deep-research

Introducing Agent Ultra, a deep research tool that uses orchestrated agents and frontier models to perform exhaustive web research, claimed to be state-of-the-art.

Meta has launched a free AI search product that uses deep research for regular search, making it difficult for other companies to compete.

Researchers at Microsoft introduced DualGraph, an architecture that splits agent memory into Outline and Knowledge Graphs to enhance autonomous research, achieving unprecedented RACE scores and cost savings with Claude Opus 5.5 and GPT-6 Sol.

The paper proposes a method called DLD-RL to enhance deep research and long-context abilities in AI agents by repurposing reinforcement learning trajectories into long-context QA data, showing significant performance improvements on benchmarks.

The article critiques open AI recreations, comparing them to chaining web search tool calls and falsely calling it open deep research.

The author shares a write-up on building reliable agentic harnesses, including benchmarks from 4,000 runs and plans to release an open-source deep research and builder application, emphasizing quality testing and insights on AI agent development.

The paper introduces Mr.LHDR, a benchmark for evaluating multimodal real-world long-horizon deep research agents, showing that current models struggle with dependency-consistent evidence integration in complex reasoning chains.

The author is building a deal screening system and faces high costs when running deep research on multiple companies, seeking advice on cost-effective strategies like funneling, reuse, or alternative tools.

This paper introduces HybridDeepResearch, a benchmark for evaluating AI agents on tasks that require integrating web search and SQL querying, revealing that even state-of-the-art models struggle with maintaining constraints across structured and unstructured data.

APEx is a hierarchical experience utilization framework that organizes agent history into memories and skills, using a closed-loop architecture to improve deep research question answering, achieving state-of-the-art performance surpassing GPT-5.4.

The paper introduces ClaimProbe, a claim-level audit framework to detect hallucinations and misattributions in deep research reports, and ClaimWriter, a hierarchical writer that reduces hallucinations by up to 4.5 times while improving fact recall.

Apodex is an open-sourced deep research harness and AI model, which is a finetune of Qwen 3.5 35B A3B, achieving performance comparable to frontier models with only 3B active parameters.

Apodex 1.1 is an AI agent designed for deep research, capable of handling complex tasks that require extensive investigation. It uses a main agent to decompose problems and asynchronously dispatches multiple sub-agents for execution, supporting long-running operations and automatic recovery.

Google is adding a dedicated student hub to Gemini with features like study notebooks, flashcards, and calendar integration, along with Deep Research in Gemini Live and new Lens educational tools.

This paper introduces ICD-Deepresearch, a system that combines EHR foundation models with language models and medical search to forecast future ICD codes for patient visits, achieving improved accuracy and physician-rated usefulness over baseline methods.

Mole is a deep research agent for the terminal that uses AI to decompose questions, search sources, and provide answers with verified citations while enforcing budgets and ensuring local data privacy.

LangChain has open-sourced open_deep_research, based on LangGraph's multi-agent architecture, a fully open-source replica of OpenAI Deep Research. It supports multiple models and search tools and can be deployed locally to generate in-depth research reports.

WebRover is an MIT-licensed open-source AI agent that uses natural language to drive the browser for web automation, cross-site data scraping, and deep research, with support for local deployment.

Introduces Deep Research Pretraining (DRP), an offline framework that generates search-open-write trajectories from citation and hyperlink evidence structures. Qwen3-14B models pretrained on 1B tokens with DRP outperform matched no-DRP baselines on deep research benchmarks, even with less supervised fine-tuning data.

Video-DeepResearch (Video-DR) extends multimodal agents from static images to continuous video streams, introducing a decoupled perception-exploration pipeline and a new benchmark Video-DR-Bench. Their Video-DeepResearch-35B-A3B model achieves 64.0% accuracy, surpassing Claude-4.5-Sonnet, GPT-5, and Gemini 2.5 Pro.