The paper proposes Multi-Head Recurrent Memory (MHM) with a select-then-update strategy to improve memory retention in long-context recurrent agents.
hub Canonical reference
Deep Research Agents: A Systematic Examination And Roadmap, September 2025
Canonical reference. 100% of citing Pith papers cite this work as background.
26
Pith papers citing it
Background
100% of classified citations
abstract
The rapid progress of Large Language Models (LLMs) has given rise to a new category of autonomous AI systems, referred to as Deep Research (DR) agents. These agents are designed to tackle complex, multi-turn informational research tasks by leveraging a combination of dynamic reasoning, adaptive long-horizon planning, multi-hop information retrieval, iterative tool use, and the generation of structured analytical reports. In this paper, we conduct a detailed analysis of the foundational technologies and architectural components that constitute Deep Research agents. We begin by reviewing information acquisition strategies, contrasting API-based retrieval methods with browser-based exploration. We then examine modular tool-use frameworks, including code execution, multimodal input processing, and the integration of Model Context Protocols (MCPs) to support extensibility and ecosystem development. To systematize existing approaches, we propose a taxonomy that differentiates between static and dynamic workflows, and we classify agent architectures based on planning strategies and agent composition, including single-agent and multi-agent configurations. We also provide a critical evaluation of current benchmarks, highlighting key limitations such as restricted access to external knowledge, sequential execution inefficiencies, and misalignment between evaluation metrics and the practical objectives of DR agents. Finally, we outline open challenges and promising directions for future research. A curated and continuously updated repository of DR agent research is available at: {https://github.com/ai-agents-2030/awesome-deep-research-agent}.
hub tools
citation-role summary
background 6
citation-polarity summary
roles
background 5polarities
background 5representative citing papers
HarnessBridge is a learnable bidirectional controller that projects observations and actions for LLM agent harnesses, matching specialized performance while cutting tokens and generalizing across model sizes.
PropGuard is a propagation-aware framework for LLM-MAS that constructs dual-view spatio-temporal graphs, employs a GE-GRPO inspector to recover suspicious subgraphs, and applies source-guided remediation to lower attack success while preserving task performance.
IGRPO allocates multi-turn LLM agent rollout budget proportional to intermediate-state information gain, inducing an exponentially tilted teacher distribution for policy optimization.
Strong medical-model task scores do not guarantee process stability: omission, contradiction, and delay mainly break contradiction detection, diagnosis updating, and hallucination self-correction while final evidence grounding can look fine.
MCompassRAG adds topic metadata to chunk representations and uses LLM distillation to train a lightweight topic-aware retriever, reporting 8.24% average information efficiency gain and over 5x lower latency than strong baselines across six benchmarks.
A design space for reading augmentation that contrasts transmission-oriented 'reading to discard' with transformation-oriented 'creative reading' and identifies two opportunities for future systems.
Helicase proposes an autonomous multi-agent LLM framework for uncertainty-guided supply chain knowledge graph construction evaluated on the new SCQA benchmark of 80 queries.
Malicious actors could use AI agents to submit large numbers of fake papers, inflating the submission count and thereby raising the acceptance odds for a small set of chosen legitimate papers under stable conference acceptance rates.
Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to cut up to 50% wasted reasoning tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance cost.
AgentEconomist is an end-to-end agentic system with idea development, experimental design, and execution stages that uses a large economics paper database to produce research ideas with better literature grounding, novelty, and insight than generic LLMs.
Multi-agent deep research systems self-optimize prompts through self-play to match or outperform expert-crafted versions.
Retrievers trained on agent trajectories via the LRAT framework improve evidence recall, task success, and efficiency in agentic search benchmarks.
An LLM agent integrated with AVEVA Process Simulation via MCP enables natural language driven flowsheet analysis, optimization, and construction for chemical separation processes.
OpenSkill bootstraps LLM agent self-evolution by pulling grounded knowledge and anchors from open-world sources, synthesizing transferable skills, and refining them on self-generated virtual tasks, achieving top benchmark pass rates without supervision.
Survey of auto-research systems identifies objective, validation, and acceptance collapses, concluding that workflow closure does not equal scientific closure and advocating non-autonomous epistemic control.
MultiSearch uses parallel multi-query retrieval plus explicit merging inside a reinforcement-learning loop to improve retrieval-augmented reasoning, outperforming baselines on seven QA benchmarks.
ViDR treats source figures as retrievable and verifiable evidence objects in multimodal deep research reports and introduces MMR Bench+ to measure improvements in visual integration and verifiability.
Web2BigTable introduces a bi-level multi-agent system that achieves new state-of-the-art results on wide-coverage and deep web-to-table search benchmarks through orchestration, coordination, and closed-loop reflection.
Generative multi-agent systems exhibit emergent collusion and conformity behaviors that cannot be prevented by existing agent-level safeguards.
IoDResearch is a private data-centric Deep Research framework that uses FAIR digital objects, atomic knowledge units, heterogeneous graph indexes, and a multi-agent system to outperform standard RAG baselines on retrieval, QA, and report generation tasks.
ERL trains LLMs to erase faulty reasoning steps and regenerate them in place, yielding gains of up to 8.48% EM on multi-hop QA benchmarks like HotpotQA.
VaseMuseum is a training-free multimodal agent that combines DeepResearch-style retrieval, source/response reliability control, and best-of-K reranking to improve citation validity and reduce hallucination for museum VQA on ancient Greek pottery.
Argues for a denoising-first paradigm in LLM-oriented information retrieval, framing challenges via a four-stage progression and providing a taxonomy of signal-to-noise optimization techniques across the pipeline.
citing papers explorer
-
Multi-Head Recurrent Memory Agents
The paper proposes Multi-Head Recurrent Memory (MHM) with a select-then-update strategy to improve memory retention in long-context recurrent agents.
-
HarnessBridge: Learnable Bidirectional Controller for LLM Agent Harness
HarnessBridge is a learnable bidirectional controller that projects observations and actions for LLM agent harnesses, matching specialized performance while cutting tokens and generalizing across model sizes.
-
PropGuard: Safeguarding LLM-MAS via Propagation-Aware Exploration and Remediation
PropGuard is a propagation-aware framework for LLM-MAS that constructs dual-view spatio-temporal graphs, employs a GE-GRPO inspector to recover suspicious subgraphs, and applies source-guided remediation to lower attack success while preserving task performance.
-
Information Gain-based Rollout Policy Optimization: An Adaptive Tree-Structured Rollout Approach for Multi-Turn LLM Agents
IGRPO allocates multi-turn LLM agent rollout budget proportional to intermediate-state information gain, inducing an exponentially tilted teacher distribution for policy optimization.
-
MedBench v5: A Dynamic, Process-Oriented, and Hallucination-Aware Benchmark for Clinical Multimodal Models
Strong medical-model task scores do not guarantee process stability: omission, contradiction, and delay mainly break contradiction detection, diagnosis updating, and hallucination self-correction while final evidence grounding can look fine.
-
MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval
MCompassRAG adds topic metadata to chunk representations and uses LLM distillation to train a lightweight topic-aware retriever, reporting 8.24% average information efficiency gain and over 5x lower latency than strong baselines across six benchmarks.
-
Creative Reading: Scaffolding Reading for Transformation
A design space for reading augmentation that contrasts transmission-oriented 'reading to discard' with transformation-oriented 'creative reading' and identifies two opportunities for future systems.
-
Helicase: Uncertainty-Guided Supply Chain Knowledge Graph Construction with Autonomous Multi-Agent LLMs
Helicase proposes an autonomous multi-agent LLM framework for uncertainty-guided supply chain knowledge graph construction evaluated on the new SCQA benchmark of 80 queries.
-
Position: Academic Conferences are Potentially Facing Denominator Gaming Caused by Fully Automated Scientific Agents
Malicious actors could use AI agents to submit large numbers of fake papers, inflating the submission count and thereby raising the acceptance odds for a small set of chosen legitimate papers under stable conference acceptance rates.
-
Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight
Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to cut up to 50% wasted reasoning tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance cost.
-
AgentEconomist: An End-to-end Agentic System Translating Economic Intuitions into Executable Computational Experiments
AgentEconomist is an end-to-end agentic system with idea development, experimental design, and execution stages that uses a large economics paper database to produce research ideas with better literature grounding, novelty, and insight than generic LLMs.
-
Self-Optimizing Multi-Agent Systems for Deep Research
Multi-agent deep research systems self-optimize prompts through self-play to match or outperform expert-crafted versions.
-
Learning to Retrieve from Agent Trajectories
Retrievers trained on agent trajectories via the LRAT framework improve evidence recall, task success, and efficiency in agentic search benchmarks.
-
Large Language Model Agent for User-friendly Chemical Process Simulations
An LLM agent integrated with AVEVA Process Simulation via MCP enables natural language driven flowsheet analysis, optimization, and construction for chemical separation processes.
-
OpenSkill: Open-World Self-Evolution for LLM Agents
OpenSkill bootstraps LLM agent self-evolution by pulling grounded knowledge and anchors from open-world sources, synthesizing transferable skills, and refining them on self-generated virtual tasks, achieving top benchmark pass rates without supervision.
-
Workflow Closure Is Not Scientific Closure in Auto-Research Systems
Survey of auto-research systems identifies objective, validation, and acceptance collapses, concluding that workflow closure does not equal scientific closure and advocating non-autonomous epistemic control.
-
Scaling Retrieval-Augmented Reasoning with Parallel Search and Explicit Merging
MultiSearch uses parallel multi-query retrieval plus explicit merging inside a reinforcement-learning loop to improve retrieval-augmented reasoning, outperforming baselines on seven QA benchmarks.
-
ViDR: Grounding Multimodal Deep Research Reports in Source Visual Evidence
ViDR treats source figures as retrievable and verifiable evidence objects in multimodal deep research reports and introduces MMR Bench+ to measure improvements in visual integration and verifiability.
-
Web2BigTable: A Bi-Level Multi-Agent LLM System for Internet-Scale Information Search and Extraction
Web2BigTable introduces a bi-level multi-agent system that achieves new state-of-the-art results on wide-coverage and deep web-to-table search benchmarks through orchestration, coordination, and closed-loop reflection.
-
Emergent Social Intelligence Risks in Generative Multi-Agent Systems
Generative multi-agent systems exhibit emergent collusion and conformity behaviors that cannot be prevented by existing agent-level safeguards.
-
IoDResearch: Deep Research on Private Heterogeneous Data via the Internet of Data
IoDResearch is a private data-centric Deep Research framework that uses FAIR digital objects, atomic knowledge units, heterogeneous graph indexes, and a multi-agent system to outperform standard RAG baselines on retrieval, QA, and report generation tasks.
-
Erase to Improve: Erasable Reinforcement Learning for Search-Augmented LLMs
ERL trains LLMs to erase faulty reasoning steps and regenerate them in place, yielding gains of up to 8.48% EM on multi-hop QA benchmarks like HotpotQA.
-
VaseMuseum: Digital Intelligent Museum for Ancient Greek Pottery
VaseMuseum is a training-free multimodal agent that combines DeepResearch-style retrieval, source/response reliability control, and best-of-K reranking to improve citation validity and reduce hallucination for museum VQA on ancient Greek pottery.
-
LLM-Oriented Information Retrieval: A Denoising-First Perspective
Argues for a denoising-first paradigm in LLM-oriented information retrieval, framing challenges via a four-stage progression and providing a taxonomy of signal-to-noise optimization techniques across the pipeline.
-
EigentSearch-Q+: Enhancing Deep Research Agents with Structured Reasoning Tools
Structured query and evidence tools added to an AI research agent improve benchmark accuracy by 0.6 to 3.8 percentage points.
-
Towards Trustworthy Report Generation: A Deep Research Agent with Progressive Confidence Estimation and Calibration
A deep research agent incorporates progressive confidence estimation and calibration to produce trustworthy reports with transparent confidence scores on claims.