Tag
论文提出 HybridResearcher，一个 4B 小型深度研究智能体，通过合成跨源轨迹与强化学习在结构化与非结构化数据之间进行推理，在 HybridDeepResearch 基准上超越更大的 Qwen3.5-397B-A17B，同时不遗忘单源能力。
GRADS 提出一种图引导的智能体数据合成框架，通过维护证据与推理依赖图实现可控难度的多跳任务生成；经 SFT 与 RL 后训练后，4B 模型在软件工程、科学与安全领域基准上匹配或超越 Sonnet 4。
LifeSearch introduces a benchmark of 203 Chinese and English tasks for evaluating deep research agents in real-world life and productivity settings, with over 4,000 expert-written rubrics and experiments on 12 advanced agents, finding that even the best frontier agent (GPT-5.6 Sol) scores only 63.4%.
DeepRewind introduces an additive control layer that models a deep-research agent's epistemic state as a typed graph and uses a prompt-based world model plus consistency monitor to block and repair premature commitments, reducing them by 59.1% on DRBench and LiveDRBench.
DAGent (NeurIPS 2026) introduces an Evaluate-then-Grow DAG-based multi-agent framework where an Orchestrator expands the task graph batch-by-batch based on node confidence, along with DAGRPO, a topology-conditioned RL algorithm that outperforms GRPO on deep research benchmarks.
The paper introduces Question's Gambit, a first-move retrieval module that improves deep research agent performance by decomposing questions and providing a ranked context before iterative search, showing significant accuracy gains on BrowseComp-Plus.
WorldCupArena is a dynamic benchmark for evaluating language models and deep-research agents on football match prediction, using the 2026 FIFA World Cup as its first evaluation. It compares model predictions against actual results, providing fine-grained scoring beyond result accuracy.
New research from Cornell University shows that a single snippet of user-generated text as short as 13 words from sites like Reddit or Wikipedia can be used to manipulate AI search tools like ChatGPT and Google AI Search, highlighting a growing vulnerability in AI-powered information retrieval.
This paper introduces a claim-centric auditing framework for identifying error spans in deep-research agent trajectories, along with a new benchmark TELBench, improving process-level reliability assessment.
DR³-Eval is a benchmark for evaluating deep research agents on multimodal, multi-file report generation with a realistic web environment simulation and comprehensive evaluation framework measuring information recall, factual accuracy, citation coverage, instruction following, and depth quality.