Research map / Large language models / Agents, code and math
Characterizing Deep Research: A Benchmark and Formal Definition
Abhinav Java, Ashmit Khandelwal, Sukruta Prakash Midigeshi, Aaron Halfaker, Amit Deshpande, Navin Goyal, Ankur Gupta, Nagarajan Natarajan, Amit Sharma
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 dedicated deep research systems instead of reasoning LLMs with web search | OpenAI Deep Research vs o4-mini · LiveDRBench NovelDS Peer Retrieval | F1: 0.345 -> 0.585 | Improved |
| use a deep research system rather than a reasoning model with low (~30 s) reasoning budget | reasoning LLMs with web search · LiveDRBench SciFacts Materials | F1: 0.000 for all reasoning and non-reasoning baselines vs 0.314 (OpenAI DR) | Hurt |
| use reasoning-enabled models instead of non-reasoning models | LLMs with web search · LiveDRBench categories | F1 | Improved |
| require both main claim and subclaim correct (material and paper) rather than either one (ablation) | OpenAI Deep Research · LiveDRBench SciFacts Materials | F1: 0.735 (title), 0.504 (material) -> 0.314 overall | Hurt |
| more sources gathered by the model | Perplexity Deep Research · LiveDRBench NovelDS | F1: over 40 sources, F1 0.41 vs OpenAI 20 sources, 0.55 | No clear effect |
| more branching and backtracking in the search trace | OpenAI Deep Research · LiveDRBench NovelDS | F1: F1 0.55 vs 0.41 and 0.35 | Improved |
| early termination with few sources | Gemini Deep Research · LiveDRBench NovelDS | F1 | Hurt |
| use GPT-4o claim agreement versus manual human verification (ablation) | evaluation pipeline · LiveDRBench | precision/recall/F1 | No clear effect |
Closest papers by meaning
- From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (NeurIPS 2026 · similarity 0.88)
- WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research (ICLR 2026 · similarity 0.88)
- DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence (ICLR 2026 · similarity 0.87)
- MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (NeurIPS 2026 · similarity 0.87)
- Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.86)
- SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery (NeurIPS 2026 · similarity 0.86)
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.
