Papers by Lei Bai

MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing (2026.acl-industry)

Challenge:	Recent studies have developed watermarking algorithms which restrict the generation process to leave an invisible trace for watermark detection.
Approach:	They propose a benchmarking procedure that compares different methods to ensure consistent watermarking strength and jointly evaluates their generation and detection performance.
Outcome:	The proposed benchmark compares 4 open-source watermarks on 2 LLMs under 2 watermarking strengths and observes the common struggles for current methods on maintaining the generation quality.

Copied to clipboard

Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Zirui Tang, Boyu Niu, Yuanhong Zheng, Dongsheng Ma, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Fangdong Wang, Jingzhou Chen, Xiaomeng Zhao, Liqun Wei, Wei Li, Shasha Wang, RuiLiang Xu, Yuanyuan Cao, Lu Chen, Qianqian Wu, Huaiyu Gu, Lindong Lu, Dechen Lin, null Shenguanlin, Xuanhe Zhou, Linfeng Zhang, Yuhang Zang, Xiaoyi Dong, Jiaqi Wang, Bo Zhang, Lei Bai, Pei Chu, Weijia Li, Jiang Wu, Lijun Wu, Zhenxiang Li, Guangyu Wang, Zhongying Tu, Chao Xu, Kai Chen, Bowen Zhou, Dahua Lin, Wentao Zhang, Conghui He

Green & Sustainable NLP

Challenge:	Document images are characterized by higher resolutions, denser content, and more complex structural layouts.
Approach:	They propose a 1.2B-parameter document parsing vision-language model that decouples layout analysis from local content recognition.
Outcome:	The proposed model surpasses general-purpose and domain-specific models on multiple benchmarks while maintaining significantly lower computational overhead.

LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks (2025.acl-long)

Copied to clipboard

Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li

WildDoc: How Far Are We from Achieving Comprehensive and Robust Document Understanding in the Wild? (2025.emnlp-main)

Challenge:	Existing benchmarks on longcontext large language models fail to reflect their deep understanding capabilities across diverse tasks.
Approach:	They propose a benchmark to assess the ability of long-context large language models to handle long-text problems.
Outcome:	The proposed model achieves 50.1% accuracy when directly answering the questions . human experts achieve only 53.7% accuracy under a 15-minute time constraint .

Copied to clipboard

An-Lan Wang, Jingqun Tang, Lei Liao, Hao Feng, Qi Liu, Xiang Fei, Jinghui Lu, Han Wang, Hao Liu, Yuliang Liu, Xiang Bai, Can Huang

Robustness in NLP

Challenge:	Existing benchmarks for document understanding in the wild are based on scanned or digital documents . however, these benchmarks fail to capture the challenges posed by documents in the real world .
Approach:	They propose a new benchmark that incorporates a diverse set of manually captured document images reflecting real-world conditions.
Outcome:	The proposed model is based on a set of manually captured document images reflecting real-world conditions and is compared with digital or scanned documents.

LLMRouterBench: A Massive Benchmark and Unified Framework for LLM Routing (2026.findings-acl)

Copied to clipboard

Hao Li, Yiqun Zhang, Zhaoyan Guo, Chenxu Wang, Shengji Tang, Qiaosheng Zhang, Yang Chen, Biqing Qi, Peng Ye, Lei Bai, Zhen Wang, Shuyue Hu

Challenge:	Large language model (LLM) routing assigns each query to the best suitable model from an ensemble.
Approach:	They introduce a large-scale benchmark and unified framework for LLM routing . they find that many routing methods exhibit similar performance under unified evaluation .
Outcome:	The proposed benchmark provides comprehensive metrics for both performance-oriented and performance-cost trade-off routing.

SURVEYFORGE : On the Outline Heuristics, Memory-Driven Generation, and Multi-dimensional Evaluation for Automated Survey Writing (2025.acl-long)

Copied to clipboard

Xiangchao Yan, Shiyang Feng, Jiakang Yuan, Renqiu Xia, Bin Wang, Lei Bai, Bo Zhang

Information Retrieval

Challenge:	SURVEYFORGE automates survey paper writing, but quality gap between LLM-generated and human-written surveys remains significant.
Approach:	They propose a survey tool that automatically generates and refines human-written surveys.
Outcome:	Experiments show that SURVEYFORGE outperforms previous work such as AutoSurvey in outline quality and content quality.

R3: End-to-End Reasoning-based Planning for Multi-step Retrosynthesis via Reinforcement Learning (2026.acl-long)

Copied to clipboard

YiFei Wang, Qizhi Pei, Jiangtao Feng, Yuntian Shi, Yi Duan, Lihao Wang, Lei Bai, Lijun Wu, Wei-Ying Ma, Hao Zhou

Nature-Inspired Population-Based Evolution of Large Language Models (2026.acl-long)

Challenge:	Experimental results show that R3 is a superior alternative to traditional search algorithms for multistep retrosynthesis planning.
Approach:	They propose a framework that reformulates multistep retrosynthetic planning as a generative reasoning task.
Outcome:	The proposed framework achieves state-of-the-art Top-1 accuracy of 43.7% on retrobench . it leverages Large Language Models to reformulate multistep retrosynthesis as a generative reasoning task.

Copied to clipboard

Yiqun Zhang, Peng Ye, Xiaocui Yang, Shi Feng, Shufei Zhang, Lei Bai, Wanli Ouyang, Shuyue Hu

GLTW: Joint Improved Graph Transformer and LLM via Three-Word Language for Knowledge Graph Completion (2025.findings-acl)

Challenge:	a new framework for population-based evolution of large language models is emerging . a population-driven evolution of LLMs is a key component of evolution, authors say .
Approach:	They propose a framework that allows for population-based evolution of large language models . they start with a population of parent LLMs and allow this population to evolve .
Outcome:	The proposed framework outperforms existing methods on 12 datasets.

Copied to clipboard

Kangyang Luo, Yuzhuo Bai, Cheng Gao, Shuzheng Si, Zhu Liu, Yingli Shen, Zhitong Wang, Cunliang Kong, Wenhao Li, Yufei Huang, Ye Tian, Xuantang Xiong, Lei Han, Maosong Sun

OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems (2024.acl-long)

Challenge:	Existing knowledge graphs lack the ability to integrate structural information into LLMs and output predictions deterministically.
Approach:	They propose a method which encodes structural information of KGs and merges it with LLMs to enhance KGC performance.
Outcome:	The proposed method improves the performance of KG Completion datasets on KGs by integrating structural information with LLMs.

Copied to clipboard

Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, Maosong Sun

MTR-DuplexBench: Towards a Comprehensive Evaluation of Multi-Round Conversations for Full-Duplex Speech Language Models (2026.findings-acl)

Challenge:	Large Language Models (LLMs) and Large Multimodal Models have exceeded general human capabilities in various tasks.
Approach:	They present an Olympiad-level bilingual multimodal scientific benchmark featuring 8,476 problems from Olympiad level mathematics and physics competitions.
Outcome:	The best performing model, GPT-4V, attains an average score of 17.97% on OlympiadBench, with a mere 10.74% in physics, highlighting the benchmark rigor and the intricacy of physical reasoning.

Copied to clipboard

Zhang He, Wenqian Cui, Haoning Xu, Xiao-Hui Li, Lei Zhu, Haoli Bai, Ma Shaohua, Irwin King

Dialogue Systems & Conversational Agents

Challenge:	Existing benchmarks focus on evaluating single-round interactions, neglecting other critical aspects.
Approach:	They propose a benchmark to evaluate full-duplex speech language models in multi-round settings . they segment continuous full-dual dialogues into discrete turns for evaluation .
Outcome:	The proposed benchmark compared full-duplex speech language models with full-dual speech models . the results show that the models perform better in multi-round settings than standard models compared to benchmarks .

FlowSearch: Advancing Deep Research with Dynamic Structured Knowledge Flow (2026.acl-long)

Copied to clipboard

Yusong Hu, Runmin Ma, Yue Fan, Jinxin Shi, Zongsheng Cao, Yuhao Zhou, Jiakang Yuan, Shuaiyu Zhang, Shiyang Feng, Xiangchao Yan, Shufei Zhang, Wenlong Zhang, Lei Bai, Bo Zhang

Multimodality

Challenge:	FlowSearch is a multi-agent framework that actively constructs and evolves a dynamic structured knowledge flow to drive subtask execution and reasoning.
Approach:	They propose a multi-agent framework that actively constructs and evolves a dynamic structured knowledge flow to drive subtask execution and reasoning.
Outcome:	The proposed framework achieves competitive performance on GAIA, HLE, GPQA and TRQA benchmarks and is available to download.

ReSo: A Reward-driven Self-organizing LLM-based Multi-Agent System for Reasoning Tasks (2025.emnlp-main)

Copied to clipboard

Heng Zhou, Hejia Geng, Xiangyuan Xue, Li Kang, Yiran Qin, Zhiyong Wang, Zhenfei Yin, Lei Bai

Cross-lingual Supervision Improves Unsupervised Neural Machine Translation (2021.naacl-industry)

Challenge:	Multi-agent systems (MAS) are limited by poor flexibility and scalability, with underdeveloped optimization strategies.
Approach:	They propose a task graph generation and a reward-driven two-stage agent selection process to integrate multi-agent systems to improve their reasoning capabilities.
Outcome:	The proposed model outperforms existing methods on Math-MAS and SciBench-MAS SciBech, while other methods completely fail.

Copied to clipboard

Mingxuan Wang, Hongxiao Bai, Hai Zhao, Lei Li

Multilinguality

Challenge:	Existing models that use only monolingual data have not been fully duplicated in the vast majority of language pairs, especially for zero-source languages.
Approach:	They propose to leverage the corpus from En-Fr and En-De to collectively train the translation from one language into many languages under one model.
Outcome:	The proposed model significantly improves translation quality with a big margin in the benchmark unsupervised translation tasks and achieves comparable performance to supervised NMT.

Pre-training Distillation for Large Language Models: A Design Space Exploration (2025.acl-long)

Copied to clipboard

Hao Peng, Xin Lv, Yushi Bai, Zijun Yao, Jiajie Zhang, Lei Hou, Juanzi Li

Easy Samples Are All You Need: Self-Evolving LLMs via Data-Efficient Reinforcement Learning (2026.findings-acl)

Challenge:	Knowledge distillation (KD) aims to transfer knowledge from a large teacher model to a smaller student model for model compression.
Approach:	They extend knowledge distillation to the pre-training phase of large language models . they first conduct an experiment using a teacher LLM to distill a 1.9B student LLM .
Outcome:	The proposed model can be used to distill a 1.9B student model using a teacher LLM.

Copied to clipboard

Zhiyin Yu, Bo Zhang, Qibin Hou, Zhonghai Wu, Xiao Luo, Lei Bai

Responsible & Trustworthy NLP

Challenge:	Experimental results show that EasyRL consistently outperforms state-of-the-art baselines due to the substantial annotation cost and issues such as model collapse or reward hacking.
Approach:	They propose a supervised RL approach with a divide-and-conquer strategy that simulates the human cognitive acquisition curve using easy labeled data.
Outcome:	The proposed approach outperforms state-of-the-art models on mathematical and scientific benchmarks using only 10% of easy labeled data.

SQUIRE: A Sequence-to-sequence Framework for Multi-hop Knowledge Graph Reasoning (2022.emnlp-main)

Copied to clipboard

Yushi Bai, Xin Lv, Juanzi Li, Lei Hou, Yincen Qu, Zelin Dai, Feiyu Xiong

MM-MATH: Advancing Multimodal Math Evaluation with Process Evaluation and Fine-grained Classification (2024.findings-emnlp)

Challenge:	Existing methods for multi-hop knowledge graph reasoning suffer from slow and poor convergence . a transformer model can be used to learn and predict in an end-to-end fashion, giving faster convergence compared to previous methods .
Approach:	They propose a Sequence-to-sequence based multi-hop reasoning framework . it uses an encoder-decoder transformer structure to translate the query to a path .
Outcome:	The proposed framework can learn and predict in an end-to-end fashion, which gives better and faster convergence.

Copied to clipboard

Kai Sun, Yushi Bai, Ji Qi, Lei Hou, Juanzi Li

Information Retrieval

Challenge:	Existing benchmarks for multimodal reasoning in large multimodal models are underperforming on multimodal tasks.
Approach:	They propose a benchmark for multimodal reasoning in large multimodal models, MM-MATH . MM's process evaluation employs LMM-as-a-judge to automatically analyze solution steps . diagram misinterpretation is the most common error, they find .
Outcome:	The proposed model achieves only 31% accuracy, compared to 82% for humans.

A Survey of Reinforcement Learning for Large Language Models under Data Scarcity: Challenges and Solutions (2026.acl-long)

Copied to clipboard

Zhiyin Yu, Yuchen Mou, Juncheng Yan, Junyu Luo, Chunchun Chen, Xing Wei, Yunhui Liu, Hongru Sun, Yuxing Zhang, Jun Xu, Yatao Bian, Ming Zhang, Wei Ye, Tieke He, Jie Yang, Guanjie Zheng, Zhonghai Wu, Bo Zhang, Lei Bai, Xiao Luo

Responsible & Trustworthy NLP

Challenge:	Existing research on reinforcement learning for LLMs under data scarcity has not been unified.
Approach:	They propose a top-up hierarchical framework built around three complementary perspectives: data-centric, training-centric and framework-centric.
Outcome:	The proposed framework provides a clear conceptual foundation for understanding the design space of data-efficient RL for large language models and to guide researchers working in this emerging area.

Aspect Is Not You Need: No-aspect Differential Sentiment Framework for Aspect-based Sentiment Analysis (2022.naacl-main)

Copied to clipboard

Jiahao Cao, Rui Liu, Huailiang Peng, Lei Jiang, Xu Bai

Sentiment Analysis

Challenge:	Existing approaches to classify aspects with aspect sentiment bias are hard to find .
Approach:	They propose a no-aspect differential sentiment framework for the ABSA task that eliminates aspect sentiment bias and uses differential sentiment loss instead of cross-entropy loss to better classify the sentiments.
Outcome:	The proposed framework can be combined with almost all traditional ABSA methods.

LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (2024.acl-long)

Copied to clipboard

Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li

Multilinguality

Challenge:	Large language models (LLMs) can only handle texts a few thousand tokens long, limiting their applications on longer sequence inputs, such as books, reports, and codebases.
Approach:	They propose a bilingual, multi-task benchmark for long context understanding that extends context windows and more sophisticated memory mechanisms to improve models' long context capabilities.
Outcome:	The proposed model outperforms open-source models but struggles on longer contexts.

LongAlign: A Recipe for Long Context Alignment of Large Language Models (2024.findings-emnlp)

Copied to clipboard

Yushi Bai, Xin Lv, Jiajie Zhang, Yuze He, Ji Qi, Lei Hou, Jie Tang, Yuxiao Dong, Juanzi Li

A Scalable Multi-LLM Collaboration System with Retrieval-based Selection and Exploration-Exploitation-Driven Enhancement (2026.acl-long)

Challenge:	Existing studies to build long context language models focus on context extension and continual training on long text.
Approach:	They propose a recipe for instruction fine-tuning on input sequences of similar length . they adopt packing and sorted batching strategies to speed up supervised fine-uning .
Outcome:	The proposed model outperforms existing recipes for LLMs in long context tasks by 30% while maintaining proficiency in handling short, generic tasks.

Copied to clipboard

Shengji Tang, Jianjian Cao, Weihao Lin, Jiale Hong, Bo Zhang, Shuyue Hu, Lei Bai, Tao Chen, Wanli Ouyang, Peng Ye

Ruler: A Model-Agnostic Method to Control Generated Length for Large Language Models (2024.findings-emnlp)

Challenge:	Existing multi-LLM collaboration systems often encounter scalability challenges when integrating new LLMs and tasks.
Approach:	They propose a Scalable Multi-LLM Collaboration System to coordinate multiple open-source LLMs.
Outcome:	The proposed system outperforms prevailing closed-source LLMs on eight mainstream benchmarks on multiple tasks.

Copied to clipboard

Jiaming Li, Lei Zhang, Yunshui Li, Ziqiang Liu, Yuelin Bai, Run Luo, Longze Chen, Min Yang

Extending Context Window of Large Language Models via Semantic Compression (2024.findings-acl)

Challenge:	Large language models struggle to meet user’s needs when required to generate responses of a specific length due to their inherent difficulty in accurately perceiving numerical constraints.
Approach:	They propose a Target Length Generation Task and propose RULER, a model-agnostic approach that controls generated length for large language models.
Outcome:	The proposed model-agnostic approach improves instruction-following ability of large language models under length-constrained instructions and can generate appropriate MLT when length constraints are not explicitly provided.

Copied to clipboard

Weizhi Fei, Xueyan Niu, Pingyi Zhou, Lu Hou, Bo Bai, Lei Deng, Wei Han

Edge-free but Structure-aware: Prototype-Guided Knowledge Distillation from GNNs to MLPs (2025.coling-main)

Challenge:	Existing models rely on a quadratic computation to generate long texts . current models impose limitations on the length of text inputs .
Approach:	They propose a semantic compression method that extends the context window of large language models . the method reduces the semantic redundancy of long inputs before passing them to the LLMs .
Outcome:	The proposed method extends the context window of large language models across tasks . it exhibits consistent fluency in text generation while reducing associated computational overhead.

Copied to clipboard

Taiqiang Wu, Zhe Zhao, Jiahao Wang, Xingyu Bai, Lei Wang, Ngai Wong, Yujiu Yang

Green & Sustainable NLP

Challenge:	Existing methods to train low-latency multilayer perceptrons (MLPs) on graph tasks are based on graph nodes and lack graph structural information.
Approach:	They propose to distill graph structural information from Graph Neural Networks (GNNs) to low-latency multilayer perceptrons (MLPs) on graph tasks.
Outcome:	The proposed method does not require graph edges (edge-free setting) yet learns structure-aware MLPs.

Faster and Better LLMs via Latency-Aware Test-Time Scaling (2025.findings-emnlp)

Copied to clipboard

Zili Wang, Tianyu Zhang, Haoli Bai, Lu Hou, Xianzhi Yu, Wulong Liu, Shiming Xiang, Lei Zhu

Challenge:	Existing research has overlooked the efficiency of TTS from a latency-sensitive perspective.
Approach:	They propose two approaches to achieve latency-optimal TTS by branch-wise parallelism and sequence-wise parallelism.
Outcome:	The proposed approach achieves latency-optimal TTS for large models . branch-wise parallelism and sequence-wise parallelism are key approaches .

From Completion to Editing: Unlocking Context-Aware Code Infilling via Search-and-Replace Instruction Tuning (2026.acl-long)

Copied to clipboard

Jiajun Zhang, Zeyu Cui, Jiaxi Yang, Lei Zhang, Yuheng Jing, Zeyao Ma, Tianyi Bai, Zilei Wang, Qiang Liu, Liang Wang, Binyuan Hui, Junyang Lin

Programming Languages in NLP

Challenge:	Fill-in-the-Middle (FIM) models suffer from performance degradation and prohibitive latency.
Approach:	They propose a search-and-replace infilling framework that integrates agentic verification and editing into a single-pass inference process.
Outcome:	The proposed framework harmonizes completion tasks with the instruction-following priors of Chat LLMs, extending the paradigm from static infilling to dynamic context-aware editing.

SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis (2025.findings-emnlp)

Copied to clipboard

Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Junjie Zhang, Fei Bai, Jia Deng, Xin Zhao, Zheng Liu, Lei Fang, Zhongyuan Wang, Ji-Rong Wen

MTRouter: Cost-Aware Multi-Turn LLM Routing with History–Model Joint Embeddings (2026.acl-long)

Challenge:	Existing approaches to deep search training lack high-quality training trajectories, prohibitive computational costs and lack of high-fidelity training data.
Approach:	They propose a framework that synthesizes high-quality training data by simulating real user interactions in live web search environments.
Outcome:	The proposed framework synthesizes high-quality training data by simulating user interactions in live web search environments.

Copied to clipboard

Yiqun Zhang, Hao Li, Zihan Wang, Shi Feng, Xiaocui Yang, Daling Wang, Bo Zhang, Lei Bai, Shuyue Hu

Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning (2026.acl-long)

Challenge:	Multi-turn, long-horizon tasks require dozens of sequential model calls per episode.
Approach:	They propose a cost-aware multi-turn LLM routing tool which encodes interaction history and candidate models into joint history–model embeddings and learns an outcome estimator from logged trajectories to predict turn-level model utility.
Outcome:	The proposed model reduces cost and performance by 58.7% on ScienceWorld and on Humanity’s Last Exam (HLE) and even reduces costs for held-out tasks.

Copied to clipboard

Zelin Tan, Hejia Geng, Xiaohang Yu, Mulei Zhang, Guancheng Wan, Yifan Zhou, Qiang He, Xiangyuan Xue, Heng Zhou, Yutao Fan, Zhong-Zhi Li, Zaibin Zhang, Guibin Zhang, Chen Zhang, Zhenfei Yin, Philip Torr, Lei Bai

Self-attention-based Graph-of-Thought for Math Problem Solving (2025.findings-acl)

Challenge:	elucidating scaling laws for large language models (LLMs) during pre-training remains unexplored.
Approach:	They characterize how model scale, data, and compute interact during pre-training . they find that large models consistently demonstrate superior compute and data efficiency .
Outcome:	The proposed scaling laws offer practical guidance for scaling reasoning capabilities through reinforcement learning post-training.

Copied to clipboard

Ruiqiao Bai, Xue Han, Shuo Lei, Junlan Feng, Yanyan Luo, Chao Deng