Yishan Jiang

I am Yishan Jiang, a final-year Computer Science undergraduate at Beijing University of Technology (2022–2026). My research bridges mechanistic interpretability with practical LLM applications, translating opaque model behaviors into measurable structural failure modes.

 

I am the co-first author of 1*KDD2026(co-first author). In this work, I formalized Topic Dilution as a structural failure in LLM-based labeling. To address this, I co-designed an ablation-and-injection framework that utilizes global discriminative signals to enforce strict semantic boundaries, directly improving downstream retrieval utility.

 

Currently, I am conducting an exploratory study in Representational Engineering (RepE). Using TransformerLens, I analyze layer-wise representations to investigate how fine-grained concepts collapse toward hypernyms in deeper layers—and how injecting discriminative context can mitigate this effect to build more reliable, grounded systems.

 

I am actively seeking a full-time Research Assistant (RA) position to construct highly controllable and robust LLM applications. By leveraging representation engineering and structural constraints, I aim to tackle real-world deployment challenges and ensure model outputs remain functionally precise in complex tasks.