Tag
DR Pixel is an open-source end-to-end multimodal deep research agent that retrieves and reasons directly over rendered document page images instead of OCR-extracted text, introducing counterfactual credit reweighting for agentic supervised finetuning. The approach outperforms OCR-based baselines on DRPixelBench across models like GPT-5.4 and Gemini-3.5-Flash, boosting retrieval recall by ~15 points and accuracy by 10+ points after training.
The paper introduces PatEx-Gym, the first framework for developing and evaluating agents for patent examination, including a 1K-task benchmark (PatEx-Bench), an annotation-free trajectory synthesis method (PASS), and a compact 4B PatEx-Agent trained via supervised fine-tuning that competes with much larger general-purpose agents.
作者探讨构建研究型 AI 智能体时的两种架构选择：直接调用一体化 deep research API（Gemini Deep Research、OpenAI、Tavily Research 等），或基于搜索/抓取原语（Firecrawl、Exa、Tavily）自建规划与合成循环，并就质量、可观测性、扩展性与成本估算向生产环境实践者提问。
开发者发布了一款免费开源的 Claude Desktop 替代工具，支持本地运行、任意 LLM、深度研究能力，并可通过 Docker 自托管连接 40+ 数据源。
提出 Self-Assessment 强化学习框架（SARL），让检索增强的 LLM 搜索智能体在训练中学会感知自身搜索状态、识别缺失信息，从而提升搜索效率而无需额外推理时评估，在多跳推理与深度研究基准上取得显著相对提升。
DiffResearch 将自回归骨干替换为扩散语言模型（dLLM）用于多智能体深度研究，在 Deep Research Bench 上实现约 6 倍端到端加速并提升报告质量，还提出 dFast 提前从去噪状态提取子查询，以 Apache-2.0 开源。
browsecomp-plus-mini 是一个 15 条查询的轻量子集，能以不到完整基准 2% 的成本，大致复现 BrowseComp-Plus 全部 830 条查询对深度研究智能体的排名（加权 Spearman 一致性 0.88），且无需 LLM 评分，仅用规范化精确匹配打分。
作者对比了六家深度研究 API（OpenAI、Perplexity、Parallel、Tavily、Gemini、Valyu），分析各自的功能支持、限制、交付物能力以及在研究工作流中的适用场景，并附上功能与定价对比表。
作者对比了 11 种可在编码 Agent 内运行或由 Agent 驱动的深度研究工具，涵盖报告生成、表格输出、来源聚合与审计层等类别，并讲述了自己 fork 其中一个项目后进行改造的过程与发现。
该论文研究了对多智能体深度研究智能体的系统提示进行反思式提示进化（GEPA），采用主题不重叠、多随机种子的评估协议，发现分解式逐维度反馈是可迁移增益的关键因素，而压缩为单一标量奖励会导致负迁移。
PANINI-DR提出了一种多智能体深度研究框架，通过生成式语义工作区分离协调、研究与验证职责，在BrowseComp-Plus等三个基准上取得最佳平均性能，BrowseComp-Plus准确率达73.3%，领先次优系统16.0个百分点。
本文提出 Incremental Open-Ended Deep Research（Incremental-OEDR）研究设定与 Structured Harness 框架，将研究报告视为可随新信息持续演化的结构化状态，在保持报告质量的同时显著降低研究成本（token 消耗减少 33%，搜索调用减少 61%）。
Web$hell 提出了一种统一的 LLM 智能体 Web 访问接口，通过 WebContract、WebCompiler、WebCache 将实时网页编译为可寻址的信息与操作对象，实现了搜索增强、浏览器操作与文件系统式访问的融合，在多项 Web 搜索、导航与深度研究基准上以更低成本取得强劲性能。
Zilliz has open-sourced DeepSearcher, an Apache-2.0 deep research framework combining LLMs, vector databases like Milvus, and agents to let companies run private, on-prem research over their own knowledge bases instead of sending data to online AI services.
A curated roundup of 8 open-source GitHub projects — including GPT Researcher, STORM, Perplexica, Crawl4AI, PaperQA2, Docling, GraphRAG, and LightRAG — that turn a research question into sourced, cross-checked, evidence-backed reports.
BAAI releases AREX-2, a 27B long-horizon self-improving agent model based on Qwen3.8 that learns to propose, measure, reflect, and revise solutions over multiple test-time rounds, achieving strong results on coding, ML engineering, and deep research benchmarks with only 27B parameters.
NVIDIA used Tavily to generate high-quality research trajectories for post-training Nemotron 3 Super, leading to top rankings on DeepResearch Benchmarks. The process highlights how specialized fine-tuning enhances research agent efficiency and performance.
LongCat-DeepResearch is a deep research system that combines an enhanced LongCat model with a multi-agent workflow for producing comprehensive, evidence-grounded reports, achieving competitive scores on benchmarks like DeepResearchBench and ResearchRubrics.
Hyperresearch is an open-source Python tool that acts as a deep research harness for Claude Code and OpenAI Codex, using agents to collect, search, and synthesize web research into a persistent, searchable wiki with adversarial auditing.
AdaTutoRank introduces a setwise reranker trained with Adaptive Tutoring Optimization for enhanced document set ranking in RAG and deep research, achieving superior performance with reduced retrieval calls.