实验室两项研究成果被 ICLR 2026 接收 Two Lab Papers Accepted by ICLR 2026
近日，实验室两项研究成果被 The Fourteenth International Conference on Learning Representations（ICLR 2026） 接收。
“Map as a Prompt: Learning Multi-Modal Spatial-Signal Foundation Models for Cross-scenario Wireless Localization” 聚焦跨场景无线定位问题。该研究提出多模态基础模型 SigMap，通过联合建模无线信号与三维地理空间信息，并以 “Map-as-Prompt” 方式引入地图先验，提升模型对不同无线环境的适应与泛化能力。
“SecP-Tuning: Efficient Privacy-Preserving Prompt Tuning for Large Language Models via MPC” 聚焦隐私敏感场景下的大语言模型高效调优问题。该研究结合 Forward-only Tuning 与安全多方计算（MPC），并优化注意力计算机制，在保护用户数据和模型信息的同时降低调优开销，为大语言模型的隐私保护部署提供了新的技术路径。
Recently, two research works from our lab were accepted by The Fourteenth International Conference on Learning Representations (ICLR 2026).
The paper “Map as a Prompt: Learning Multi-Modal Spatial-Signal Foundation Models for Cross-scenario Wireless Localization” focuses on cross-scenario wireless localization. It proposes SigMap, a multi-modal foundation model that jointly models wireless signals and 3D geospatial information, introducing map priors through a “Map-as-Prompt” mechanism to improve adaptation and generalization across diverse wireless environments.
The paper “SecP-Tuning: Efficient Privacy-Preserving Prompt Tuning for Large Language Models via MPC” focuses on efficient LLM tuning in privacy-sensitive scenarios. By combining Forward-only Tuning with secure multi-party computation (MPC) and optimizing attention computation, SecP-Tuning reduces tuning overhead while protecting both user data and model information, providing a practical approach to privacy-preserving LLM deployment.