Research map / Large language models / Agents, code and math
DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents
Mingxuan Du, Benfeng Xu, Chiwei Zhu, Licheng Zhang, Xiaorui Wang, Zhendong Mao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use reference-based scoring (add reference report) in RACE (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 66.56 -> 71.33 | Improved |
| use RACE instead of vanilla direct-scoring prompt (ablation) | Gemini-2.5-pro judge · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 58.89 -> 71.33 | Improved |
| add adaptive task-specific criteria instead of static criteria (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 68.33 -> 71.33 | Improved |
| add dynamic criteria weights (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 70.67 -> 71.33 | Improved |
| add dynamic dimension weights (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 70.89 -> 71.33 | Improved |
| keep reference as second position (vs swapping reference and target positions) (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 69.56 -> 71.33 | Improved |
| use Gemini 2.5 Pro Preview as judge instead of o3 (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%): 68.11 -> 71.33 | Improved |
| use cheaper o4-mini as judge instead of Gemini 2.5 Pro Preview (ablation) | RACE · Human consistency (50 Chinese tasks) | pairwise agreement rate (%) / cost: 71.33 -> 70.89; cost 0.13 -> 0.04 | Mixed |
| deep research agent design (vs LLM with search tools) for report quality | Gemini-2.5-Pro · DeepResearch Bench (RACE) | RACE overall score: 35.12 -> 48.88 | Improved |
| allow multi-turn web searches for Claude-3.7 vs single-turn search models | Claude-3.7-Sonnet w/Search · DeepResearch Bench (RACE) | RACE overall score: 28.48 -> 40.67 | Improved |
Closest papers by meaning
- Deep Research Arena: The First Exam of LLMs’ Research Abilities via Seminar-Grounded Tasks (AAAI 2026 · similarity 0.91)
- SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery (NeurIPS 2026 · similarity 0.90)
- From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (NeurIPS 2026 · similarity 0.90)
- MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (NeurIPS 2026 · similarity 0.89)
- DRBench: A Realistic Benchmark for Enterprise Deep Research (ICLR 2026 · similarity 0.89)
- WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
