Papers by Zining Liu
PipeSpec: Breaking Stage Dependencies in Hierarchical LLM Decoding (2025.findings-acl)
Copied to clipboard
| Challenge: | Speculative decoding is limited by sequential stage dependencies that prevent full hardware utilization. |
| Approach: | They propose a framework that generalizes speculative decoding to use multiple models arranged in a hierarchical pipeline and enables asynchronous execution with lightweight coordination for prediction verification and rollback. |
| Outcome: | The proposed framework achieves 2.25 tokens/unit through pipelined parallelism with multiple models arranged in a hierarchical pipeline. |
Mitigating Sequential Dependencies: A Survey of Algorithms and Systems for Generation-Refinement Frameworks in Autoregressive Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Sequential dependencies present a fundamental bottleneck in deploying large-scale autoregressive models . |
| Approach: | They analyze methods based on generation strategies and refinement mechanisms . they examine deployment strategies across computing environments and explore applications spanning text, images, and speech generation. |
| Outcome: | The proposed frameworks can be used to improve the quality of autoregressive models. |
Adaptive Contrastive Knowledge Distillation for BERT Compression (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods for BERT implicitly learn discriminative student features by mimicking the teacher features. |
| Approach: | They propose a new knowledge distillation approach called adaptive contrastive knowledge distilling for BERT compression using hidden state features in BERT as explicit supervision to learn discriminative student features. |
| Outcome: | The proposed approach improves on multiple natural language processing tasks. |
DREAM-S: Speculative Decoding with Searchable Drafting and Target-Aware Refinement for Multimodal Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Speculative decoding (SD) has proven to be effective for autoregressive generation in large language models (LLMs), however its application to vision-language models (VLMs) remains relatively unexplored. |
| Approach: | They propose a Speculative Decoding framework for vision-language models that integrates a neural architecture search framework and target-aware supernet training to identify optimal interaction strategies. |
| Outcome: | DREAM-S achieves 3.85 speedup compared to baselines on well-established vision-language models. |
Half-S: Halving the Scale for Near-Lossless 4-Bit LLM Training (2026.findings-acl)
Copied to clipboard
Jinyang Du, Ruihao Gong, Linghan Ai, Zining Wang, Yunke Peng, Yao Wang, Lei Yan, null Wxuefei, Yaoyuan Wang, Jinyang Guo, Dahua Lin, Xianglong Liu
| Challenge: | Existing 4-bit training pipelines rely on max-scaling, which causes representation collapse . despite this, there are limitations in the accuracy of 4-bit LLM training . |
| Approach: | They propose a scaling strategy that uses half-scaling as a hardware-friendly default . they propose fp4 support that allows for a faster scaling of large language models . |
| Outcome: | The proposed scaling strategy narrows the gap between theoretical optimum and BF16 while maintaining the efficiency benefits of 4-bit training. |