Papers by Shuohuan Wang
ERNIE-Doc: A Retrospective Long-Document Modeling Transformer (2021.acl-long)
Copied to clipboard
| Challenge: | Existing models for document-level language pretraining are not suitable for long documents due to their quadratically increasing memory and time consumption. |
| Approach: | They propose a document-level language pretraining model based on Recurrence Transformers. |
| Outcome: | The proposed model outperforms existing models on language understanding tasks. |
Uncertainty-Aware Routing for Principled Alignment with MoE Dynamics (2026.acl-long)
Copied to clipboard
Yilong Chen, Junyuan Shang, Yuchen Feng, Zhenyu Zhang, Naibin Gu, Ziqi Wang, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang
| Challenge: | Mixture-of-Experts (MoE) is a cornerstone for scaling LLMs, yet its training dynamics remain poorly understood, often leading to sub-optimal specialization. |
| Approach: | They propose to use Helmholtz Free Energy and Router Entropy to study the MoE lifecycle and identify a universal Three-Stage Phase Transition . |
| Outcome: | The proposed model reduces perplexity and improves expert distinctiveness, offering a principled path toward thermodynamically aligned computation. |
LEMON: Reviving Stronger and Smaller LMs from Larger LMs with Linear Parameter Fusion (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods to train a stronger and smaller model with the help of large models are limited by the model size and performance. |
| Approach: | They propose to learn competent initial points for smaller models by fusing parameters from larger models and introduce controllable receptive fields to model prior parameter characteristics. |
| Outcome: | The proposed method outperforms baselines in terms of effectiveness and efficiency. |
Clip-Tuning: Towards Derivative-free Prompt Learning with a Mixture of Rewards (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work does not take full advantage of over-parameterized characteristics of large pre-trained language models. |
| Approach: | They propose a method that uses frozen "thinned" networks to obtain a mixture of rewards and advance the derivative-free prompt learning. |
| Outcome: | The proposed method outperforms previous gradient-free prompt learning methods and achieves parity with gradient-based counterparts on seven language understanding benchmarks under few-shot settings. |
ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for pretraining cross-lingual models are limited in their size due to the limited amount of parallel corpora. |
| Approach: | They propose a method that encourages the model to align multiple languages with monolingual corpora to overcome the constraint of the parallel corpus size. |
| Outcome: | The proposed method outperforms existing cross-lingual models and delivers new state-of-the-art results in various cross-linguistic downstream tasks. |
Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking (2025.acl-long)
Copied to clipboard
Yilong Chen, Junyuan Shang, Zhenyu Zhang, Yanxi Xie, Jiawei Sheng, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang
| Challenge: | Large language models face inherent performance bottlenecks under parameter constraints . challenging tokens induce abrupt gradient spikes across layers, exposing stress points . |
| Approach: | They propose an inner thinking transformer that reimagines layer computations as implicit thinking steps. |
| Outcome: | Empirical results show that ITT outperforms Transformer/Loop variants in 11 benchmarks. |
NACL: A General and Effective KV Cache Eviction Framework for LLM at Inference Time (2024.acl-long)
Copied to clipboard
Yilong Chen, Guoxia Wang, Junyuan Shang, Shiyao Cui, Zhenyu Zhang, Tingwen Liu, Shuohuan Wang, Yu Sun, Dianhai Yu, Hua Wu
| Challenge: | Large Language Models (LLMs) with extended context windows are expensive and infeasible on fixed memory hardware due to the surprisingly large memory consumption of KV Cache. |
| Approach: | They propose a general framework for long-context KV cache eviction that achieves more optimal and efficient evict in a single operation during the encoding phase. |
| Outcome: | The proposed framework improves performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 5 with over 95% performance maintenance. |
Autoregressive Pre-Training on Pixels and Texts (2024.emnlp-main)
Copied to clipboard
| Challenge: | pixel-based language modeling integrates visual and textual data to improve performance of language models. |
| Approach: | They propose a method that integrates visual and textual data into an autoregressive framework. |
| Outcome: | The proposed method improves performance of pixel-based language models by incorporating visual and textual data. |
HFT: Half Fine-Tuning for Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) with one or more fine-tuning phases can unlock various capabilities, but can be catastrophic forgetting during sequential training. |
| Approach: | They propose a method to regularly reset partial parameters to mitigate forgetting issues by using half fine-tuning instead of full fine-uning. |
| Outcome: | The proposed approach reduces the risk of catastrophic forgetting during training and the parametric knowledge lost during training may be overwhelmed by incoming training data. |
On Training Data Influence of GPT Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | generative language models have redefined performance standards across tasks . current research on the influence of training data on autoregressivity remains underexplored . |
| Approach: | They propose a parameterized simulation to assess the impact of training examples on the training dynamics of GPT models. |
| Outcome: | The proposed approach compares existing methods with existing methods across training scenarios in generative language models, spanning tasks across 14 million to 2.8 billion parameters. |
Correcting Chinese Spelling Errors with Phonetic Pre-training (2021.findings-acl)
Copied to clipboard
Ruiqing Zhang, Chao Pang, Chuanqiang Zhang, Shuohuan Wang, Zhongjun He, Yu Sun, Hua Wu, Haifeng Wang
| Challenge: | Existing methods for Chinese spelling correction only use pre-trained language model or incorporate phonological information as external knowledge. |
| Approach: | They propose a phonetic Chinese spelling correction model that integrates phonetic features into language model by leveraging pre-training and fine-tuning methods. |
| Outcome: | The proposed model outperforms existing methods on SIGHAN datasets and improves on other datasets. |
ERNIE-Code: Beyond English-Centric Cross-lingual Pretraining for Programming Languages (2023.findings-acl)
Copied to clipboard
| Challenge: | ERNIE-Code is a unified pre-trained language model for 116 NLs and 6 PLs. |
| Approach: | They propose a unified pre-trained language model for 116 NLs and 6 PLs . they employ span-corruption language modeling that learns patterns from monolingual NL or PL . |
| Outcome: | The proposed model outperforms previous multilingual models for NL or NL across end tasks. |
Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping (2026.findings-acl)
Copied to clipboard
Yao Chen, Yilong Chen, Yinqi Yang, Junyuan Shang, Zhenyu Zhang, Zefeng Zhang, Shuaiyi Nie, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang, Tingwen Liu
| Challenge: | Existing approaches to increasing effective depth of LLMs rely on parameter reuse, extending computation through recursive execution. |
| Approach: | They propose a training-time sparse depth allocation framework that progressively increases depth for a small subset of parameters as training evolves. |
| Outcome: | The proposed model outperforms existing approaches to increasing the effective depth of language models while reducing training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone. |
BeamLoRA: Beam-Constraint Low-Rank Adaptation (2025.acl-long)
Copied to clipboard
Naibin Gu, Zhenyu Zhang, Xiyu Liu, Peng Fu, Zheng Lin, Shuohuan Wang, Yu Sun, Hua Wu, Weiping Wang, Haifeng Wang
| Challenge: | Low-Rank Adaptation (LoRA) is one of the most efficient parameter-efficient fine-tuning methods. |
| Approach: | They propose to conceptualize each LoRA module as a beam where each rank corresponds to a potential sub-solution. |
| Outcome: | The proposed method improves performance on three base models and 12 datasets. |
Curiosity-Driven Reinforcement Learning from Human Feedback (2025.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models with human preferences, but often at the cost of reduced output diversity. |
| Approach: | They propose a framework that incorporates intrinsic rewards for novel states alongside traditional sparse extrinsic rewards to optimize both output diversity and alignment quality. |
| Outcome: | The proposed framework achieves significant gains in diversity on multiple diversity-oriented metrics while maintaining alignment with human preferences comparable to standard RLHF. |
Upcycling Instruction Tuning from Dense to Mixture-of-Experts via Parameter Merging (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for tuning large language models from dense to MoE face significant data requirements and require large-scale post-training. |
| Approach: | They propose an upcycling instruction tuning approach for tuning a dense pre-trained model into a MoE instruction model using genetic algorithm and parameter merging. |
| Outcome: | The proposed approach improves the performance of large language models with a small amount of seed data and improves their scaling. |