Papers by Fei Ma
Are Large Language Models Reliable Reviewers? A Benchmark for Error Detection in Financial Documents (2026.findings-acl)
Copied to clipboard
Ying He, Zhouhong Gu, Zhecheng Hu, Yubo Zhou, Hao Shen, Jiaqing Liang, Zhaoqian Dai, Ma Shuguang, Fei Yu, Yanghua Xiao, Zhixu Li
| Challenge: | Existing LLMs struggle to identify errors in financial documents, a study shows . 18% of financial practitioners make errors daily, one-third make errors several times weekly, and 59% make errors multiple times monthly. |
| Approach: | They introduce FinED-Bench, a publicly available Benchmark for financial error detection . it covers nine real-world financial scenarios and includes over 900 documents in 2025 . supervised fine-tuning can significantly improve the performance of weaker LLMs, they show . |
| Outcome: | The proposed benchmark covers nine real-world financial scenarios and includes over 900 documents reported in 2025 that are unseen by existing language models. |
The Mirage of Model Editing: Revisiting Evaluation in the Wild (2025.acl-long)
Copied to clipboard
| Challenge: | despite near-perfect results, effectiveness of model editing in real-world applications remains unclear. |
| Approach: | They propose QAEdit and WILD to better reflect real-world use of model editing . they propose a benchmark aligned with widely used question answering datasets and a task-agnostic evaluation framework . |
| Outcome: | The proposed QAEdit benchmark and WILD evaluation framework show that current models perform worse than previously reported. |
Robustness via Referencing: Defending against Prompt Injection Attacks by Referencing the Executed Instruction (2026.findings-acl)
Copied to clipboard
Yulin Chen, Haoran Li, Yuan Sui, Yue Liu, Yufei He, Xiaoling Bai, Chi Fei, Li Yabo, Haozhe Ma, Yangqiu Song, Bryan Hooi
| Challenge: | Prompt injection attacks manipulate large language models (LLMs) by misleading them to deviate from the original input instructions and execute maliciously injected instructions. |
| Approach: | They propose a prompt injection defense method that suppresses the model's instruction-following tendencies rather than suppressing them. |
| Outcome: | The proposed method outperforms prompt-engineering-based approaches and fine-tuning methods and reduces the ASR to nearly 0% in some scenarios. |
PairCoder: Pair Programming-Inspired Two-Agent Collaboration for Code Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing multi agent frameworks for large language models are brittle on code generation tasks. |
| Approach: | They propose a framework that brings pair programming to autonomous LLM collaboration. |
| Outcome: | Using PairCoder, large language models achieve better results on code generation tasks and reduce token usage by 40% to 70% on eight representative backbones. |
Automated Fine-Grained Mixture-of-Experts Quantization (2025.findings-acl)
Copied to clipboard
| Challenge: | specialized quantization framework for Mixture of Experts architectures is inadequate for model compression. |
| Approach: | They propose a specialized quantization framework for Mixture of Experts architectures . they find that expert networks exhibit distinctive channel-wise outlier distributions ." |
| Outcome: | The proposed framework improves on the Mixtral-8x7b-v0.1 architecture while maintaining minimal computational overhead. |
PFDial: A Structured Dialogue Instruction Fine-tuning Method Based on UML Flowcharts (2025.findings-acl)
Copied to clipboard
Ming Zhang, Yuhui Wang, Yujiong Shen, Tingyi Yang, Changhao Jiang, Yilong Wu, Shihan Dou, Qinhao Chen, Zhiheng Xi, Zhihao Zhang, Yi Dong, Zhen Wang, Zhihui Fei, Mingyang Wan, Tao Liang, Guojun Ma, Qi Zhang, Tao Gui, Xuanjing Huang
| Challenge: | Large Language Models (LLMs) have shown remarkable progress in dialogue and reasoning, but they struggle to solve strictly constrained dialogue tasks. |
| Approach: | They construct a dataset that contains 12,705 high-quality Chinese dialogue instructions from 440 flowcharts containing 5,055 process nodes. |
| Outcome: | The proposed model outperforms GPT-4o models on backward transitions and outperformed GPT-42 models on the same dataset. |
Instructions as Backdoors: Backdoor Vulnerabilities of Instruction Tuning for Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | et al., 2021) show that instruction models can be trained on crowdsourced datasets with task instructions to achieve superior performance. |
| Approach: | They examine security concerns of emergent instruction tuning paradigm that models are trained on crowdsourced datasets with task instructions to achieve superior performance. |
| Outcome: | The proposed model can achieve 90% success rate across four commonly used datasets. |
KS-Lottery: Finding Certified Lottery Tickets for Multilingual Transfer in Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that a small subset of parameters is highly effective in fine-tuning . prior work shows that there are a few additional parameters corresponding to an intrinsic dimension in a well-trained Large Language Model. |
| Approach: | They propose a method to identify a small subset of LLM parameters highly effective in multilingual fine-tuning. |
| Outcome: | The proposed method can find the certified winning tickets in the embedding layer, and fine-tuning on the found parameters is guaranteed to perform as well as full fine- tuning. |
Second Language (Arabic) Acquisition of LLMs via Progressive Vocabulary Expansion (2025.acl-long)
Copied to clipboard
Jianqing Zhu, Huang Huang, Zhihang Lin, Juhao Liang, Zhengyang Tang, Khalid Almubarak, Mosen Alharthi, Bang An, Juncai He, Xiangbo Wu, Fei Yu, Junying Chen, Ma Zhuoheng, Yuhao Du, He Zhang, Saied Alshahrani, Emad A. Alghamdi, Lian Zhang, Ruoyu Sun, Haizhou Li, Benyou Wang, Jinchao Xu
| Challenge: | In the evolving landscape of large language models, the predominant focus has been on English and Chinese. |
| Approach: | They propose to utilize Arabic-specific vocabulary in the tokenizer to accelerate decoding. |
| Outcome: | The proposed model achieves decent performance comparable to the best Arabic LLMs across various Arabic benchmarks. |
Don’t Tell the Answer, Truly Guide the Reasoning During RL Rollouts (2026.findings-acl)
Copied to clipboard
Xinyi Wang, Jinyi Han, Zishang Jiang, Tingyun li, Jiaqing Liang, Sihang Jiang, Zhaoqian Dai, Ma Shuguang, Fei Yu, Yanghua Xiao
| Challenge: | Existing methods such as GRPO often break down when task difficulty exceeds the model’s capacity, resulting in sparse rewards and inefficient training. |
| Approach: | They propose to measure the compatibility between external guidance and a model's intrinsic policy by introducing an adaptive framework to enhance reasoning performance while explicitly preserving high Affinity. |
| Outcome: | The proposed framework outperforms baseline models while maintaining high Affinity. |
Speech-Text Pre-training for Spoken Dialog Understanding with Explicit Cross-Modal Alignment (2023.acl-long)
Copied to clipboard
Tianshu Yu, Haoyu Gao, Ting-En Lin, Min Yang, Yuchuan Wu, Wentao Ma, Chao Wang, Fei Huang, Yongbin Li
| Challenge: | Existing speech-text pre-training methods are limited to one or two specific tasks, despite their success in speech-language processing tasks. |
| Approach: | They propose a temporal position prediction task to capture the speech-text alignment . they use a textual dialog pre-training task to generalize a response selection task . |
| Outcome: | The proposed model is superior in learning speech-text alignment and multi-turn dialog context. |
ADaPT: Token-Level Decoupling for Efficient Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Tingyun li, Zishang Jiang, Jinyi Han, Xinyi Wang, Sihang Jiang, Han Xia, Zhaoqian Dai, Ma Shuguang, Fei Yu, Jiaqing Liang, Yanghua Xiao
| Challenge: | Existing efficiency-oriented methods attempt to shorten or mix reasoning strategies, yet often degrade reasoning capability. |
| Approach: | They propose a token-level dual-process framework that explicitly decouples efficiency and correctness signals during training. |
| Outcome: | The proposed framework reduces inference cost while maintaining strong reasoning ability across multiple benchmarks. |
Toward Automated Robustness Evaluation of Mathematical Reasoning (2026.findings-acl)
Copied to clipboard
Yutao Hou, Zeguan Xiao, Fei Yu, Yihan Jiang, Ma Shuguang, Zhaoqian Dai, Hailiang Huang, Yun Chen, Guanhua Chen
| Challenge: | Existing robustness evaluations rely on hand-crafted templates or a limited set of perturbation rules, resulting in model failure. |
| Approach: | They propose a framework inspired by software stress testing that generates adversarial variants via a multi-round rewrite-verify loop, ensuring semantic consistency while successfully inducing model failure. |
| Outcome: | The proposed framework generates adversarial variants dynamically for each LLM, minimizing the risk of data contamination. |
PrAd: Prompt Adaptive Tuning for Decoder-only Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Prompt-based methods suffer from increased input lengths and sensitivity to weight initialization . adapter-based approaches can substantially increase inference time . |
| Approach: | a new paradigm for prompt-based tuning addresses the problem of fine tuning pretrained models . prompt--based methods suffer from increased input lengths and sensitivity to weight initialization . a prompt-oriented approach employs adapters for flexible input transformation . |
| Outcome: | a proposed framework can achieve comparable or better performance and higher inference efficiency even in multi-task scenarios. |
EPO: Explicit Policy Optimization for Strategic Reasoning in LLMs via Reinforcement Learning (2025.acl-long)
Copied to clipboard
Xiaoqian Liu, Ke Wang, Yongbin Li, Yuchuan Wu, Wentao Ma, Aobo Kong, Fei Huang, Jianbin Jiao, Junge Zhang
| Challenge: | Existing methods for strategic reasoning face challenges in adaptability, scalability, and transferring strategies to new contexts. |
| Approach: | They propose an explicit policy optimization model that provides strategies in open-ended action space and can be plugged into arbitrary LLM agents to motivate goal-directed behavior. |
| Outcome: | The proposed model provides strategies in open-ended action space and can be plugged into arbitrary LLM agents to motivate goal-directed behavior. |
VReST: Enhancing Reasoning in Large Vision-Language Models through Tree Search and Self-Reward Mechanism (2025.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) have shown exceptional performance in multimodal tasks, but their effectiveness in complex visual reasoning is constrained. |
| Approach: | They propose a training-free approach that enhances Reasoning in Large Vision-Language Models . they propose integrating Monte Carlo Tree Search and Self-Reward mechanisms into the reasoning tree . |
| Outcome: | The proposed approach surpasses current prompting methods and secures state-of-the-art performance across three multimodal reasoning benchmarks. |
UI-Copilot: Advancing Long-Horizon GUI Automation via Tool-Integrated Policy Optimization (2026.acl-long)
Copied to clipboard
Zhengxi Lu, Fei Tang, Guangyi Liu, Jin Ma, Kaitao Song, Xu Tan, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
| Challenge: | Experimental results show that UI-Copilot-7B achieves state-of-the-art performance on challenging MemGUI-Bench, outperforming strong 7B-scale GUI agents such as GUI-Owl-7B and UITARS-1.5-7B. |
| Approach: | They propose a collaborative framework where the GUI agent focuses on task execution while a lightweight copilot provides on-demand assistance for memory retrieval and numerical computation. |
| Outcome: | The proposed framework outperforms GUI-Owl-7B and UI-TARS-1.5-7B on MemGUI-Bench and delivers 17.1% improvement on AndroidWorld over the base Qwen model. |
LongWeave: A Long-Form Generation Benchmark Bridging Real-World Relevance and Verifiability (2025.findings-emnlp)
Copied to clipboard
Zikai Xiao, Fei Huang, Jianhong Tu, Jianhui Wei, Wen Ma, Yuxuan Zhou, Jian Wu, Bowen Yu, Zuozhu Liu, Junyang Lin
| Challenge: | Existing benchmarks for long-form generation assess real-world queries with hard-to-verify metrics or use synthetic setups that overlook real-life intricacies. |
| Approach: | They propose a new approach that balances verifiable and real-world assessment with Target-Anchored Evaluation. |
| Outcome: | The proposed model balances real-world and verifiable assessment with Target-Anchored Evaluation (TAE) it generates queries, textual materials, and anchors based on verifier targets within real-life scenarios . |
Improving Factual Consistency of News Summarization by Contrastive Preference Optimization (2024.findings-emnlp)
Copied to clipboard
Huawen Feng, Yan Fan, Xiong Liu, Ting-En Lin, Zekun Yao, Yuchuan Wu, Fei Huang, Yongbin Li, Qianli Ma
| Challenge: | Recent advances in news summarization have created problems with “hallucinations” that are factually inconsistent with the source text. |
| Approach: | They propose to disentangle LLMs’ propensities to generate faithful and fake content by adopting a probing-based specific training method to improve their capacity of distinguishing two types of propensity. |
| Outcome: | The proposed method disentangles LLMs’ propensities to generate faithful and fake content and improves their ability to distinguish between two types of propensity. |
CPO: Addressing Reward Ambiguity in Role-playing Dialogue via Comparative Policy Optimization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Comparative Policy Optimization (CPO) redefines the reward evaluation paradigm by shifting from sample-wise scoring to comparative group-wise score. |
| Approach: | They propose a method to optimize subjective tasks by shifting from sample-wise to comparative group-wise scoring. |
| Outcome: | The proposed framework shifts from sample-wise scoring to comparative group-wise score . it minimizes contextual bias and enables more robust and fair performance evaluation. |
SDPO: Segment-Level Direct Preference Optimization for Social Agents (2025.acl-long)
Copied to clipboard
Aobo Kong, Wentao Ma, Shiwan Zhao, Yongbin Li, Yuchuan Wu, Ke Wang, Xiaoqian Liu, Qicheng Li, Yong Qin, Fei Huang
| Challenge: | Direct Preference Optimization (DPO) has proven effective in aligning LLM behavior with human preferences across various tasks, but is limited in multi-turn social interactions. |
| Approach: | They propose a method which dynamically selects key segments within interactions to optimize multi-turn agent behavior. |
| Outcome: | The proposed methods outperform existing methods and proprietary LLMs on the SOTOPIA benchmark and show that they can improve social intelligence. |
TextFusion: Privacy-Preserving Pre-trained Model Inference via Token Fusion (2022.emnlp-main)
Copied to clipboard
Xin Zhou, Jinzhu Lu, Tao Gui, Ruotian Ma, Zichu Fei, Yuran Wang, Yong Ding, Yibo Cheung, Qi Zhang, Xuanjing Huang
| Challenge: | Existing methods to preserve inference privacy are available as cloud services . however, the risk of privacy leakage remains, according to recent studies . |
| Approach: | They propose a method to preserve inference privacy by fusing token representations in the cloud. |
| Outcome: | The proposed method preserves inference privacy without sacrificing performance on different scenarios. |
The Butterfly Effect of Model Editing: Few Edits Can Trigger Large Language Models Collapse (2024.findings-acl)
Copied to clipboard
| Challenge: | Even a single edit can trigger model collapse, manifesting as significant performance degradation in various benchmark tasks. |
| Approach: | They propose to use perplexity as a surrogate metric to determine whether an edited model's performance is affected by a single edit. |
| Outcome: | The proposed method shows that even a single edit can cause model collapse, manifesting as significant performance degradation in various benchmark tasks. |
A Law Reasoning Benchmark for LLM with Tree-Organized Structures including Factum Probandum, Evidence and Experiences (2025.findings-acl)
Copied to clipboard
Jiaxin Shen, Jinan Xu, Huiqi Hu, Luyi Lin, Guoyang Ma, Fei Zheng, Fandong Meng, Jie Zhou, Wenjuan Han
| Challenge: | a recent study focuses on generating impartial and interpretable judicial judgments based on established criminal fact. |
| Approach: | They propose a law reasoning schema enriched with hierarchical factum probandum, evidence, and implicit experience that enables public scrutiny and preventing bias. |
| Outcome: | The proposed schema enables public scrutiny and prevents bias in the "Intelligent Court" it employs a suite of legal analysis tools to address the challenge task. |
MuVER: Improving First-Stage Entity Retrieval with Multi-View Entity Representations (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in entity retrieval ignore the property that meanings of entity mentions diverge in different contexts and are related to various portions of descriptions. |
| Approach: | They propose a novel approach that constructs multi-view representations for entity descriptions and approximates the optimal view for mentions via a heuristic searching method. |
| Outcome: | The proposed approach achieves state-of-the-art performance on ZESHEL and improves quality of candidates on three standard Entity Linking datasets. |
Self-Explanation Prompting Improves Dialogue Understanding in Large Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have achieved great success in various NLP tasks, but the vast model parameters pose challenges in downstream fine-tuning. |
| Approach: | They propose a task-agnostic prompting strategy that analyzes each dialogue utterance before task execution to enhance LLMs' comprehension in multi-turn dialogues. |
| Outcome: | The proposed strategy outperforms other zero-shot prompts and matches or exceeds efficacy of few-shot ones. |
Instructional Fingerprinting of Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) require considerable cost to train from scratch . fingerprinting is essential to protect intellectual property and to ensure downstream users and developers adhere to their license terms. |
| Approach: | They propose a method for fingerprinting Large language models that implants a private key into the model to generate specific text when the key is present. |
| Outcome: | The proposed method prevents publisher overclaim and maintains robustness against fingerprint guessing and parameter-efficient training. |
FlowBench: Revisiting and Benchmarking Workflow-Guided Planning for LLM-based Agents (2024.findings-emnlp)
Copied to clipboard
| Challenge: | LLM-based agents are susceptible to undesired planning hallucinations when lacking specific knowledge for expertise-intensive tasks. |
| Approach: | They propose a benchmark to evaluate the efficacy of workflow-guided agent planning by formalizing different formats of workflow knowledge. |
| Outcome: | The proposed benchmark aims to improve the planning reliability of LLM-based agents by incorporating external workflow-related knowledge. |
CATS: A Pragmatic Chinese Answer-to-Sequence Dataset with Large Scale and High Quality (2023.acl-long)
Copied to clipboard
Liang Li, Ruiying Geng, Chengyang Fang, Bing Li, Can Ma, Rongyu Cao, Binhua Li, Fei Huang, Yongbin Li
| Challenge: | Current datasets bias in the English language while leaving other languages underexplored. |
| Approach: | They propose a Chinese answer-to-sequence dataset with high quality and large scale . they propose encoding space for two hybrid knowledge resources to convert this task to a graph-totext problem. |
| Outcome: | The proposed method is effective in generating textual descriptions for the Chinese answer-to-sequence dataset. |
The Fall of ROME: Understanding the Collapse of LLMs in Model Editing (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have found that model editing methods can cause large language models to collapse with just a single edit. |
| Approach: | They propose a method that uses prefixed keys and adds prefixes during testing to prevent model collapse. |
| Outcome: | The proposed method prevents model collapse while maintaining effectiveness, the authors show . Rank-One Model Editing (ROME) has been found to cause model collapse with just a single edit . |
TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing (2021.acl-demo)
Copied to clipboard
Xiao Wang, Qin Liu, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, Qinzhuo Wu, Zhengyan Li, Chong Zhang, Ruotian Ma, Zichu Fei, Ruijian Cai, Jun Zhao, Xingwu Hu, Zhiheng Yan, Yiding Tan, Yuan Hu, Qiyuan Bian, Zhihua Liu, Shan Qin, Bolin Zhu, Xiaoyu Xing, Jinlan Fu, Yue Zhang, Minlong Peng, Xiaoqing Zheng, Yaqian Zhou, Zhongyu Wei, Xipeng Qiu, Xuanjing Huang
| Challenge: | Existing approaches to textual robustness evaluation focus on slightly modifying the input data, which maintains the original meaning and results in a different prediction. |
| Approach: | They propose a multilingual robustness evaluation toolkit for NLP that integrates universal text transformations, task-specific transformations and adversarial attack. |
| Outcome: | The toolkit includes universal text transformation, task-specific transformation, adversarial attack, subpopulation, and their combinations to provide comprehensive robustness analyses. |
VisualRWKV: Exploring Recurrent Neural Networks for Visual Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Visual Language Models (VLMs) have been gaining popularity with large language models, but few attempts have been made to incorporate efficient linear Recurrent Neural Networks (RNNs) into VLMs. |
| Approach: | They propose a linear RNN model with a data-dependent recurrence and sandwich prompts to enhance modeling capabilities and a 2D image scanning mechanism to enrich the processing of visual sequences. |
| Outcome: | The proposed model achieves competitive performance compared to Transformer-based models on various benchmarks. |