Papers by Zechun Liu

8 papers
Mixture-of-Supernets: Improving Weight-Sharing Supernet Training with Architecture-Routed Mixture-of-Experts (2024.findings-acl)

Copied to clipboard

Challenge: Neural architecture search (NAS) uses weight-sharing supernets to generate diverse subnetworks without retraining.
Approach: They propose a weight-sharing supernet that leverages mixture-of-experts to enhance supernet model expressiveness with minimal training overhead.
Outcome: The proposed method achieves state-of-the-art (SoTA) performance in NAS for fast machine translation models, surpassing NAS-BERT and AutoDistil across various model sizes.
Binary and Ternary Natural Language Generation (2023.acl-long)

Copied to clipboard

Challenge: ternary and binary neural networks have proven difficult to optimize since both parameter and output space are discretized . authors demonstrate ternaries and binary models on downstream tasks of summarization and machine translation .
Approach: They propose to use ternary and binary neural networks to optimize for multiplication-free computation . they propose to apply statistics-based quantization for the weights and elastic quantization of the activations to the transformer text generation model.
Outcome: The proposed model outperforms the best existing models on machine translation tasks.
Scaling Parameter-Constrained Language Models with Quality Data (2024.emnlp-industry)

Copied to clipboard

Challenge: Scaling laws in language modeling quantify training loss as a function of dataset size and model parameters, but neglect the critical role of data quality in model generalization.
Approach: They propose to use effective training tokens as a combination of text diversity and syntheticity as measured by a teacher model to calculate scaling laws.
Outcome: The proposed term effective training tokens is a combination of two readily-computed indicators of text diversity and syntheticity as measured by a teacher model.
Target-Aware Language Modeling via Granular Data Sampling (2024.emnlp-main)

Copied to clipboard

Challenge: Language model pretraining is the cornerstone of universal language models (LMs), creating generalpurpose representations to excel across a variety of downstream tasks.
Approach: They propose to use multi-granular tokens to sample large-scale language models for domain-specific use cases.
Outcome: The proposed model outperforms random sampled samples on eight benchmarks with 1% of the data and performs on par with the full RefinedWeb data.
LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Several post-training quantization methods have been shown to perform well down to 8-bits.
Approach: They propose a data-free distillation method that leverages generations produced by the pre-trained model to quantize any generative model independent of its training data.
Outcome: The proposed method outperforms SoTA PTQ and LLaMA models at low bit precision.
LLM-FP4: 4-Bit Floating-Point Quantized Transformers (2023.emnlp-main)

Copied to clipboard

Challenge: Existing quantization solutions are integer-based and struggle with bit widths below 8 bits.
Approach: They propose a method for quantizing weights and activations in large language models down to 4-bit floating-point values in a post-training manner.
Outcome: The proposed method outperforms existing methods on common sense zero-shot reasoning tasks by 12.7 points.
MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale Deployment (2026.acl-industry)

Copied to clipboard

Challenge: MobileLLM-Flash is a family of foundation models for efficient on-device use with strong capabilities.
Approach: They propose a method for designing on-device large language models under mobile latency constraints using hardware-in-the-loop architecture search.
Outcome: The proposed model is amenable to industry-scale deployment and is compatible with mobile runtimes like Executorch.
RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization (2024.findings-emnlp)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) improves training efficiency by updating only a small portion of the weights in Large Language Models.
Approach: They propose a rotation-aware scheme to fine-tune rotated outlier-free LLMs for effective weight-activation quantization.
Outcome: The proposed method improves low-bit LoRA convergence and post-training quantization robustness.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations