Jianhe Xie

Zhaocheng Liu, Quan Tu, Wen Ye et al.

Recently, large language models have shown great potential to transform online medical consultation. Despite this, most research targets improving diagnostic accuracy with ample information, often overlooking the inquiry phase. Some studies try to evaluate or refine doctor models by using prompt-engineered patient agents. However, prompt engineering alone falls short in accurately simulating real patients. We need to explore new paradigms for patient simulation. Furthermore, the relationship between inquiry and diagnosis remains unexplored. This paper extracts dialogue strategies from real doctor-patient conversations to guide the training of a patient simulator. Our simulator shows higher anthropomorphism and lower hallucination rates, using dynamic dialogue strategies. This innovation offers a more accurate evaluation of diagnostic models and generates realistic synthetic data. We conduct extensive experiments on the relationship between inquiry and diagnosis, showing they adhere to Liebig's law: poor inquiry limits diagnosis effectiveness, regardless of diagnostic skill, and vice versa. The experiments also reveal substantial differences in inquiry performance among models. To delve into this phenomenon, the inquiry process is categorized into four distinct types. Analyzing the distribution of inquiries across these types helps explain the performance differences. The weights of our patient simulator are available https://github.com/PatientSimulator/PatientSimulator.

Boyu Gou, Zanming Huang, Yuting Ning et al. · microsoft-research

Agentic search such as Deep Research systems-where agents autonomously browse the web, synthesize information, and return comprehensive citation-backed answers-represents a major shift in how users interact with web-scale information. While promising greater efficiency and cognitive offloading, the growing complexity and open-endedness of agentic search have outpaced existing evaluation benchmarks and methodologies, which largely assume short search horizons and static answers. In this paper, we introduce Mind2Web 2, a benchmark of 130 realistic, high-quality, and long-horizon tasks that require real-time web browsing and extensive information synthesis, constructed with over 1000 hours of human labor. To address the challenge of evaluating time-varying and complex answers, we propose a novel Agent-as-a-Judge framework. Our method constructs task-specific judge agents based on a tree-structured rubric design to automatically assess both answer correctness and source attribution. We conduct a comprehensive evaluation of ten frontier agentic search systems and human performance, along with a detailed error analysis to draw insights for future development. The best-performing system, OpenAI Deep Research, can already achieve 50-70% of human performance while spending half the time, highlighting its great potential. Altogether, Mind2Web 2 provides a rigorous foundation for developing and benchmarking the next generation of agentic search systems.

Kai Zhang, Xiangchao Chen, Bo Liu et al. · microsoft-research

A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.