Postdoc, City University of Hong Kong
2 papers at NeurIPS 2025
Based on systematic empirical analysis on post-training compression, we propose a calibration data curation framework to help pruning and quantization methods better preserve critical LLM capabilities.
This paper proposes a novel framework that generates better contrastive pairs for contrastive learning by integrating LLM-based semantic retrieval with a learnable sample synthesizer.
