Papers by Damai Dai
Bi-Drop: Enhancing Fine-tuning Generalization via Synchronous sub-net Estimation and Optimization (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained language models can be fine-tuned on limited training data, which can overfit and thus diminish performance. |
| Approach: | They propose a fine-tuning strategy that selectively updates model parameters using gradients from various sub-nets dynamically generated by dropout. |
| Outcome: | The proposed method outperforms existing methods on the GLUE benchmark and exhibits excellent generalization ability and robustness for domain transfer, data imbalance, and low-resource scenarios. |
Denoising Bottleneck with Mutual Information Maximization for Video Multimodal Fusion (2023.acl-long)
Copied to clipboard
| Challenge: | Prior denoising methods suppress redundant and noisy information at risk of losing critical information. |
| Approach: | They propose a denoising bottleneck fusion model for fine-grained video multimodal fusion . they employ a bottleneck mechanism to filter out noise and redundancy with a restrained receptive field . |
| Outcome: | The proposed model improves on state-of-the-art video multimodal fusion benchmarks. |
DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models (2024.acl-long)
Copied to clipboard
Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, Wenfeng Liang
| Challenge: | Mixture-of-Experts (MoE) architectures face challenges in ensuring expert specialization . despite the promising performance, scaling language models to an extremely large scale is associated with exceedingly high computational costs. |
| Approach: | They propose an architecture that allows for ultimate expert specialization by segmenting experts into mN ones and activating mK from them. |
| Outcome: | The proposed architecture achieves comparable performance with GShard with 2B parameters and computation. |
Hierarchical Curriculum Learning for AMR Parsing (2022.acl-short)
Copied to clipboard
| Challenge: | Abstract Meaning Representation (AMR) parsing aims to translate sentences to semantic representation with a hierarchical structure, but there is a gap between their flat training objective and the hierarchic structure, which limits the model generalization. |
| Approach: | They propose a Hierarchical Curriculum Learning framework with Structure-level (SC) and Instance-level curricula (IC) that aims to translate sentences to semantic representation with a hierarchical structure. |
| Outcome: | Experiments on AMR2.0, AMR3.0, structure-complex and out-of-distribution situations confirm the effectiveness of the proposed framework. |
Knowledge Neurons in Pretrained Transformers (2022.acl-long)
Copied to clipboard
| Challenge: | Existing studies show that pretrained language models are good at recalling factual knowledge without fine-tuning. |
| Approach: | They propose a method to identify neurons that express factual knowledge in pretrained Transformers by filling-in-the-blank cloze queries. |
| Outcome: | The proposed method can be used to edit, erase, and update factual knowledge without fine-tuning. |
Leveraging Word-Formation Knowledge for Chinese Word Sense Disambiguation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Word sense disambiguation (WSD) methods have not explored word-formations in parataxis languages like Chinese. |
| Approach: | They propose to leverage word-formation knowledge to enhance Chinese WSD by incorporating word-forms into sense disambiguation models. |
| Outcome: | The proposed model improves on baselines in Chinese word sense disambiguation (WSD) with word-formation knowledge, the results show. |
Why Can GPT Learn In-Context? Language Models Secretly Perform Gradient Descent as Meta-Optimizers (2023.findings-acl)
Copied to clipboard
| Challenge: | Large pretrained language models have shown surprising in-context learning ability . despite the great success in performance, its working mechanism remains unclear . |
| Approach: | They explain language models as meta-optimizers and understand in-context learning as implicit finetuning . they find that Transformer attention has a dual form of gradient descent . |
| Outcome: | The proposed model can predict labels for unseen inputs without parameter updates . the proposed model outperforms smaller models with a single parameter update . |
StableMoE: Stable Routing Strategy for Mixture of Experts (2022.acl-long)
Copied to clipboard
| Challenge: | Existing learning-to-route methods suffer from the routing fluctuation issue . with the model scale growing, training speed will go slower and memory requirements are heavy . |
| Approach: | They propose a Mixture-of-Experts technique that can scale up the model size of Transformers with an affordable computational overhead. |
| Outcome: | The proposed method outperforms existing learning-to-route methods on language modeling and multilingual machine translation. |
Calibrating Factual Knowledge in Pretrained Language Models (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies show that Pretrained Language Models can store factual knowledge, but facts stored in PLMs are not always correct. |
| Approach: | They propose a lightweight method to calibrate factual knowledge in PLMs without re-training from scratch. |
| Outcome: | The proposed method can be used to calibrate factual knowledge in PLMs without re-training from scratch. |
A Survey on In-context Learning (2024.emnlp-main)
Copied to clipboard
Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, Zhifang Sui
| Challenge: | In-context learning (ICL) is a new paradigm for natural language processing . large language models (LLMs) demonstrate the ability to learn from a few examples . |
| Approach: | They propose to explore ICL to evaluate and extrapolate the ability of large language models. |
| Outcome: | The proposed methods can be used to evaluate and extrapolate the ability of large language models. |
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention (2025.acl-long)
Copied to clipboard
Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Yuxing Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, Wangding Zeng
| Challenge: | Long-context modeling is crucial for next-generation language models, but high computational cost of standard attention mechanisms poses significant computational challenges. |
| Approach: | They propose a natively trained Sparse Attention mechanism that integrates algorithms with hardware-aligned optimizations to achieve efficient long-context modeling. |
| Outcome: | The proposed model maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. |
Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning (2023.emnlp-main)
Copied to clipboard
| Challenge: | In-context learning (ICL) is a promising capability for large language models (LLMs) but its underlying mechanism remains unexplored. |
| Approach: | They propose a demonstration compression technique to expedite inference and an analysis framework for diagnosing ICL errors in GPT2-XL. |
| Outcome: | The proposed method improves ICL performance and expedites inference. |
Not All Demonstration Examples are Equally Beneficial: Reweighting Demonstration Examples for In-Context Learning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have recently gained the In-Context Learning ability . however, the quality of demonstration examples is usually uneven . |
| Approach: | They propose to determine optimal weights for demonstration examples and apply them during ICL. |
| Outcome: | The proposed approach outperforms conventional ICL on 8 classification tasks. |
Learning to Control the Fine-grained Sentiment for Story Ending Generation (P19-1)
Copied to clipboard
| Challenge: | Existing studies focus on controlling the sentiment of story endings. |
| Approach: | They propose a generic and novel framework which controls fine-grained sentiment intensity for automatic story ending generation without manually annotating sentiment labels. |
| Outcome: | The proposed framework can generate story endings which meet the given sentiment intensity better. |
Language Models Encode the Value of Numbers Linearly (2025.coling-main)
Copied to clipboard
| Challenge: | Existing studies show that large language models encode the value of numbers linearly. |
| Approach: | They construct a large language model and use linear probes to read out input numbers from hidden states. |
| Outcome: | The proposed model encodes the value of numbers linearly, and can store the outputs via simple vector additions. |
Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for process-oriented math reward models rely on manual annotation. |
| Approach: | They propose a process-oriented math process reward model called Math-shepherd which assigns a reward score to each step of math problem solutions. |
| Outcome: | The proposed model breaks the bottleneck of manual supervision in two scenarios. |
Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models (2026.acl-long)
Copied to clipboard
Xin Cheng, Wangding Zeng, Damai Dai, Qinyu Chen, Bingxuan Wang, Zhenda Xie, Kezhao Huang, Xingkai Yu, Zhewen Hao, Han Zhang, Yu-Kun Li, Huishuai Zhang, Dongyan Zhao, Wenfeng Liang
| Challenge: | Mixture-of-Experts (MoE) scales capacity via conditional computation, but lacks knowledge lookup primitive. |
| Approach: | They propose a conditional memory instantiated via Deep Sparse Embedding (DSE) they propose 'u-shaped scaling law' that identifies optimal balance between MoE experts and DSE memory . |
| Outcome: | The proposed model outperforms an iso-parameter and isoFLOPs MoE baseline across knowledge and reasoning benchmarks and is infrastructure-efficient. |
Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on parameter-efficient fine-tuning (PEFT) for dense-architecture LLMs are lacking. |
| Approach: | They propose an expert-specialized fine-tuning method that tunes the experts most relevant to downstream tasks while freezing the other experts. |
| Outcome: | The proposed method matches or surpasses full-parameter fine-tuning. |
Decompose, Fuse and Generate: A Formation-Informed Method for Chinese Definition Generation (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing definition generation methods take the source word as an indecomposable semantic unit, but in parataxis languages like Chinese, word meanings can be composed using the word formation process. |
| Approach: | They propose to use word formation features to enhance Definition Generation (DG) in Chinese to generate an explanatory text. |
| Outcome: | The proposed model enhances Definition Generation (DG) in Chinese by decomposing the word meaning into different semantic components. |