Papers by Feiyu Wang
CodeFlowBench: A Multi-turn, Iterative Benchmark for Complex Code Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Modern software development demands code that is maintainable, testable, and scalable by organizing the implementation into modular components with iterative reuse of existing codes. |
| Approach: | They propose a benchmark to evaluate LLMs' ability to perform codeflow by reusing existing functions over multiple turns. |
| Outcome: | The proposed benchmarks show that LLMs perform significantly worse in multi-turn codeflow scenarios and that their performance inversely correlates with dependency complexity. |
UHGEval: Benchmarking the Hallucination of Chinese Large Language Models via Unconstrained Generation (2024.acl-long)
Copied to clipboard
Xun Liang, Shichao Song, Simin Niu, Zhiyu Li, Feiyu Xiong, Bo Tang, Yezhaohui Wang, Dawei He, Cheng Peng, Zhonghao Wang, Haiying Deng
| Challenge: | Large language models (LLMs) produce hallucinated text, compromising their practical utility in professional contexts. |
| Approach: | They have developed an unconstrained hallucination generation evaluation benchmark that contains hallucines generated by large language models with minimal restrictions. |
| Outcome: | The proposed benchmarks are based on a Chinese-language dataset that is lacking in the field. |
PositionID: LLMs can Control Lengths, Copy and Paste with Explicit Positional Awareness (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have impressive capabilities across various domains, including role-playing, creative writing, mathematical reasoning, and coding. |
| Approach: | They propose two methods to improve the model’s adherence to length constraints and copy-paste accuracy without compromising response quality. |
| Outcome: | The proposed methods improve the model’s adherence to length constraints and copy-paste accuracy without compromising response quality. |
FRAME: Boosting LLMs with A Four-Quadrant Multi-Stage Pretraining Strategy (2025.findings-acl)
Copied to clipboard
Xuemiao Zhang, Feiyu Duan, Xu Liangyu, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, Xunliang Cai
| Challenge: | Multi-stage pretraining methods lack quantitative criteria for data partitioning and instead rely on intuitive heuristics. |
| Approach: | They propose a Four-quadRAnt Multi-stage prEtraining strategy that partitions data into four quadrants to achieve significant loss reductions four times. |
| Outcome: | The proposed strategy achieves 16.8% improvement over random across MMLU and CMMLU for the 3B model. |
Controlled Text Generation for Large Language Model with Dynamic Attribute Graphs (2024.findings-acl)
Copied to clipboard
| Challenge: | Controlled Text Generation (CTG) aims to produce texts that exhibit specific desired attributes. |
| Approach: | They propose a pluggable CTG framework for Large Language Models to control text . they use attribute scorers to evaluate attributes of sentences and construct dynamic attribute graphs . |
| Outcome: | The proposed framework achieves a peak improvement of 19.29% over baseline methods in two tasks. |
DeMAC: Enhancing Multi-Agent Coordination with Dynamic DAG and Manager-Player Feedback (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Multi-agent systems (MAS) powered by large language models struggle to adapt to evolving task dependencies and to handle uncertainties. |
| Approach: | They propose a Dynamic Environment-Aware Manager-Player Agents Coordination framework that enhances multi-agent coordination through long-term strategic planning. |
| Outcome: | The proposed framework outperforms traditional reinforcement learning and human-agent collaboration in the Overcooked simulation. |
LLMs Know What They Need: Leveraging a Missing Information Guided Framework to Empower Retrieval-Augmented Generation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing solutions to improve the accuracy of RAG are based on retrieval-augmented generation . however, RAG still faces several challenges in tackling complex multi-hop queries . |
| Approach: | They propose a Missing Information Guided Retrieve-Extraction-Solving paradigm that leverages the identification of missing information to generate a targeted query. |
| Outcome: | The proposed method can extract information from retrieved knowledge and know what is still missing. |
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%. |
FIRE: Flexible Integration of Data Quality Ratings for Effective Pretraining (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to evaluate data quality rely on heuristic techniques or single quality signals. |
| Approach: | They propose a framework for integrating multiple data quality raters that integrates multiple quality signals into a unified space and provides a comprehensive quality signal for each data point. |
| Outcome: | The proposed framework outperforms existing methods and boosts model performance across a wide range of downstream tasks while requiring less than 37.5% tokens to reach the target performance. |
Attribution and Application of Multiple Neurons in Multimodal Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to identify multimodal neurons in MLLMs are insufficiently understood . previous studies focused on identifying neurons corresponding to single-tokens . |
| Approach: | They propose a method to identify multimodal neurons in Transformer-based MLLMs . they introduce fuzzy set theory to model the complex relationship between neurons and semantic concepts . |
| Outcome: | The proposed method improves performance on the Visual Question Answering task. |
MoC: Mixtures of Text Chunking Learners for Retrieval-Augmented Generation System (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text chunking are limited by text chunks and lack of domain-specific knowledge. |
| Approach: | They propose a dual-metric evaluation method to quantify text chunking quality . they aim to generate a structured list of chunking regular expressions . |
| Outcome: | The proposed method enables direct quantification of chunking quality . it substantiates the need to integrate LLMs into chunking process . |
Text2Mem: A Unified Memory Operation Language for Memory Operating System (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing memory frameworks lack a formal, executable specification for memory control. |
| Approach: | They propose a unified memory operation language that standardizes translation of natural-language instructions into reliable execution. |
| Outcome: | The proposed language standardizes translation of natural-language instructions into reliable execution. |
H-MAS: Hierarchical Multi-Agent Scheduling for Multi-Tenant LLM Serving (2026.findings-acl)
Copied to clipboard
Yuhan Liu, Cong Xu, Qi Jia, Yihua Wang, Feiyu Chen, Liang Jin, Lu Liu, Yaqian Zhao, Yuting Ding, Xiang Li
| Challenge: | Multi-tenant Model-as-a-Service (MaaS) workloads exhibit non-stationarity across multiple time scales . existing request schedulers often rely on a fixed policy that remains unchanged at runtime . |
| Approach: | They propose a hierarchical multi-agent scheduler that operates in a layered closed loop . they propose to maintain 1.2–3.0 higher Goodput than SGLang and vLLM . |
| Outcome: | Experiments show that H-MAS achieves 1.2–3.0 higher Goodput than SGLang and vLLM . it maintains more stable QoS under diverse request lengths and heterogeneous SLO targets . |
Preference Curriculum: LLMs Should Always Be Pretrained on Their Preferred Data (2025.findings-acl)
Copied to clipboard
Xuemiao Zhang, Xu Liangyu, Feiyu Duan, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, Xunliang Cai
| Challenge: | Existing methods of uniformly sampling data throughout the pretraining process are suboptimal because they overlook the model's evolving data preferences. |
| Approach: | They propose a Perplexity Difference (PD) based Preference Curriculum learning framework which perceives and uses the data preferred by LLMs as their capabilities improve . they propose PDPC to complete the arrangement of the dataset offline and ensure continuous training without interruption. |
| Outcome: | The proposed framework surpasses baselines on 1.3B and 3B models and achieves an increased average accuracy of over 8.1% across MMLU and CMMLU. |
QAEncoder: Towards Aligned Representation Learning in Question Answering Systems (2025.acl-long)
Copied to clipboard
Zhengren Wang, Qinhan Yu, Shida Wei, Zhiyu Li, Feiyu Xiong, Xiaoxing Wang, Simin Niu, Hao Liang, Wentao Zhang
| Challenge: | Modern QA systems entail retrieval-augmented generation (RAG) for accurate and trustworthy responses, but the inherent gap between user queries and relevant documents hinders precise matching. |
| Approach: | They propose a retrieval-augmented generation (RAG)-based approach to bridge this gap by attaching document fingerprints to the embedding to estimate the expectation of potential queries. |
| Outcome: | Experiments across diverse datasets, languages, and embedding models confirm the proposed solution is simple-yet-effective with zero additional index storage, retrieval latency, training costs, or catastrophic forgetting and hallucination issues. |
NewsBench: A Systematic Evaluation Framework for Assessing Editorial Capabilities of Large Language Models in Chinese Journalism (2024.acl-long)
Copied to clipboard
Miao Li, Ming-Bin Chen, Bo Tang, ShengbinHou ShengbinHou, Pengyu Wang, Haiying Deng, Zhiyu Li, Feiyu Xiong, Keming Mao, Cheng Peng, Yi Luo
| Challenge: | a novel evaluation framework assesses the capabilities of Large Language Models (LLMs) for editorial capabilities in Chinese journalism. |
| Approach: | They propose to use a benchmark dataset to assess the capabilities of Large Language Models (LLMs) for editorial capabilities in Chinese journalism. |
| Outcome: | The proposed evaluation framework is based on a dataset of 1,267 test samples in 24 news domains. |
HopRAG: Multi-Hop Reasoning for Logic-Aware Retrieval-Augmented Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Traditional retrieval systems focus on lexical or semantic similarity rather than logical relevance. |
| Approach: | They propose a new RAG framework that augments retrieval with logical reasoning . hopRAG uses a retrieve-reason-prune mechanism to explore multi-hop neighbors . |
| Outcome: | The proposed framework outperforms conventional retrieval systems and state-of-the-art benchmarks on multi-hop QA tasks. |
SafeRAG: Benchmarking Security in Retrieval-Augmented Generation of Large Language Model (2025.acl-long)
Copied to clipboard
Xun Liang, Simin Niu, Zhiyu Li, Sensen Zhang, Hanyu Wang, Feiyu Xiong, Zhaoxin Fan, Bo Tang, Jihao Zhao, Jiawei Yang, Shichao Song, Mengwei Wang
| Challenge: | Existing approaches to integrating external knowledge into large language models (LLMs) however, the incorporation of external knowledge increases the vulnerability of LLMs . |
| Approach: | They propose a benchmark to evaluate the RAG security using a dataset . they classify attack tasks into silver noise, inter-context conflict, soft ad, and white Denial-of-Service . |
| Outcome: | The proposed benchmark evaluates the security of RAG against 14 representative RAG components. |