Token-level confidence gain as a training-free signal for detecting toxic retrieval and switching generation modes. → GOING, ACM TIST (in submission)

以 token 级置信度增益作为免训练信号，检测有害检索并切换生成模式。→ GOING, ACM TIST（投稿中）

Every generated sentence must cite the exact spatial tokens it draws from, enforcing evidence commitment by construction. → MARIE, ACM MM '26. Extending the same reliability lens to multi-property molecule optimization with instruction-tuned LLMs for lead optimization in drug discovery — score-free candidate selection when property predictors are noisy. → NeurIPS / AAAI '26 (in preparation). The token-gating idea also parallels grounding VLA planning outputs in BEV tokens for autonomous driving and robotic manipulation.

每条生成句子须引用其所依赖的空间 token，从设计层面保障证据承诺。→ MARIE, ACM MM '26。同一可靠性视角也延伸至药物发现中基于指令微调 LLM 的多属性分子优化——在属性预测噪声下进行无评分候选选择。→ NeurIPS / AAAI '26（准备中）

GAN/GRU-CNN architectures for stock prediction; ESG-driven bond issuance forecasting with XGBoost and LightGBM.

GAN/GRU-CNN 架构用于股价预测；基于 XGBoost 与 LightGBM 的 ESG 债券发行预测。

A future direction I am actively pursuing: applying spatial token grounding techniques to medical AI, embodied AI, and autonomous driving, where VLA models face the same core challenge of anchoring planning outputs to explicit perceptual evidence. The MARIE evidence commitment mechanism provides a direct conceptual bridge.

我正在积极探索的未来方向：将空间 token 定位技术延伸至医疗、具身 AI 与自动驾驶领域。VLA 模型面临与 MARIE 相同的核心挑战，即如何将规划输出锚定于明确的感知证据；MARIE 的证据承诺机制为此提供了直接的概念桥梁。