Papers by Cao Xiao
Length-Induced Embedding Collapse in PLM-based Models (2025.acl-long)
Copied to clipboard
| Challenge: | In text embeddings from PLMs are essential for many NLP applications, but performance degrades on longer texts. |
| Approach: | They propose a method which mitigates the phenomenon of Length Collapse . they propose TempScale to ensure more consistent embeddings across different text lengths . |
| Outcome: | The proposed method improves performance on MTEB and LongEmbed by 0.94% on short and 1.10% on long texts. |
Unity in Diversity: Collaborative Pre-training Across Multimodal Medical Sources (2024.acl-long)
Copied to clipboard
Xiaochen Wang, Junyu Luo, Jiaqi Wang, Yuan Zhong, Xiaokun Zhang, Yaqing Wang, Parminder Bhatia, Cao Xiao, Fenglong Ma
| Challenge: | Current pre-training techniques rely on a limited scope of medical data, limiting the range of downstream tasks. |
| Approach: | They propose a pre-training strategy that unifies patient data within individual sources and captures explicit and implicit correlations between patients across different sources. |
| Outcome: | The proposed strategy bridges the gap between multimodal medical sources by aggregating patient data within individual sources and capturing explicit and implicit correlations between patients across sources. |
RetroMAE-2: Duplex Masked Auto-Encoder For Pre-Training Retrieval-Oriented Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods for retrieval-oriented language models focus on contextualized embedding of the [CLS] token, but recent study shows that ordinary tokens besides [CLL] may provide extra information, which help to produce a better representation effect. |
| Approach: | They propose a method where all contextualized embeddings of pre-trained model can be jointly pre-trained for retrieval tasks. |
| Outcome: | The proposed method improves the quality of representation where all contextualized embeddings of the pre-trained model can be leveraged. |
Program Transfer for Answering Complex Questions over Knowledge Bases (2022.acl-long)
Copied to clipboard
| Challenge: | Program induction for complex questions over knowledge bases relies on a large number of parallel question-program pairs for the given KB, but the gold program annotations are usually lacking, making learning difficult. |
| Approach: | They propose an approach to leverage program annotations on rich KBs as external supervision signals to aid program induction for low-resourced KB. |
| Outcome: | The proposed approach outperforms SOTA methods on ComplexWebQuestions and WebQuestionSP. |
Incomplete In-context Learning (2026.acl-long)
Copied to clipboard
Wenqiang Wang, Wen Yujia, Yan Xiao, Zhifeng Chen, Yangshijie Zhang, Peng Chen, Mingbo Yang, Xiaochun Cao
| Challenge: | Existing in-context learning assumes the retrieval dataset contains demonstrations for all output label spaces. |
| Approach: | They propose a framework with train-free and train-based variants to address IICL . they propose to integrate a dataset with labeled demonstrations for each output space . |
| Outcome: | The proposed framework outperforms existing methods under incomplete retrieval datasets and even outperformed ICL with complete labels. |
Unlocking Memorization in Large Language Models with Dynamic Soft Prompting (2024.emnlp-main)
Copied to clipboard
Zhepeng Wang, Runxue Bao, Yawen Wu, Jackson Taylor, Cao Xiao, Feng Zheng, Weiwen Jiang, Shangqian Gao, Yanfu Zhang
| Challenge: | Pretrained large language models excel in a variety of natural language processing tasks . however, they pose significant security risks due to their tendency to memorize training data . |
| Approach: | They propose a method to estimate LLM memorization using dynamic, prefix-dependent soft prompts. |
| Outcome: | The proposed method can achieve maximum relative improvement of 135.3% and 39.8% over baseline compared to state-of-the-art methods. |
RankPrompt: Step-by-Step Comparisons Make Language Models Better Reasoners (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing solutions to reasoning tasks require extensive human annotations or fail in scenarios with inconsistent responses. |
| Approach: | They propose a new method that enables LLMs to self-rank their responses without additional resources. |
| Outcome: | The proposed method improves reasoning performance of ChatGPT and GPT-4 with 13% improvement over existing methods. |
Dynamic Uncertainty Ranking: Enhancing Retrieval-Augmented In-Context Learning for Long-Tail Knowledge in LLMs (2025.naacl-long)
Copied to clipboard
| Challenge: | Prior work has shown that in-context learning (ICL) with retriever augmentation can help LLMs better capture long-tail knowledge, reducing their reliance on pre-trained data. |
| Approach: | They propose a reinforcement learning-based dynamic uncertainty ranking method that accounts for the varying impact of each retrieved sample on LLM predictions. |
| Outcome: | The proposed method outperforms baseline models on question-answering datasets by 2.76% and 5.96% on long-tail questions that elude zero-shot inference. |
CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging (2026.findings-acl)
Copied to clipboard
Jie Cao, Zhenxuan Fan, Zhuonan Wang, Tianwei Lin, Ziyuan Zhao, Rolan Yan, Wenqiao Zhang, Feifei Shao, Hongwei Wang, Jun Xiao, Siliang Tang
| Challenge: | Existing PEFT methods suffer from limited parameter efficiency and coarse-grained adaptation due to proliferation of LoRA experts and instance-level routing. |
| Approach: | They propose a new MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. |
| Outcome: | The proposed framework outperforms existing methods on multiple tasks while maintaining parameter efficiency. |
An Augmented Benchmark Dataset for Geometric Question Answering through Dual Parallel Text Encoding (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods for solving geometric problems are limited due to lack of high-quality datasets and efficient neural solvers. |
| Approach: | They propose to annotate 2,518 geometric problems with richer types and greater difficulty using a benchmark dataset. |
| Outcome: | The proposed method improves the accuracy of automatic geometric problem solving to 66.09%. |
Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies focus on inconsistency issues within a single LLM, while we explore the inter-consistencies among multiple LLMs for collaboration. |
| Approach: | They propose a formal debate framework to examine whether LLMs can collaborate effectively to achieve a consensus for a shared goal. |
| Outcome: | The proposed framework enables LLMs to achieve consensus in three real-world debate scenarios with real-time scenarios aligned to the LLM's goals. |
SCALER: Synthetic Scalable Adaptive Learning Environment for Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning (RL) is a principled way to enhance the reasoning capabilities of large language models, yet its effectiveness hinges on training signals that remain informative as models evolve. |
| Approach: | They propose a framework that sustains effective learning signals through adaptive environment design that transforms real-world programming problems into verifiable reasoning environments with controllable difficulty and unbounded instance generation. |
| Outcome: | The proposed framework outperforms baselines across diverse reasoning benchmarks and exhibits more stable, long-horizon training dynamics. |
TriSum: Learning Summarization Ability from Large Language Models with Structured Rationale (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have advanced tasks like text summarization, but their size and computational demands limit their use in resource-constrained and privacy-centric settings. |
| Approach: | They propose a framework for distilling LLMs’ text summarization abilities into a compact, local model using a curriculum learning strategy that evolves from simple to complex tasks. |
| Outcome: | The proposed framework outperforms baseline models on CNN/DailyMail, XSum, and ClinicalTrial, and improves interpretability by providing insights into the summarization rationale. |
Revealing the Parallel Multilingual Learning within Large Language Models (2024.emnlp-main)
Copied to clipboard
Yongyu Mu, Peinan Feng, Zhiquan Cao, Yuzhang Wu, Bei Li, Chenglong Wang, Tong Xiao, Kai Song, Tongran Liu, Chunliang Zhang, JingBo Zhu
| Challenge: | Large language models (LLMs) can handle multilingual and cross-lingual text within a single input; however, previous studies focusing on using English as the pivot language to enhance language understanding and reasoning focus on using multiple languages. |
| Approach: | They propose to use parallel multilingual input to enhance the model's comprehension of the input and to examine how multilingual processing affects prediction. |
| Outcome: | The proposed model can handle multilingual and cross-lingual text within a single input, but previous studies focused on using English as the pivot language to enhance language understanding and reasoning. |
Privacy Risks of Intermediate Representations: Attribute Inference in Distributed LLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Distributed LLMs avoid raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy. |
| Approach: | They propose a distributed inference framework that transmits intermediate hidden states to avoid sending raw inputs by exposing sensitive user attributes. |
| Outcome: | The proposed approach achieves Top-1 accuracy of 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR. |
AutoTrial: Prompting Language Models for Clinical Trial Design (2023.emnlp-main)
Copied to clipboard
| Challenge: | Generative large language models (LLMs) are a popular tool for creating coherent and human-like documents for clinical trials. |
| Approach: | They propose to generate clinical eligibility criteria using language models by a hybrid of discrete and neural prompting and scalable knowledge incorporation via in-context learning. |
| Outcome: | The proposed method generates high-quality criteria texts fluent and coherent with high accuracy against the GPT-3.5 baselines. |
ChatSOP: An SOP-Guided MCTS Planning Framework for Controllable LLM Dialogue Agents (2025.acl-long)
Copied to clipboard
Zhigen Li, Jianxiang Peng, Yanmeng Wang, Yong Cao, Tianhao Shen, Minghui Zhang, Linxi Su, Shang Wu, Yihang Wu, YuQian Wang, Ye Wang, Wei Hu, Jianfeng Li, Shaojun Wang, Jing Xiao, Deyi Xiong
| Challenge: | Existing models that use Large Language Models (LLMs) show superior performance in various tasks, but lack of controllability leads to unfocused conversations or task failure. |
| Approach: | They propose a standard operating procedure (SOP) framework to regulate dialogue flow by integrating Chain of Thought reasoning and supervised fine-tuning for SOP prediction. |
| Outcome: | The proposed method achieves a 27.95% improvement in action accuracy compared to baseline models based on GPT-3.5 and also shows notable gains for open-source models. |
FRSUM: Towards Faithful Abstractive Summarization via Enhancing Factual Robustness (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models of abstractive summarization are able to generate fluent and coherent summaries, but they still suffer from the unfaithful generation problem. |
| Approach: | They propose to improve the faithfulness of existing models by enhancing their factual robustness by using a novel training strategy, namely FRSUM, which teaches the model to defend against both explicit adversarial samples and implicit factual adversarials. |
| Outcome: | The proposed training strategy improves faithfulness of various models, such as T5, BART, and T5 . |
Attribution-Based Analysis and Optimization of Modular Agentic Workflows (2026.findings-acl)
Copied to clipboard
Yingxuan Yang, Bo Huang, Siyuan Qi, Chao Feng, Haoyi Hu, Yuxuan Zhu, Jinbo Hu, Haoran Zhao, Ziyi He, Xiao Liu, ZongYu Wang, Muning Wen, Lin Qiu, Xuezhi Cao, Xunliang Cai, Yong Yu, Weinan Zhang
| Challenge: | Large Language Models (LLMs) have driven the rise of agentic workflows . yet, how can we attribute performance gains to individual upgrades and their interactions? |
| Approach: | They propose a game-theoretic framework that models component upgrades as players and evaluates component coalitions to compute Shapley values. |
| Outcome: | The proposed framework provides interaction-aware attribution and recommendation for model allocation under a fixed workflow structure. |
Focus on What Matters: Enhancing Medical Vision-Language Models with Automatic Attention Alignment Tuning (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods rely on inference-time interventions, which are limited in attention adaptation or require additional supervision. |
| Approach: | They propose a framework for automatic attention alignment tuning that leverages weak labels from SAM and selectively modifies visually-critical attention heads to improve alignment while minimizing interference. |
| Outcome: | The proposed framework outperforms state-of-the-art models on medical VQA and report generation benchmarks. |
Multi-task Adversarial Attacks against Black-box Model with Few-shot Queries (2025.acl-long)
Copied to clipboard
| Challenge: | Existing adversarial text attacks rely on abundant access to shared internal features and numerous queries, limited to a single task type. |
| Approach: | They propose a black-box attack that exploits the transferability of adversarial texts . they use a deep-level substitute model trained in a plug-and-play manner for text classification . |
| Outcome: | The proposed attack can target multiple tasks with minimal perturbations . it can target commercial APIs, large language models, and image-generation models . |
PILOT: Legal Case Outcome Prediction with Case Law (2024.naacl-long)
Copied to clipboard
| Challenge: | predicting legal case outcomes requires identifying relevant precedent cases . predicting case outcomes in case law systems presents unique challenges . |
| Approach: | They propose a framework for making legal case outcome predictions with case law . they propose to use two modules for relevant case retrieval and temporal pattern handling . |
| Outcome: | The proposed framework shows significant improvement over previous models based on civil law cases . it is crucial to identify relevant precedent cases that serve as evidence for judges . |
AdaReTaKe: Adaptive Redundancy Reduction to Perceive Longer for Video-language Understanding (2025.findings-acl)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) are limited by context length when processing long videos. |
| Approach: | They propose a training-free method that flexibly reduces redundancy by allocating compression ratios among time and model layers with theoretical guarantees. |
| Outcome: | Experiments on videoMME, MLVU, LongVideoBench, and LVBench show that AdaRETAKE outperforms existing methods by 2.3% and 2.8% for 7B and 72B models. |
RetroMAE: Pre-Training Retrieval-oriented Language Models Via Masked Auto-Encoder (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for dense retrieval are not effective, but there are still challenges. |
| Approach: | They propose a retrieval oriented pre-training paradigm based on Masked Auto-Encoder (MAE) where the sentence embedding is generated from the encoder’s masked input and the original sentence is recovered based upon the sentence embedded and decoded input via mangled language modeling. |
| Outcome: | The proposed model significantly improves the SOTA performance on a wide range of NLP benchmarks, like BEIR and MS MARCO. |
CQR-SQL: Conversational Question Reformulation Enhanced Context-Dependent Text-to-SQL Parsers (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing text-to-SQL methods focus on making full use of history context, but neglect to explicitly comprehend the schema and conversational dependency. |
| Approach: | They propose a CQR-SQL that explicitly exploits schema and conversational dependency for multi-turn SQL parsing. |
| Outcome: | The proposed method exploits schema and contextual dependency for multi-turn SQL parsing. |
ChemAmp: Amplified Chemistry Tools via Composable Agents (2026.findings-acl)
Copied to clipboard
Zhucong Li, Powei Chang, Jin Xiao, Zhijian Zhou, Qianyu He, Jiaqing Liang, Fenglei Cao, Xu Yinghui, Yuan Qi
| Challenge: | LLM-based agents are powerful tools for automating complex scientific workflows, especially in chemistry, but their single-task performance is limited by tool constraints. |
| Approach: | They propose a framework that optimizes the collective capabilities of specialized tools by dynamic coordination within individual tasks. |
| Outcome: | The proposed framework outperforms chemistry-specialized models, generalist LLMs, and agent systems with tool orchestration. |
Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Autoregressive models assign low probabilities to tokens that need corrections . grammatical error correction (GEC) is widely applied to natural language processing tasks . |
| Approach: | They propose to use a non-autoregressive model as an auxiliary model to train GEC models to correct grammatical errors in sentences. |
| Outcome: | The proposed method outperforms baselines on English and Chinese GEC tasks significantly. |
Zero-Resource Hallucination Prevention for Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for detecting hallucinations post-generation suffer from inconsistent performance due to the influence of instruction format and model style. |
| Approach: | They propose a new technique that evaluates the model’s familiarity with the concepts present in the input instruction and withholding the generation of response in case of unfamiliar concepts under the zero-resource setting. |
| Outcome: | The proposed technique shows superior performance across four different large language models and demonstrates that it can be used to mitigate hallucinations in LLMs. |
Task-Related In-Context Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Standard in-context learning assumes identical output spaces between test and retrieval datasets . however, in practice, these datasets can be fully aligned, partially alignes, or fully disjoint in label space . |
| Approach: | They propose a framework for in-context learning under output-space mismatch . they identify demonstrations relevant to the test label space via a Bayesian probabilistic criterion . |
| Outcome: | The proposed framework achieves state-of-the-art results across three LLMs, three task types, and four datasets. |
Jailbreaking Prompt Attack: A Controllable Adversarial Attack against Diffusion Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Text-to-image (T2I) models can be used to generate harmful content such as sexually explicit, unfaithful, and misleading or Not-Safe-for-Work (NSFW) images. |
| Approach: | They propose a more practical and universal attack that does not require the presence of a target model. |
| Outcome: | The proposed attack bypasses both text and image safety checkers while preserving high semantic alignment with the target prompt. |
Reinforcing Agentic Search Via Reward Density Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) is a promising approach for enhancing agentic search, but its performance is often hindered by reward sparsity . |
| Approach: | They propose a new research problem to improve the reward obtained per unit of exploration cost by using a system that decomposes long-horizon tasks into intermediate objectives and assigns process-level rewards to provide denser learning signals. |
| Outcome: | The proposed framework outperforms strong baselines on several agentic search benchmarks and achieves comparable performance to that of advanced proprietary models. |
Augmenting Large Language Model Translators via Translation Memories (2023.findings-acl)
Copied to clipboard
Yongyu Mu, Abudurexiti Reheman, Zhiquan Cao, Yuchun Fan, Bei Li, Yinqiao Li, Tong Xiao, Chunliang Zhang, Jingbo Zhu
| Challenge: | Using translation memories (TMs) as prompts is a promising approach to in-context learning of machine translation models. |
| Approach: | They propose to use translation memories (TMs) as prompts to prompt large language models (LLMs) they find that the ability of LLMs to "understand" prompts is helpful . |
| Outcome: | The results are comparable to state-of-the-art NMT systems with bilingual data and are tuned on downstream tasks. |
Com2 : A Causal-Guided Benchmark for Exploring Complex Commonsense Reasoning in Large Language Models (2025.acl-long)
Copied to clipboard
Kai Xiong, Xiao Ding, Yixin Cao, Yuxiong Yan, Li Du, Yufei Zhang, Jinglong Gao, Jiaqian Liu, Bing Qin, Ting Liu
| Challenge: | Existing works focus on complex tasks like math and code, while complex commonsense reasoning remains underexplored due to its uncertainty and lack of structure. |
| Approach: | They propose to build a benchmark for large language models based on complex commonsense reasoning based upon causal event graphs and causal theory. |
| Outcome: | The proposed benchmark combines a complex commonsense reasoning benchmark with a detective story to achieve a more challenging subset. |
Benchmarking Automated Clinical Language Simplification: Dataset, Algorithm, and Evaluation (2022.coling-1)
Copied to clipboard
| Challenge: | Existing studies to translate medical jargon into layperson-understandable language focus on accuracy and readability aspects of clinical language. |
| Approach: | They propose to construct a dataset to support automated clinical language simplification and propose a model that mimics the human annotation procedure. |
| Outcome: | The proposed model matches human annotation procedures and achieves state-of-the-art performance compared with baselines. |
BASS: Boosting Abstractive Summarization with Unified Semantic Graph (2021.acl-long)
Copied to clipboard
| Challenge: | Abstractive summarization for long-document or multi-document remains challenging for Seq2Seq as it does not analyze long-distance relations in text. |
| Approach: | They propose a framework for Boosting Abstractive Summarization based on a unified Semantic graph which aggregates co-referent phrases distributing across a long range of context and conveys rich relations between phrases. |
| Outcome: | The proposed framework improves document representation and summary generation process by leveraging the graph structure. |
Complex Logical Query Answering by Calibrating Knowledge Graph Completion Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to solve complex logical queries are not well-calibrated . CKGC is lightweight and effective, allowing the model to quickly converge . |
| Approach: | They propose a method for calibrating KGC models to adapt to complex logical queries . they map the values of predictions of KGC to the range [0, 1] . |
| Outcome: | The proposed method can significantly boost model performance in complex logical query answering task. |
Teaching Language Models to Self-Improve by Learning from Language Feedback (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) generate content that can be untruthful or harmful. |
| Approach: | They propose a method that leverages model feedback for alignment . they use a base language model to generate initial responses, critiqued and refined . |
| Outcome: | The proposed method outperforms strong baselines across diverse tasks and model sizes. |
Dialect-SQL: An Adaptive Framework for Bridging the Dialect Gap in Text-to-SQL (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing Text-to-SQL research focuses on specific database systems, limiting adaptability to different dialects. |
| Approach: | They propose a framework that employs Object Relational Mapping (ORM) code as an intermediate language to bridge this gap. |
| Outcome: | The proposed framework outperforms existing methods that generate SQL queries directly. |
Plug-and-Play Document Modules for Pre-trained Models (2023.acl-long)
Copied to clipboard
Chaojun Xiao, Zhengyan Zhang, Xu Han, Chi-Min Chan, Yankai Lin, Zhiyuan Liu, Xiangyang Li, Zhonghua Li, Zhao Cao, Maosong Sun
| Challenge: | Large-scale pre-trained models have been widely adopted for document-oriented NLP tasks, such as question answering. |
| Approach: | They propose to decouple document encoding from downstream tasks by introducing a document plugin into the backbone of a PTM. |
| Outcome: | The proposed model can encode documents once and for all across different scenarios. |
BIPEFT: Budget-Guided Iterative Search for Parameter Efficient Fine-Tuning of Large Pretrained Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for parameter-efficient fine-tuning are limited by computational and storage requirements. |
| Approach: | They propose a budget-guided iterative search strategy to disentangle binary module and rank dimension search spaces and early selection strategies based on parameter budgets. |
| Outcome: | The proposed method significantly improves search efficiency on public benchmarks. |
Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvement (2025.findings-naacl)
Copied to clipboard
Xiyao Wang, Jiuhai Chen, Zhaoyang Wang, Yuhang Zhou, Yiyang Zhou, Huaxiu Yao, Tianyi Zhou, Tom Goldstein, Parminder Bhatia, Taha Kass-Hout, Furong Huang, Cao Xiao
| Challenge: | Existing methods for visual and language alignment depend on external models or data, leading to uncontrollable and unstable results. |
| Approach: | They propose a framework that enhances visual and language alignment without external dependencies by incorporating an in-context self-critic mechanism that constructs preference pairs for tuning. |
| Outcome: | The proposed framework outperforms existing methods and improves performance on 14 hallucination and comprehensive benchmarks. |
Fusion: Towards Automated ICD Coding via Feature Compression (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to assign ICD codes from unstructured clinical notes are noisy and prone to errors. |
| Approach: | They propose a feature compressed ICD coding model called Fusion to address this problem. |
| Outcome: | The proposed model outperforms existing models on two widely used datasets. |