Papers by Pengyu Zhang
LECO: Improving Early Exiting via Learned Exits and Comparison-based Exiting Mechanism (2023.acl-srw)
Copied to clipboard
| Challenge: | Recent work on dynamic early exiting has neglected the intermediate exits’ architectural designs. |
| Approach: | They propose a framework for learning exits and COmparison-based early exiting to improve PTMs’ early exit performance. |
| Outcome: | The proposed framework achieves the SOTA performance on multi-exit BERT training and dynamic early exiting on pre-trained models. |
Fusion Training for Mathematical Generalization in Large Language Models (2026.acl-srw)
Copied to clipboard
| Challenge: | Existing efforts to improve reasoning efficiency have focused on switching between simple queries and complex problems. |
| Approach: | They analyze the effects of the training schedule and data ratio between thinking and non-thinking modes and construct a benchmark to test their theory. |
| Outcome: | The proposed model unifies a thinking mode and a non-thinking mode within a single model. |
UnifiedMLLM: Enabling Unified Representation for Multi-modal Multi-tasks With Large Language Model (2025.findings-naacl)
Copied to clipboard
Zhaowei Li, Wei Wang, YiQing Cai, Qi Xu, Pengyu Wang, Dong Zhang, Hang Song, Botian Jiang, Zhida Huang, Tao Wang
| Challenge: | Representative models like LLaVA and MiniGPT-4 have great capabilities in various tasks. |
| Approach: | They propose a unified model to represent various multi-modal tasks using a single representation. |
| Outcome: | The proposed model outperforms existing models in a variety of tasks while maintaining generality and scalability. |
SeqXGPT: Sentence-Level AI-Generated Text Detection (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for sentence-level AIGT detection are weak . large language models (LLMs) can generate human-like content . |
| Approach: | They propose a sentence-level AIGT detection challenge using LLMs as log probability lists . they propose 'check' GPT' method that uses log probability list features to detect AIGT . |
| Outcome: | The proposed method surpasses baseline methods in sentence- and document-level detection challenges. |
OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding (2026.acl-long)
Copied to clipboard
Deming Ding, Shichun Liu, Enhui Yang, Jiahang Lin, Ziying Chen, Shihan Dou, Honglin Guo, Weiyu Cheng, Pengyu Zhao, Chengjun Xiao, Qunhong Zeng, Qi Zhang, Xuanjing Huang, Qidi Xu, Tao Gui
| Challenge: | coding scaffolds that follow heterogeneous instructions remain under-examined in software engineering . coding models are capable software agents, but their ability to follow constraints remains under-explored . |
| Approach: | They introduce OctoBench, which benchmarks scaffold-aware instruction following in agentic coding. |
| Outcome: | The proposed benchmark aims to accelerate the development of more scaffold-aware agents. |
DemonAgent: Dynamically Encrypted Multi-Backdoor Implantation Attack on LLM-based Agent (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new method for detecting advanced backdoors is proposed to bypass safety audits. |
| Approach: | They propose a backdoor implantation strategy that introduces dynamic encryption to bypass safety audits. |
| Outcome: | The proposed method achieves an attack success rate approaching 100% while maintaining a detection rate of 0%. |
InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance (2024.emnlp-main)
Copied to clipboard
Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, Xipeng Qiu
| Challenge: | Existing methods for enhancing harmlessness and helpfulness of large language models (LLMs) involve complex and resource-intensive training processes. |
| Approach: | They propose a method that decouples harmlessness from helpfulness during inference phase. |
| Outcome: | The proposed method significantly reduces the attack success rate (ASR) of harmful instructions and jailbreak instructions while maintaining almost unchanged performance in downstream tasks. |
MARCH: Multi-Agent Reinforced Check for Hallucination (2026.acl-long)
Copied to clipboard
Zhuo Li, Yupeng Zhang, Pengyu Cheng, Jiajun Song, Mengyu Zhou, Hao Li, Shujie Hu, Yu Qin, null Erchao.zec, Xiaoxi Jiang, null Guanjunjiang
| Challenge: | Existing methods to detect hallucinations suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. |
| Approach: | They propose a framework that enforces rigorous factual alignment by leveraging deliberate *information asymmetry* by combining a pipeline of three specialized agents: a Solver, a Proposer, and a Checker. |
| Outcome: | Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucinism rates. |
CARD: Cross-modal Agent Framework for Generative and Editable Residential Design (2025.emnlp-main)
Copied to clipboard
| Challenge: | Architectural design automation has made significant progress, but the complexity of open-world environments makes residential design a challenging task. |
| Approach: | They propose a framework that leverages a system of specialized cross-modal agents to adapt to open-world residential design. |
| Outcome: | The proposed framework enables users to generate and edit residential design without requiring specialized expertise. |
PhysPRM: A Generative Process Reward Model with Fine-grained Diagnosis for Physics Problem Solving (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing Large Language Models (LLMs) struggle with physics problem solving due to difficulties in decoding implicit constraints and maintaining physical consistency. |
| Approach: | They propose a Generative PRM that treats evaluation as a generative task . it produces fine-grained diagnoses comprising critiques, final judgments, and specific error types . |
| Outcome: | The proposed model improves performance across seven benchmarks in Best-of-N and critique refinement strategies. |
FloorPlan-LLaMa: Aligning Architects’ Feedback and Domain Knowledge in Architectural Floor Plan Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing evaluation methods for floor plan generation rely on statistical metrics like FID, GED, and PSNR, which fail to evaluate using domain knowledge. |
| Approach: | They propose to use a first floor plan dataset to train a floor plan generation model based on a multi-dimensional preference score and a textual analysis to integrate architects’ professional expertise and preferences. |
| Outcome: | The proposed model outperforms baseline models in text-conditional and class-condition tasks and is more rational and aligns better with human preferences. |
MetaAlign: Align Large Language Models with Diverse Preferences during Inference Time (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to align large language models with human preferences often result in a static alignment that cannot account for the diversity of human preferences in practical applications. |
| Approach: | They propose a method to help large language models dynamically align with various explicit or implicit preferences specified at inference time. |
| Outcome: | The proposed method can help LLMs dynamically align with various explicit or implicit preferences specified at the inference stage, validating the feasibility of MetaAlign. |
WildGraphBench: Benchmarking GraphRAG with Wild-Source Corpora (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for Graph-based Retrieval-Augmented Generation (GraphRAG) rely on short, curated passages as external knowledge, failing to adequately evaluate systems in realistic settings involving long contexts and large-scale heterogeneous documents. |
| Approach: | They propose a benchmark to assess GraphRAG performance in the wild using Wikipedia's unique structure where cohesive narratives are grounded in long and heterogeneous external reference documents. |
| Outcome: | Experiments with articles across 12 top-level topics show that GraphRAG performs better in the wild than existing methods. |
Decoupled Proxy Alignment: Mitigating Language Prior Conflict for Multimodal Alignment in MLLMs (2025.findings-emnlp)
Copied to clipboard
Chenkun Tan, Pengyu Wang, Shaojun Zhou, Botian Jiang, Zhaowei Li, Dong Zhang, Xinghao Wang, Yaqian Zhou, Xipeng Qiu
| Challenge: | Recent advances in multimodal large language models focus on improving performance . however, language prior conflict leads to suboptimal vision-language alignment . |
| Approach: | They propose a method to decouple the alignment process from language prior interference . they use a proxy LLM to detach from language interference during pretraining . |
| Outcome: | The proposed method improves training performance and generalizes training data. |
AGTAO: Robust and Stabilized LLM Unlearning via Adversarial Gating Training with Adaptive Orthogonality (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) unintentionally memorize sensitive data, posing privacy and security risks. |
| Approach: | They propose a framework that reconciles unlearning efficacy and utility preservation by using a latent-space gating mechanism to simulate internal recovery attempts. |
| Outcome: | The proposed framework achieves superior trade-off between unlearning efficacy and model utility. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |
SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing multi-modal large language models typically adopt the cascade paradigm, preventing inter-modal knowledge transfer. |
| Approach: | They propose a large language model with intrinsic cross-modal conversational abilities . they construct a cross-text speech instruction dataset and employ a three-stage training strategy . |
| Outcome: | The proposed model can follow cross-modal human instructions and handle multiple modalities with one model. |
Impact of Evaluation Methodologies on Code Summarization (2022.acl-long)
Copied to clipboard
| Challenge: | Existing evaluation methodologies for code summarization tasks do not consider timestamps of code and comments. |
| Approach: | They propose a time-segmented evaluation methodology for code summarization that considers timestamps of code and comments during evaluation. |
| Outcome: | The proposed evaluation methodology compares with other evaluation methodologies that have been widely used. |
Reviving Cultural Heritage: A Novel Approach for Comprehensive Historical Document Restoration (2025.acl-long)
Copied to clipboard
Yuyi Zhang, Peirong Zhang, Zhenhua Yang, Pengyu Yan, Yongxin Shi, Pengwei Liu, Fengjun Guo, Lianwen Jin
| Challenge: | Existing methods for historical document restoration focus on single modality or limited-size restoration, failing to meet practical needs. |
| Approach: | They propose a full-page HDR dataset and an automated HDR solution to replace manual restoration methods. |
| Outcome: | The proposed solution improves OCR accuracy from 46.83% to 84.05% when processing severely damaged documents, with enhancement to 94.25% through human-machine collaboration. |
Improving Disentangled Text Representation Learning with Information-Theoretic Guidance (2020.acl-main)
Copied to clipboard
Pengyu Cheng, Martin Renqiang Min, Dinghan Shen, Christopher Malon, Yizhe Zhang, Yitong Li, Lawrence Carin
| 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. |
DecIF: Improving Instruction-Following through Decomposition (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to obtain high-quality instruction-following data rely heavily on existing documents and existing methods. |
| Approach: | They propose a data synthesis framework, DecIF, which automatically generates accurate and diverse instruction-following data from scratch for supervised fine-tuning and reinforcement learning. |
| Outcome: | Extensive experiments show that the proposed framework can synthesize accurate instruction-following data for both SFT and RL paradigms compared to baselines. |
Learning Compressed Sentence Representations for On-Device Text Processing (P19-1)
Copied to clipboard
Dinghan Shen, Pengyu Cheng, Dhanasekar Sundararaman, Xinyuan Zhang, Qian Yang, Meng Tang, Asli Celikyilmaz, Lawrence Carin
| Challenge: | Existing methods for learning sentence embeddings assume they are continuous and real-valued. |
| Approach: | They propose four different strategies to transform continuous and generic sentence embeddings into a binarized form while preserving their rich semantic information. |
| Outcome: | The proposed methods reduce storage requirements by over 98% and improve performance on downstream tasks. |
UnifiedVisual: A Framework for Constructing Unified Vision-Language Datasets (2025.emnlp-main)
Copied to clipboard
Pengyu Wang, Shaojun Zhou, Chenkun Tan, Xinghao Wang, Wei Huang, Zhen Ye, Zhaowei Li, Botian Jiang, Dong Zhang, Xipeng Qiu
| Challenge: | Existing datasets address understanding and generation in isolation, limiting the performance of unified vision large language models. |
| Approach: | They propose a dataset that facilitates mutual enhancement between multimodal understanding and generation. |
| Outcome: | The proposed framework integrates diverse visual and textual inputs and outputs, enabling comprehensive cross-modal reasoning and precise text-to-image alignment. |
Improving Textual Network Embedding with Global Attention via Optimal Transport (P19-1)
Copied to clipboard
Liqun Chen, Guoyin Wang, Chenyang Tao, Dinghan Shen, Pengyu Cheng, Xinyuan Zhang, Wenlin Wang, Yizhe Zhang, Lawrence Carin
| 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. |