Yixin Zhang (张益新)

Designing training-free Executor-Analyst architectures that decouple schema-bound tool execution from high-level clinical synthesis. Rather than exposing a monolithic LLM to long-horizon context drift (>12k tokens) across hundreds of noisy API definitions, the Executor performs strict atomic tool calls while the Analyst operates exclusively over compact, structured factual evidence.

• Stratified Multi-Agent Ensemble Reasoning: Combining complementary tool-augmented trajectories with internal parametric knowledge to mitigate single-path hallucination.
• Biomedical Tool Orchestration (211 Tools): High-precision retrieval and verification across PubMed, OpenFDA, ChEMBL, and ClinicalTrials.gov for complex therapeutic and regulatory decision-making.

构建免训练的执行器-分析器（Executor-Analyst）解耦架构，将底层的严格工具调用与高层临床证据综合完全隔离。通过识别并消除超长上下文（>12k tokens）下的注意力漂移（Attention Drift），由短上下文 Executor 负责原子级 API 校验与容错调用，由 Analyst 基于高信噪比结构化事实卡片完成高阶医学推理。

• 分层多智能体集成推理（Stratified Ensemble）：融合多样化工具增强轨迹与大模型内生医学先验，显著降低复杂临床决策中的幻觉率与单路径偏差。
• 211 项生物医学工具协同编排：深度打通 PubMed、OpenFDA、ChEMBL 与 ClinicalTrials.gov 等权威数据库，实现药物靶点、不良反应与临床指南的闭环核验。

Formalizing proactive turn-taking strictly at Transition Relevance Places (TRPs, ±200 ms) under the Principle of Deliberate Silence. We replace brittle silence timers with the Three-Factor Multiplicative Utility Model:

• Multiplicative Veto & Three-Question Validity Filter: Unifies acoustic-syntactic TRP probability A(t), semantic grounding & voiceprint confidence S(t), and Gricean net utility N(t) gated by New Information (Q₁), Immediate Need (Q₂), and Timeliness Superiority (Q₃) (>96.67% neutral state space defaults to silence).
• Full-Duplex Runtime State Machines: Hysteresis VAD segmentation, L2-normalized speaker voiceprint crosstalk veto, and zero-loss pre-roll user barge-in frame replay (60 ms detection).

在“蓄意沉默原则（Principle of Deliberate Silence）”下形式化建模话轮转换关联点（TRP, ±200 ms）的主动话轮接管机制。针对传统固定静音阈值（VAD）“要么迟钝、要么打断思考停顿”的交互困境，提出三因子乘性效用模型：

• 乘性一票否决与三问有效性过滤器：耦合声学-句法完成度 A(t)、语义锚定与声纹置信度 S(t) 以及受格赖斯准则（新信息 Q₁、即时需求 Q₂、时机优越性 Q₃）门控的净发言效用 N(t)，在中性先验下使 96.67% 状态空间默认保持安静。
• 全双工零丢失打断状态机：融合迟滞双阈值 VAD、非注册说话人声纹一票否决及 60 ms 零丢字预卷（Pre-roll）帧回放机制。

Bridging multi-simulator reinforcement learning (Isaac Gym / MuJoCo) and physical humanoid deployment on Unitree G1 through automotive-grade deterministic verification pipelines and hardware-in-the-loop safety gates.

• Four-Layer Blameless Postmortem Attribution: Reconstructing timestamped failure timelines across (1) Simulation Environment, (2) Interface Protocol, (3) Hardware Safety, and (4) Policy/Model Generalization to isolate root causes from contributing factors.
• Disturbance-Resilient Locomotion & Safety Gates: Rapid baseline-to-expert policy distillation with real-time joint torque/thermal telemetry and millisecond-level emergency disengagement on physical humanoid platforms.

将车规级“比特级 I/O 校验与节点隔离 SOP”引入具身智能领域，打通多仿真器强化学习（Isaac Gym / MuJoCo）到 Unitree G1（宇树 G1）人形机器人真机部署的确定性评测线束与硬件在环安全门控。

• 四层无责事故复盘归因（4-Layer Blameless Postmortem）：基于严格时间戳日志重建“仿真 → 接口 → 真机”事件链，将异常精准归因于仿真环境、接口协议、硬件安全、模型策略四个正交层级。
• 抗扰动全身运动控制与安全门：实现从基线控制到技能专家网络的高效蒸馏，在真机上完成站立平衡、连续步态、外力推搡恢复及跌倒自恢复，并配备毫秒级急停与黑匣子回放保护。