Mingcheng Presents at ICML 2026 in Seoul

Mingcheng Presents MedTPE at ICML 2026 in Seoul

On 10 July 2026, Mingcheng attended the International Conference on Machine Learning (ICML 2026) in Seoul to present his work, From Token to Token Pair: Efficient Prompt Compression for Large Language Models in Clinical Prediction. The conference provided an excellent opportunity to share the work with the international machine learning community and discuss recent advances in efficient and clinically applicable large language models (LLMs).

The paper addresses the challenge of processing long electronic health record (EHR) sequences with LLMs. Conventional tokenisers often divide medical terms into multiple fragments, creating unnecessarily long inputs and increasing computational cost. To address this issue, the team developed Medical Token-Pair Encoding (MedTPE), a prompt-compression framework that merges frequently co-occurring medical tokens using a dependency-aware replacement strategy and lightweight self-supervised fine-tuning. MedTPE reduced input sequence length by up to 31% and inference latency by 34 to 63%, while maintaining or improving performance across clinical prediction tasks.

Mingcheng’s presentation highlighted the potential of MedTPE to support more efficient and scalable use of large language models in real-world healthcare settings. The conference also offered valuable opportunities to exchange ideas with researchers working on efficient language modelling, clinical AI, and electronic health record analysis.

For more information, please refer to the paper.