Papers by Denny Zhou
Symbol tuning improves in-context learning in language models (2023.emnlp-main)
Copied to clipboard
Jerry Wei, Le Hou, Andrew Lampinen, Xiangning Chen, Da Huang, Yi Tay, Xinyun Chen, Yifeng Lu, Denny Zhou, Tengyu Ma, Quoc Le
| Challenge: | Language models are sensitive to the way that prompts are given, indicating that they are not reasoning in a robust manner. |
| Approach: | They propose to fine tune language models on in-context input-label pairs where natural language labels are replaced with arbitrary symbols. |
| Outcome: | The proposed model is much stronger at reasoning tasks and more robust to underspecified prompts than the standard model. |
Token Dropping for Efficient BERT Pretraining (2022.acl-long)
Copied to clipboard
| Challenge: | Existing methods to accelerate pretraining of transformer-based models are computationally expensive and degrade performance on downstream tasks. |
| Approach: | They propose a "token dropping" method to accelerate the pretraining of transformer-based models by 25% . they leverage the already built-in masked language modeling loss to identify unimportant tokens with practically no computational overhead. |
| Outcome: | The proposed method reduces the pretraining cost of BERT models by 25% while achieving similar overall performance on downstream tasks. |
FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation (2024.findings-acl)
Copied to clipboard
Tu Vu, Mohit Iyyer, Xuezhi Wang, Noah Constant, Jerry Wei, Jason Wei, Chris Tar, Yun-Hsuan Sung, Denny Zhou, Quoc Le, Thang Luong
| Challenge: | Modern large language models often "hallucinate" plausible but factually incorrect information, which reduces their trustworthiness especially in settings where accurate and up-to-date information is critical. |
| Approach: | They develop a human evaluation procedure to measure correctness and hallucination and use it to benchmark both closed and open-source LLMs. |
| Outcome: | The proposed method outperforms both competing search engine-augmented prompting methods and commercial systems on search-augmented QA. |
Transcending Scaling Laws with 0.1% Extra Compute (2023.emnlp-main)
Copied to clipboard
Yi Tay, Jason Wei, Hyung Chung, Vinh Tran, David So, Siamak Shakeri, Xavier Garcia, Steven Zheng, Jinfeng Rao, Aakanksha Chowdhery, Denny Zhou, Donald Metzler, Slav Petrov, Neil Houlsby, Quoc Le, Mostafa Dehghani
| Challenge: | Existing scaling of language models is expensive and requires significant computational costs. |
| Approach: | They propose a method that substantially improves existing language models and their scaling curves with a relatively tiny amount of extra compute. |
| Outcome: | The proposed method significantly improves existing language models and their scaling curves with a relatively tiny amount of extra compute. |
MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices (2020.acl-main)
Copied to clipboard
| Challenge: | Empirical studies show that MobileBERT is 4.3x smaller and 5.5x faster than BERT_BASE . BERT is one of the largest models ever in NLP, but suffers from heavy model size and high latency . |
| Approach: | They propose a tool to compress and accelerate the popular BERT model by task-agnostic application. |
| Outcome: | The proposed model is 4.3x smaller and 5.5x faster than BERT_BASE . it achieves competitive results on well-known benchmarks . |
Fast WordPiece Tokenization (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for tokenization of text are not efficient, but they are based on Aho-Corasick's algorithm. |
| Approach: | They propose an efficient algorithm for WordPiece tokenization using a longest-match-first strategy . they propose an algorithm whose tokenization complexity is strictly O(n) |
| Outcome: | The proposed method is 8.2x faster than HuggingFace Tokenizers and 5.1x faster on average for general text tokenization. |
Extremely Small BERT Models from Mixed-Vocabulary Training (2021.eacl-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods cannot be directly applied to train student models with reduced vocabulary and embedding dimensions. |
| Approach: | They propose a method to align teacher and student embeddings via mixed-vocabulary training. |
| Outcome: | The proposed method compresses BERT-LARGE to a task-agnostic model with smaller vocabulary and hidden dimensions, which is an order of magnitude smaller than other distilled models. |
A Pretrainer’s Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, & Toxicity (2024.naacl-long)
Copied to clipboard
Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, Daphne Ippolito
| Challenge: | a large number of pretraining data design practices are under-documented, authors say . authors: strong performance of modern language models depends on selfsupervised pretraining . |
| Approach: | They propose to pretrain models on data curated at different collection times . they find temporal shift between evaluation data and pretraining data leads to performance degradation . |
| Outcome: | The results validate, quantify, and expose many undocumented intuitions about text pretraining . authors say this practice has outperformed other models in the field . |
Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them (2023.findings-acl)
Copied to clipboard
Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc Le, Ed Chi, Denny Zhou, Jason Wei
| Challenge: | Language models have already made good progress on this benchmark, with the best model outperforming average reported human-rater results on 65% of the BIG-Bench tasks. |
| Approach: | They propose to use chain-of-thought prompting to challenge language models on 23 challenging BIG-Bench tasks which they call BIG-Bench Hard. |
| Outcome: | The proposed language models outperform the average human-rater on 65% of the BIG-Bench tasks. |