GLM–4.7	52.0 / 66.6*	- / 67.5*	-	42.8	-
DeepSeek–V3.2	51.4 / 67.6*	- / 65.0*	-	40.8	-
LongCat–Flash–Thinking	56.6 / 73.1*	69.0 / 77.7*	-	-	-
Open-source 30B–A3B Agents
WebResearcher–30B	37.3	45.2	-	28.8	-
WebSailorV2–30B	35.3	44.1	74.1	30.6	46.0
Tongyi DeepResearch–30B	43.4	46.7	70.9	32.9	48.5
GLM–4.7–Flash	42.8	-	-	-	-
REDSearcher	42.1 / 57.4*	49.8 / 58.2*	80.1	33.3	51.3

* Results with Context Management (CM). Best results in bold.

Main results on multimodal search benchmarks

Model	MM-BC	BC-VL	MMS+	MMS	LiveVQA	HLE-T	HLE-VL	BC	BC-ZH
Proprietary Deep Research Agents
Gemini-2.5-Flash	5.6	44.6	19.9	64.0	73.0	-	-	-	-
Gemini-2.5-Pro	7.1	49.9	22.2	69.0	76.0	-	-	7.6	27.3
Seed1.8	46.3	-	-	-	-	40.9	31.5	67.6	81.3
Seed1.8^†	21.4	54.1	11.0	69.7	62.4	-	-	-	-
GPT-5	-	46.1	17.2	63.7	73.3	41.7	-	54.9	63.0
Gemini-3-Pro^†	28.5	56.4	38.1	73.0	79.9	45.8*	36.0*	37.8*	51.6*
Multimodal Agent Flow
Qwen2.5-VL	1.8	10.2	-	29.2	35.7	-	4.9	-	-
Qwen3-VL (30B)	10.7	37.1	11.0	59.7	64.8	8.8	8.7	0.2	7.2
Qwen3-VL (235B)	12.1	43.1	17.4	63.3	70.2	14.5	14.1	0.3	18.6
Multimodal DeepResearch Agent
MMSearch-R1	-	-	-	53.8	48.4	-	-	-	-
WebWatcher	-	27.0	-	55.3	58.7	-	13.6	-	-
DeepEyesV2	-	-	-	63.7	-	-	-	-	-
Vision-DeepResearch	-	53.7	28.5	69.6	77.6	-	-	-	-
REDSearcher-MM-SFT	25.3	55.3	20.2	70.3	78.5	24.4	24.2	30.1	43.1
REDSearcher-MM-RL	23.5	57.2	26.6	72.9	79.3	25.3	25.6	31.2	44.5

† denotes results evaluated using the same evaluation tools as ours, and * denotes results taken from the original papers.

BibTeX

@article{redsearcher2026,
  title={REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents},
  author={Zheng Chu and Xiao Wang and Jack Hong and Huiming Fan and Yuqi Huang and Yue Yang and Guohai Xu and Shengchao Hu and Dongdong Kuang and Chenxiao Zhao and Cheng Xiang and Ming Liu and Bing Qin and Xing Yu},
  journal={arXiv preprint arXiv:2602.14234},
  url={https://arxiv.org/pdf/2602.14234},
  year={2026}
}

REDSearcher: A Scalable and Cost-EfficientFramework for Long-Horizon Search Agents

Our Pipeline

Performance

BibTeX

REDSearcher: A Scalable and Cost-Efficient
Framework for Long-Horizon Search Agents