deep research
A deep research agent is an AI system that can go online, look up documents, read them, and then produce a report with citations. The core idea is that the agent can search for information itself, but this creates a problem: the order in which it finds things can bias its conclusions, and if it stops looking too early, it might miss important evidence. Because it makes its own decisions about what to read and when to stop, the documents it sees are not necessarily a fair sample of all available information.
The papers examine this process from multiple angles. Some show that agents get stuck on early findings or pursue weak paths, and propose fixes like exploring multiple hypotheses in parallel before committing. Others find that even good search can produce reports with errors traced to a single component, or that accuracy drops sharply when questions require long chains of evidence. There is also evidence that training such agents is fragile: they can stall or waste resources, and failures at production scale are structural, not just technical.
- Papers
- 24
- Released code
- 3
- First seen
- Aug 2026
- Latest
- Nov 2026
18 papers in the last two months, against 6 in the two before.
Start here
Three papers to read first, in the order they came out- 1Web research benchmark exposes persistent limits in autonomous search agents
On OpenDS-World’s live-Web tasks, the best of 12 tested models reaches only 39% accuracy.
Oct 2026 - 2Defect-guided fuzzing finds deep research AI failures three times faster
Deep Research systems have nine recurring defects that EMBERS exploits to detect failures at 3.21 times the baseline rate.
Nov 2026 - 3AI reward models struggle to judge research reports; language model judges outperformNov 2026
How fast it is moving
Often studied with
agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.6 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.5 togetherevidence retrievalPulling specific, relevant information from a large corpus to support an answer, decision, or reasoning step.5 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Top Universitycs.AI
AI reward models struggle to judge research reports; language model judges outperform
University of Macau, Monash University · Nov 2026
- Chinese Techcs.AI
Defect-guided fuzzing finds deep research AI failures three times faster
Deep Research systems have nine recurring defects that EMBERS exploits to detect failures at 3.21 times the baseline rate.
Tianjin University, Huawei Technologies Co., Ltd. · Nov 2026
- Big Techcs.IR
Web research benchmark exposes persistent limits in autonomous search agents
On OpenDS-World’s live-Web tasks, the best of 12 tested models reaches only 39% accuracy.
HKUST(GZ), FDU · Oct 2026
- Academiccs.AI
Correcting search bias makes deep research conclusions more representative
CESS corrects evidence selection bias in deep research agents by weighting documents by their selection probabilities.
Beijing Normal University, Ke Holdings · Oct 2026
- Independentcs.CV
Training long-horizon multimodal research agents with reinforcement learning at scale
Long-MDR makes 128k-context, 75-turn multimodal reinforcement learning more efficient and stable through distillation warmup, gradual horizon expansion, and entropy-based recovery.
Oct 2026
- Big Tech
Production-scale AutoResearch exposes five failure modes and a three-principle fix
AutoResearch at production scale reveals five failure modes fixed by a prevent-persist-redirect multi-agent framework.
Amazon · Sept 2026
- Chinese Techcs.CL
Separating planning from synthesis improves long-horizon search agents
IterSynth separates planning and synthesis in deep search agents using a summary state, outperforming prior small agents.
Zhejiang University, Tencent · Sept 2026
- Chinese Techcs.CL
Rubric-based process rewards improve deep research agents’ evidence gathering
Rubric-grounded process credit, which tracks accumulated support per rubric, improves deep research agent performance and enables an 8B model to match proprietary systems.
University of Chinese Academy of Sciences, State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) · Sept 2026
- Top Universitycs.AIcode
Deep research agents struggle with long, multimodal evidence chains
Current deep research agents struggle with sustained, dependency-consistent evidence integration across long multimodal chains.
Mohamed bin Zayed University of Artificial Intelligence, University of Science and Technology of China · Sept 2026
- Top Universitycs.CL
Branching Across Hypotheses Helps Research Agents Avoid Early Search Traps
HypoSearch parallelizes multiple hypotheses before commitment to outperform single-trajectory deep-research agents.
City University of Hong Kong, Meituan · Sept 2026
- Academiccs.CLcode
In AI-Q, most report errors originate in the orchestrator
Most final-report errors in agentic deep research originate at the orchestrator, with simple interventions raising citation recall by 5%.
Bar-Ilan University, UNC Chapel Hill · Aug 2026
- Top Universitycs.CL
Bounded citation-graph search improves recall while cutting scholarly search costs
Crase prunes citation edges whose claims lack entailment support and ranks remaining papers with a recency-aware random walk, outperforming deep research agents at lower cost.
National University of Singapore, TCG CREST · Aug 2026
- Big Techcs.AI
Evidence graphs produce synthetic research tasks with controllable difficulty
GRADS generates synthetic question-answer data with graph-regulated difficulty, producing verified tasks where empirical hardness matches requested hardness and improves agent training.
Johns Hopkins University, Nanjing University · Oct 2026
- Chinese Techcs.AI
Research agents struggle to sustain gains in multimodal model training
Autonomous research agents for multimodal post-training often degrade performance and fail to sustain improvements.
University of Science and Technology of China, Alibaba Token Hub, Alibaba Group · Oct 2026
- Big Techcs.AI
Managing research ideas helps autonomous agents search faster
AIM automates idea-driven research, beating baselines by up to 4.9% and reaching best performance 3.1x faster.
Google Cloud AI Research, University of Wisconsin-Madison · Oct 2026
- Chinese Techcs.CV
Adaptive tool use improves video research agents’ accuracy and efficiency
AdaVDR adapts tool use and reflection in video deep research to cut costs and errors.
Accio Team, Alibaba Group, Beijing Key Laboratory of Intelligent Information Technology · Aug 2026
- Chinese Techcs.LG
Evidence annotations improve step-level credit for deep search agents
Evidence-guided step-level process rewards derived from annotated evidence claims improve credit assignment and outperform standard GRPO on deep search benchmarks.
Renmin University of China, Key Laboratory of Data Engineering and Knowledge Engineering, MOE · Oct 2026
- Academiccs.CL
Dual-system agent couples fast evidence checks with deliberative planning for efficient research
JevGPT couples fast evidence processing with deliberative planning to achieve leading performance, lower cost, and faster completion.
King Abdullah University of Science and Technology · Oct 2026
- Chinese Techcs.CLcode
Adaptive tutoring helps rerankers assemble better evidence sets for research
AdaTutoRank, a setwise reranker with adaptive tutoring hints matched to rollout quality, tops ten RAG and deep research benchmarks while using fewer retrieval calls.
University of Science and Technology of China, Yuanbao Team, Tencent · Sept 2026
- Top Universitycs.AI
Benchmark reveals AI systems struggle to retrieve papers that inspire new research
Current retrieval systems recover at most 48% of catalyst papers that authors credit.
Stanford University, Seoul National University · Oct 2026
- Chinese Techcs.AI
Research agents can reuse experience to avoid unnecessary experiments
Experimental Experience Modeling improves autonomous research by reusing accumulated experimental experience and acquiring new experience only when needed.
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · Oct 2026
- Top Universitycs.CL
Specialized AI agents improve clinical trial design risk assessment
A multi-agent system that coordinates specialized AI agents to detect clinical trial design flaws and predict regulatory success matches or beats current methods.
University of Illinois Urbana-Champaign, Keiji AI Inc · Sept 2026
- Independentcs.AI
AI agents fail end-to-end physics research due to shallow reasoning and missing intuition
An end-to-end physics research benchmark finds frontier AI agents fail most tasks, with main weaknesses insufficiently rigorous reasoning and limited expert intuition.
Nov 2026
- Top Universitycs.CL
Generating Agent Harnesses on Demand Improves Models Across Benchmarks
JIT-Agent trains a meta-model to generate task-specific agent harnesses on the fly, improving off-the-shelf LLMs' agent performance.
LV-NUS Lab · Aug 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.