Papers by Yizhe Zhang

Microsoft Icecaps: An Open-Source Toolkit for Conversation Modeling (P19-3)

Challenge:	Missing sentence generation fosters a wide range of applications in natural language generation . Developing models for sentence infilling can potentially facilitate many text generation applications .
Approach:	They propose a framework to decouple the problem from natural language processing . they propose generating missing sentences that can syntactically and semantically bridge context .
Outcome:	The proposed model learns a sentence representation and generates 'missing sentences' the proposed model can be used for document auto-completion and meeting note expansion .

Copied to clipboard

Vighnesh Leonardo Shiv, Chris Quirk, Anshuman Suri, Xiang Gao, Khuram Shahid, Nithya Govindarajan, Yizhe Zhang, Jianfeng Gao, Michel Galley, Chris Brockett, Tulasi Menon, Bill Dolan

DIALOGPT : Large-Scale Generative Pre-training for Conversational Response Generation (2020.acl-demos)

Challenge:	upcoming open-source natural language processing repository aims to train conversational agents for multi-turn situations.
Approach:	They present the Intelligent Conversation Engine: Code and Pre-trained Systems (ICECAPS) the framework wraps TensorFlow functionality in a modular component-based architecture.
Outcome:	The Intelligent Conversation Engine: Code and Pre-trained Systems (ICECAPS) is an open-source natural language processing repository.

Copied to clipboard

Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan

Decoding-Unlearning: Fact Forgetting via Entropy-Guided Inference (2026.acl-long)

Challenge:	DIALOGPT is a large, tunable neural conversational response generation model . trained on 147M conversation-like exchanges extracted from Reddit comment chains .
Approach:	They present a large, tunable neural conversational response generation model, DIALOGPT . the model is trained on 147M conversation-like exchanges extracted from Reddit comment chains .
Outcome:	The proposed model can generate more relevant, contentful and context-consistent responses than baseline systems.

Copied to clipboard

Jingwen Pu, Mingjun Shi, Xinrui Ren, Yizhe Wang, Xinyu Zhang, Zhaokun Wang, Kun She

Optimus: Organizing Sentences via Pre-trained Modeling of a Latent Space (2020.emnlp-main)

Challenge:	Existing methods for large-scale modeling memorize sensitive information . however, they are limited in real-world scenarios and require updating parameters .
Approach:	They propose a training-free, plug-and-play inference-time unlearning strategy that uses a probe to detect queries involving forgettable concepts and applies entropy-guided decoding to suppress target knowledge.
Outcome:	Experiments on MUSE, RWKU, and WMDP datasets show that SEGUE outperforms existing methods.

Copied to clipboard

Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, Jianfeng Gao

Joint Embedding of Words and Labels for Text Classification (P18-1)

Challenge:	Existing models for language understanding and understanding can be trained to provide contextualized representations of words based on text data.
Approach:	They propose a large-scale language VAE model Optimus that is pre-trained on large text corpus and fine-tuned for various language generation and understanding tasks.
Outcome:	The proposed model achieves new state-of-the-art on VAE language modeling benchmarks.

Copied to clipboard

Guoyin Wang, Chunyuan Li, Wenlin Wang, Yizhe Zhang, Dinghan Shen, Xinyuan Zhang, Ricardo Henao, Lawrence Carin

MMAU: A Holistic Benchmark of Agent Capabilities Across Diverse Domains (2025.findings-naacl)

Challenge:	Existing approaches to text classification use word embeddings to capture semantic regularities between words.
Approach:	They propose to view text classification as a label-word joint embedding problem . they use a framework that measures compatibility between text sequences and labels .
Outcome:	The proposed framework outperforms the state-of-the-art methods on large text datasets.

Copied to clipboard

Guoli Yin, Haoping Bai, Shuang Ma, Feng Nan, Yanchao Sun, Zhaoyang Xu, Shen Ma, Jiarui Lu, Xiang Kong, Aonan Zhang, Dian Ang Yap, Yizhe Zhang, Karsten Ahnert, Vik Kamath, Mathias Berglund, Dominic Walsh, Tobias Gindele, Juergen Wiest, Zhengfeng Lai, Xiaoming Simon Wang, Jiulong Shan, Meng Cao, Ruoming Pang, Zirui Wang

Challenge:	Existing benchmarks focus on specific application scenarios, emphasizing task completion but failing to dissect the underlying skills that drive these outcomes.
Approach:	They propose a Massive Multitask Agent Understanding benchmark that evaluates LLMs across five domains and offline tasks.
Outcome:	The Massive Multitask Agent Understanding (MMAU) benchmark evaluates models across five domains including Tool-use, Directed Acyclic Graph (DAG) QA, Data Science and Machine Learning coding, Contest-level programming and Mathematics.

Generating a Common Question from Multiple Documents using Multi-source Encoder-Decoder Models (D19-56)

Copied to clipboard

Woon Sang Cho, Yizhe Zhang, Sudha Rao, Chris Brockett, Sungjin Lee

Bridging the Training-Inference Gap for Dense Phrase Retrieval (2022.findings-emnlp)

Challenge:	Ambiguous user queries can result in multiple topics being retrieved from search engines.
Approach:	They propose a task of generating a common question from multiple documents by training an RNN-based single encoder-decoder generator from document pairs and then a model that aggregates these word distributions to generate a question.
Outcome:	The proposed model significantly outperforms existing models when evaluated using automated metrics and human judgments on the MS-MARCO-QA dataset.

Copied to clipboard

Gyuwan Kim, Jinhyuk Lee, Barlas Oguz, Wenhan Xiong, Yizhe Zhang, Yashar Mehdad, William Yang Wang

CriticLean: Critic-Guided Reinforcement Learning for Mathematical Formalization (2026.acl-long)

Challenge:	Existing methods for building dense retrievers are often misaligned and do not reflect retrieval scenario at inference time.
Approach:	They propose a way to validate dense retrievers using a small subset of the entire corpus.
Outcome:	The proposed model improves top-1 phrase retrieval accuracy by 2 3 points and top-20 passage retrieval by 2 4 points for open-domain question answering.

Copied to clipboard

Zhongyuan Peng, Yifan Yao, Kaijing Ma, Shuyue Guo, Yizhe Li, Yichi Zhang, Chenchen Zhang, Yifan Zhang, Zhouliang Yu, Luming Li, Minghao Liu, Yihang Xia, Jiawei Shen, Yuchen Wu, Yixin Cao, Zhaoxiang Zhang, Wenhao Huang, Jiaheng Liu, Ge Zhang

Reasoning

Challenge:	Existing approaches to formalizing mathematical statements face limitations in accuracy, especially in the context of complex, highlevel problems that involve sophisticated mathematical reasoning.
Approach:	They propose a CriticLean framework that elevates the role of the critic from a passive validator to an active learning component and introduce a benchmark to measure models’ ability to distinguish semantically correct from incorrect formalizations.
Outcome:	The proposed framework outperforms open- and closed-source benchmarks and shows that it significantly outperformed existing models.

ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities (2025.findings-naacl)

Copied to clipboard

Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Haoping Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, Zirui Wang, Ruoming Pang

Probing the Multi-turn Planning Capabilities of LLMs via 20 Question Games (2024.acl-long)

Challenge:	Recent advances in large language models have led to a growing interest in tool assisted LLMs . toolSandbox includes stateful tool execution, implicit state dependencies between tools .
Approach:	a new tool-based evaluation tool is released to help LLMs evaluate their tool-use capabilities. a tool-driven evaluation tool includes stateful tool execution, implicit state dependencies between tools and a built-in user simulator.
Outcome:	the toolSandbox evaluation benchmark shows that open source and proprietary models have a performance gap . the benchmarks show that even the most capable LLMs are challenged by state dependent tasks .

Copied to clipboard

Yizhe Zhang, Jiarui Lu, Navdeep Jaitly

Natural Language Interfaces

Challenge:	Large language models (LLMs) are effective at answering clear questions but when faced with ambiguous queries they act unpredictably and produce incorrect outputs.
Approach:	They propose to use a surrogate problem to assess an LLMs’s ability to deduce an entity unknown to itself, but revealed to a judge, by asking the judge a series of queries.
Outcome:	The proposed model outperforms human players on the entity-deducing task by a large margin.

Finetuning Pretrained Transformers into RNNs (2021.emnlp-main)

Copied to clipboard

Jungo Kasai, Hao Peng, Yizhe Zhang, Dani Yogatama, Gabriel Ilharco, Nikolaos Pappas, Yi Mao, Weizhu Chen, Noah A. Smith

Dialogue Response Ranking Training with Large-Scale Human Feedback Data (2020.emnlp-main)

Challenge:	Efficient transformers outperform recurrent neural networks in natural language generation, but this comes with significant computational cost and memory footprint during generation.
Approach:	They propose to convert a pretrained transformer into its efficient recurrent counterpart, improving efficiency while maintaining accuracy.
Outcome:	The proposed transformers outperform recurrent neural networks in natural language generation but come with significant computational and memory footprint during generation.

Copied to clipboard

Xiang Gao, Yizhe Zhang, Michel Galley, Chris Brockett, Bill Dolan

A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation (2022.acl-long)

Challenge:	Existing open-domain dialog models can minimize the perplexity of target human responses . however, some human responses are more engaging than others, spawning more followup interactions .
Approach:	They train open-domain dialog models to minimize perplexity of target human responses . they use social media feedback data to train models to predict engaging dialog turns .
Outcome:	The proposed model outperforms existing models on 133M human feedback pairs . it also outperformed the conventional dialog perplexity baseline model .

Copied to clipboard

Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, Bill Dolan

Reasoning

Challenge:	Existing work on pre-trained generative models often fails to detect non-existent or incorrect content . Existing studies have attempted to detect hallucinations based on oracle references .
Approach:	They propose a token-level, reference-free hallucination detection task based on Wikipedia annotations to detect non-existent or incorrect content.
Outcome:	The proposed task is token-level, reference-free hallucination detection task and dataset . authors argue that the proposed task can be used in real-time to detect hallucines .

Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling (2024.acl-long)

Copied to clipboard

Pratyush Maini, Skyler Seto, Richard Bai, David Grangier, Yizhe Zhang, Navdeep Jaitly

Automatic Document Sketching: Generating Drafts from Analogous Texts (2021.findings-acl)

Challenge:	Large language model pre-training is infeasible due to the large compute costs and duration associated with pre- training and the impending scarcity of high-quality data on the web.
Approach:	They propose to use an off-the-shelf instruction-tuned model prompted to paraphrase documents on the web in specific styles such as “like Wikipedia” or in “question-answer format” to jointly pre-train LLMs on real and synthetic rephrases.
Outcome:	The proposed model speeds up pre-training by 3x on the C4 dataset, and improves perplexity by 50% on average across different subsets of the Pile.

Copied to clipboard

Zeqiu Wu, Michel Galley, Chris Brockett, Yizhe Zhang, Bill Dolan

POINTER: Constrained Progressive Text Generation via Insertion-based Generative Pre-training (2020.emnlp-main)

Challenge:	Large pre-trained language models have made it possible to make high-quality predictions on how to add or change a sentence in a document.
Approach:	They propose a task to generate entire draft documents for the writer to review and revise.
Outcome:	The proposed model can make high-quality predictions on how to add or change a sentence in a document, but it lacks the branching factor to offer useful editing suggestions at a global or document level.

Copied to clipboard

Yizhe Zhang, Guoyin Wang, Chunyuan Li, Zhe Gan, Chris Brockett, Bill Dolan

Breaking the Stage Barrier: A Novel Single-Stage Approach to Long Context Extension for Large Language Models (2025.coling-main)

Challenge:	Existing pre-trained language models cannot be directly employed to generate text under specified lexical constraints.
Approach:	They propose a method for insertion-based text generation that inserts tokens between existing tokens in a parallel manner.
Outcome:	The proposed method is intuitive and interpretable on Wikipedia and Yelp datasets.

Copied to clipboard

Haoran Lian, Junmin Chen, Wei Huang, Yizhe Xiong, Wenping Hu, Guiguang Ding, Hui Chen, Jianwei Niu, Zijia Lin, Fuzheng Zhang, Di Zhang

Domain Adaptive Text Style Transfer (D19-1)

Challenge:	Recent studies show that Large language models struggle with handling long token sequences due to limited training context size.
Approach:	They propose a single-stage continual pretraining method to equip LLMs with long context modeling capabilities.
Outcome:	The proposed method outperforms existing methods on 4 language modeling benchmarks.

Copied to clipboard

Dianqi Li, Yizhe Zhang, Zhe Gan, Yu Cheng, Chris Brockett, Bill Dolan, Ming-Ting Sun

DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs (2025.emnlp-main)

Challenge:	Text style transfer without parallel data is a promising method for learning, but in the scenario where less data is available, it may yield poor performance.
Approach:	They propose to leverage available data to learn domain-adaptive text style transfer models . they evaluate two style transfer tasks where only limited non-parallel data is available .
Outcome:	The proposed models learn from the source domain to: (i) distinguish stylized information and generic content information; (ii) maximally preserve content information and (iv) adaptively transfer the styles in a domain-aware manner.

Copied to clipboard

Minxuan Lv, Zhenpeng Su, Leiyu Pan, Yizhe Xiong, Zijia Lin, Hui Chen, Wei Zhou, Jungong Han, Guiguang Ding, Wenwu Ou, Di Zhang, Kun Gai, Songlin Hu

Improving Disentangled Text Representation Learning with Information-Theoretic Guidance (2020.acl-main)

Challenge:	Existing sparsification methods like pruning can lose model knowledge through parameter removal.
Approach:	They propose a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks.
Outcome:	The proposed approach achieves superior performance across language modeling and downstream tasks under equivalent computational constraints.

Copied to clipboard

Pengyu Cheng, Martin Renqiang Min, Dinghan Shen, Christopher Malon, Yizhe Zhang, Yitong Li, Lawrence Carin

Representation Learning

Challenge:	Disentangled representation learning (DRL) maps different aspects of data into distinct and independent low-dimensional latent vector spaces.
Approach:	They propose a method that manifests disentangled representations of text without supervision on semantics by minimizing the upper bound between style and content.
Outcome:	The proposed method improves on conditional text generation and text-style transfer tasks and improves style preservation.

Baseline Needs More Love: On Simple Word-Embedding-Based Models and Associated Pooling Mechanisms (P18-1)

Copied to clipboard

Dinghan Shen, Guoyin Wang, Wenlin Wang, Martin Renqiang Min, Qinliang Su, Yizhe Zhang, Chunyuan Li, Ricardo Henao, Lawrence Carin

MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning (2026.acl-long)

Challenge:	Existing deep learning architectures to model compositionality in text sequences require a large number of parameters and expensive computations.
Approach:	They propose two additional pooling strategies over word embeddings for improved interpretability and hierarchical pooling for spatial (n-gram) information within text sequences.
Outcome:	The proposed pooling strategies improve interpretability and preserve spatial (n-gram) information within text sequences.

Copied to clipboard

Yizhe Zeng, Wei Zhang, Yunpeng Li, Juxin Xiao, Xiao Wang, Yuling Liu

Robustness in NLP

Challenge:	Existing CoT backdoor attacks manipulate intermediate reasoning steps to steer the model toward incorrect answers, but these corrupted reasoning traces are readily detected by prevalent process-monitoring defenses.
Approach:	They propose a backdoor attack that exploits the model's post-output space to preserve clean CoTs while selectively steering the final answer toward a specific target.
Outcome:	Experiments show that MirageBD achieves over 90% success rate across four datasets and five models with a poison ratio of only 5%.

Divide-or-Conquer? Which Part Should You Distill Your LLM? (2024.findings-emnlp)

Copied to clipboard

Zhuofeng Wu, Richard Bai, Aonan Zhang, Jiatao Gu, V.G.Vinod Vydiswaran, Navdeep Jaitly, Yizhe Zhang

Eliciting In-context Retrieval and Reasoning for Long-context Large Language Models (2025.findings-acl)

Challenge:	Recent studies have shown that Large Language Models (LLMs) can solve reasoning tasks better when they are encouraged to solve subtasks of the main task first.
Approach:	They propose a strategy that breaks down reasoning tasks into a problem decomposition phase and a solution phase and propose 'smaller' models that can achieve good generalization.
Outcome:	The proposed approach outperforms a single stage solution in two tasks and their impact on reasoning outcomes and inference cost.

Copied to clipboard

Yifu Qiu, Varun R. Embar, Yizhe Zhang, Navdeep Jaitly, Shay B Cohen, Benjamin Han

Towards More Efficient Insertion Transformer with Fractional Positional Encoding (2023.eacl-main)

Challenge:	Existing benchmarks like LOFT often overestimate LCLM performance by providing overly simplified contexts.
Approach:	They propose to use retrieval-attention-probing to filter and de-noise long contexts during decoding and joint retrieval head training alongside the generation head to improve LCLM performance.
Outcome:	The proposed approach outperforms RAG and GPT-4-Turbo on most tasks despite being a much smaller model.

Copied to clipboard

Zhisong Zhang, Yizhe Zhang, Bill Dolan

Structuring Latent Spaces for Stylized Response Generation (D19-1)

Challenge:	Empirical studies on text generation tasks demonstrate the effectiveness of insertion-based models.
Approach:	They propose a reusable positional encoding scheme for insertion transformers that allows reusing representations calculated in previous steps.
Outcome:	Empirical studies show that the proposed model reduces the time required to generate a token and improves decoding efficiency.

Copied to clipboard

Xiang Gao, Yizhe Zhang, Sungjin Lee, Michel Galley, Chris Brockett, Jianfeng Gao, Bill Dolan

Mis-prompt: Benchmarking Large Language Models for Proactive Error Handling (2025.acl-long)

Challenge:	Existing methods for generating responses in a targeted style are limited by the lack of parallel data.
Approach:	They propose a method that bridges conversation modeling and non-parallel style transfer by sharing a structured latent space.
Outcome:	The proposed system generates responses of the targeted style and outperforms baselines without sacrificing appropriateness.

Copied to clipboard

Jiayi Zeng, Yizhe Feng, Mengliang He, Wenhui Lei, Wei Zhang, Zeming Liu, Xiaoming Shi, Aimin Zhou

On the Limited Generalization Capability of the Implicit Reward Model Induced by Direct Preference Optimization (2024.findings-emnlp)

Challenge:	Current error-handling works are performed in a passive manner, with explicit error- handling instructions.
Approach:	They propose a new benchmark to analyze LLMs' performance on a mis-prompt benchmark and a dataset to promote further research.
Outcome:	The proposed benchmark shows that current LLMs show poor performance on proactive error handling, and that SFT improves on error handling instances.

Copied to clipboard

Yong Lin, Skyler Seto, Maartje Ter Hoeve, Katherine Metcalf, Barry-John Theobald, Xuan Wang, Yizhe Zhang, Chen Huang, Tong Zhang

Challenge:	Reinforcement Learning from Human Feedback (RLHF) is an effective approach for aligning language models to human preferences.
Approach:	They compare the accuracy of DPORM and EXRM with a reward function for scoring human preferences.
Outcome:	The proposed methods can approximate an EXRM on the limit infinite samples, but it is unclear how effective they are in practice.

Contextual Text Style Transfer (2020.findings-emnlp)

Copied to clipboard

Yu Cheng, Zhe Gan, Yizhe Zhang, Oussama Elachqar, Dianqi Li, Jingjing Liu

Improving Text Generation with Student-Forcing Optimal Transport (2020.emnlp-main)

Challenge:	Existing methods for text style transfer are limited by the lack of parallel data.
Approach:	They propose a task to translate a sentence into a desired style with its surrounding context taken into account.
Outcome:	The proposed model outperforms state-of-the-art methods across style accuracy, content preservation and contextual consistency metrics.

Copied to clipboard

Jianqiao Li, Chunyuan Li, Guoyin Wang, Hao Fu, Yuhchen Lin, Liqun Chen, Yizhe Zhang, Chenyang Tao, Ruiyi Zhang, Wenlin Wang, Dinghan Shen, Qian Yang, Lawrence Carin

Improving Textual Network Embedding with Global Attention via Optimal Transport (P19-1)

Challenge:	Maximum likelihood estimation (MLE) is used to train models, but during testing, the model is conditioned on previously generated tokens, resulting in exposure bias.
Approach:	They propose to use optimal transport to match the sequences generated in MLE and test modes to reduce exposure bias.
Outcome:	The proposed method is validated on machine translation, text summarization, and text generation tasks.

Copied to clipboard

Liqun Chen, Guoyin Wang, Chenyang Tao, Dinghan Shen, Pengyu Cheng, Xinyuan Zhang, Wenlin Wang, Yizhe Zhang, Lawrence Carin

Representation Learning

Challenge:	Existing methods for learning textual network embeddings are noisy and sparse.
Approach:	They propose to use text-based attention parsing to learn context-aware network embeddings.
Outcome:	The proposed model outperforms state-of-the-art methods in a number of domains.

Linearizing Transformer with Key-Value Memory (2022.emnlp-main)

Copied to clipboard

Yizhe Zhang, Deng Cai

EvoWiki: Evaluating LLMs on Evolving Knowledge (2025.acl-long)

Challenge:	Efficient transformer variants with linear time complexity have been developed to mitigate the quadratic computational overhead of the vanilla transformer.
Approach:	They propose a linear time complexity transformer variant that reduces the quadratic computational overhead of the vanilla transformer by using a recurrent-style incremental computation similar to kernel-based transformers.
Outcome:	The proposed method reduces the performance gap while achieving the same efficiency even with short generation.

Copied to clipboard

Wei Tang, Yixin Cao, Yang Deng, Jiahao Ying, Bo Wang, Yizhe Yang, Yuyue Zhao, Qi Zhang, Xuanjing Huang, Yu-Gang Jiang, Yong Liao

Challenge:	Existing knowledge evolution benchmarks are static and fail to capture the evolving nature of LLMs and knowledge.
Approach:	They propose an evolving dataset that categorizes information into stable, evolved, and uncharted states.
Outcome:	The proposed dataset is auto-updatable and enables evaluation of continuously changing knowledge and newly released LLMs.

Towards Generating Long and Coherent Text with Multi-Level Latent Variable Models (P19-1)

Copied to clipboard

Dinghan Shen, Asli Celikyilmaz, Yizhe Zhang, Liqun Chen, Xin Wang, Jianfeng Gao, Lawrence Carin

Jointly Optimizing Diversity and Relevance in Neural Response Generation (N19-1)

Challenge:	Variational autoencoders (VAEs) have received much attention as an end-to-end architecture for text generation with latent variables.
Approach:	They propose to leverage several multi-level structures to learn a variational autoencoder model for generating long, and coherent text.
Outcome:	The proposed model produces more coherent and less repetitive long text compared to baselines and mitigates posterior collapse issue.

Copied to clipboard

Xiang Gao, Sungjin Lee, Yizhe Zhang, Chris Brockett, Michel Galley, Jianfeng Gao, Bill Dolan

Contrastive Multi-document Question Generation (2021.eacl-main)

Challenge:	Recent neural conversation models often generate bland and generic responses . however, the improvement often comes at the cost of decreased relevance .
Approach:	They propose a spacefusion model to jointly optimize diversity and relevance that fuses the latent space of a sequence-to-sequence model and that of an autoencoder model by leveraging novel regularization terms.
Outcome:	The proposed model improves diversity and relevance compared to baselines in both diversity and diversity.

Copied to clipboard

Woon Sang Cho, Yizhe Zhang, Sudha Rao, Asli Celikyilmaz, Chenyan Xiong, Jianfeng Gao, Mengdi Wang, Bill Dolan