Papers by Fei Mi

Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning (2025.acl-long)

Challenge:	Recent work shows that large-scale pretrained language models (PLMs) are effective few-shot learners.
Approach:	They propose a method that treats few-shotlearners as crowdsourcing workers . they propose to use these workers to train models that solve a task well .
Outcome:	The proposed approach treats few-shotlearners as crowdsourcing workers . the resulting annotations can be utilized to train models that solve the task well .

Copied to clipboard

Erxin Yu, Jing Li, Ming Liao, Qi Zhu, Boyang Xue, Minghui Xu, Baojun Wang, Lanqing Hong, Fei Mi, Lifeng Shang

Self-training Improves Pre-training for Few-shot Learning in Task-oriented Dialog Systems (2021.emnlp-main)

Challenge:	Existing approaches to learning from errors synthesize training data by extrapolating from isolated bad cases, thereby failing to generalize the extensive patterns inherent within these cases.
Approach:	They propose a framework that synthesizes more generalized training data from isolated bad cases by extrapolating from isolated cases.
Outcome:	The proposed framework synthesizes more generalized training data to address these model weaknesses.

Copied to clipboard

Fei Mi, Wanhao Zhou, Lingjing Kong, Fengyu Cai, Minlie Huang, Boi Faltings

How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study (2026.acl-long)

Challenge:	Large-scale pre-trained language models have shown promising results for few-shot learning in task-oriented dialog (ToD) systems.
Approach:	They propose a self-training approach that iteratively labels the most confident unlabeled data to train a stronger Student model.
Outcome:	The proposed approach improves state-of-the-art pre-trained models in few-shot learning scenarios for task-oriented dialog (ToD) systems when only a small number of labeled data are available.

Copied to clipboard

Zhexin Zhang, Xian Qi Loye, Victor Shea-Jay Huang, Junxiao Yang, Qi Zhu, Shiyao Cui, Fei Mi, Lifeng Shang, Yingkang Wang, Hongning Wang, Minlie Huang

Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues (2023.findings-emnlp)

Challenge:	Large Reasoning Models have achieved remarkable success on reasoning-intensive tasks, but their enhanced reasoning capabilities do not translate to improved safety performance.
Approach:	They propose to use supervised fine tuning to enhance the safety of Large Reasoning Models.
Outcome:	The proposed method improves the safety of large reasoning models on reasoning-intensive tasks.

Copied to clipboard

Hongru Wang, Minda Hu, Yang Deng, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, Kam-Fai Wong

Modeling Complex Dialogue Mappings via Sentence Semantic Segmentation Guided Conditional Variational Auto-Encoder (2022.findings-emnlp)

Challenge:	Existing knowledge-grounded dialogue systems focus on a single knowledge source or ignore the dependency between multiple knowledge sources.
Approach:	They propose a framework that integrates multiple knowledge sources and dependencies between them.
Outcome:	The proposed framework can produce persona-consistent and knowledge-enhanced responses on a knowledge-grounded dialogue dataset.

Copied to clipboard

Bin Sun, Shaoxiong Feng, Yiwei Li, Weichao Wang, Fei Mi, Yitong Li, Kan Li

Constructing Highly Inductive Contexts for Dialogue Safety through Controllable Reverse Generation (2022.findings-emnlp)

Challenge:	Existing efforts to identify and avoid CDM to facilitate dialogue learning failed to solve the problem.
Approach:	They propose a Sentence Semantic Segmentation guided Conditional Variational Auto-Encoder which can model and take advantage of the CDM data.
Outcome:	The proposed method can model and take advantages of the CDM data.

Copied to clipboard

Zhexin Zhang, Jiale Cheng, Hao Sun, Jiawen Deng, Fei Mi, Yasheng Wang, Lifeng Shang, Minlie Huang

DecompEval: Evaluating Generated Texts as Unsupervised Decomposed Question Answering (2023.acl-long)

Challenge:	Existing methods to detect toxic generation of pretrained language models rely on templates, data extraction, crowdsourcing workers or automatic generation.
Approach:	They propose a method to construct adversarial contexts conditioned on a given response . they augment existing dataset BAD+ and construct a new dataset B AD+ .
Outcome:	The proposed method can detect toxic or biased content in large pretrained language models.

Copied to clipboard

Pei Ke, Fei Huang, Fei Mi, Yasheng Wang, Qun Liu, Xiaoyan Zhu, Minlie Huang

Responsible & Trustworthy NLP

Challenge:	Existing evaluation metrics for natural language generation (NLG) tasks face the challenges on generalization ability and interpretability.
Approach:	They propose a metric that evaluates natural language generation tasks as an instruction-style question answering task and utilizes instruction-tuned pre-trained language models without training on evaluation datasets.
Outcome:	The proposed metric achieves state-of-the-art performance in untrained metrics for evaluating text summarization and dialogue generation, which exhibits strong dimension-level / task-level generalization ability and interpretability.

One Cannot Stand for Everyone! Leveraging Multiple User Simulators to train Task-oriented Dialogue Systems (2023.acl-long)

Copied to clipboard

Yajiao Liu, Xin Jiang, Yichun Yin, Yasheng Wang, Fei Mi, Qun Liu, Xiang Wan, Benyou Wang

Cue-CoT: Chain-of-thought Prompting for Responding to In-depth Dialogue Questions with LLMs (2023.findings-emnlp)

Challenge:	Recent studies have found that Task-oriented Dialogue systems can be more suitable for human users.
Approach:	They propose a framework to optimize ToD systems by leveraging Multiple User SimulaTors.
Outcome:	The proposed framework improves performance on multiWOZ with human evaluations and automatic evaluations.

Copied to clipboard

Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, Kam-Fai Wong

ClusterUCB: Efficient Gradient-Based Data Selection for Targeted Fine-Tuning of LLMs (2025.findings-emnlp)

Challenge:	Existing LLMs generate responses based on the dialogue context, overlooking the underlying linguistic cues about the user status exhibited in the context.
Approach:	They propose a linguistic cue-based chain-of-thoughts method which enhances the LLMs inference with an intermediate reasoning step to find cues exhibited in the dialogue.
Outcome:	The proposed method outperforms standard prompting methods on in-depth dialogue questions and linguistic cues exhibited in the context.

Copied to clipboard

Zige Wang, Qi Zhu, Fei Mi, Minghui Xu, Ruochun Jin, Wenjing Yang

Continual Learning for Natural Language Generation in Task-oriented Dialog Systems (2020.findings-emnlp)

Challenge:	Gradient-based data influence approximation is not feasible in practice.
Approach:	They propose a gradient-based data selection framework with clustering and a modified Upper Confidence Bound algorithm to solve this problem.
Outcome:	The proposed framework can achieve comparable results to the original gradient-based data selection methods while reducing computational consumption.

Copied to clipboard

Fei Mi, Liangwei Chen, Mengjie Zhao, Minlie Huang, Boi Faltings

Enhancing Large Language Models Against Inductive Instructions with Dual-critique Prompting (2024.naacl-long)

Challenge:	Existing neural approaches for natural language generation are typically developed offline for specific domains.
Approach:	They propose a method to expand NLG knowledge incrementally to new domains . major challenge is catastrophic forgetting, meaning a model forgets the knowledge it has learned before .
Outcome:	The proposed method outperforms other methods by effectively mitigating catastrophic forgetting issue.

Copied to clipboard

Rui Wang, Hongru Wang, Fei Mi, Boyang Xue, Yi Chen, Kam-Fai Wong, Ruifeng Xu

Responsible & Trustworthy NLP

Challenge:	Existing studies have focused on how LLMs handle inductive instructions, which may stem from users’ false beliefs or malicious intents.
Approach:	They propose a benchmark of Inductive Instructions where false knowledge is incorporated into instructions in multiple different styles.
Outcome:	The proposed model improves robustness against inductive instructions, despite different inductive styles and complexity.

UniRetriever: Multi-task Candidates Selection for Various Context-Adaptive Conversational Retrieval (2024.lrec-main)

Copied to clipboard

Hongru Wang, Boyang Xue, Baohang Zhou, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Kam-Fai Wong

Low-Resource NLP

Challenge:	Existing methods for retrieving information from a large corpus of data are sub-optimal and low efficiency.
Approach:	They propose a multi-task framework that functions as a universal retriever for three dominant retrieval tasks during the conversation.
Outcome:	The proposed framework can perform persona selection, knowledge selection, and response selection tasks simultaneously.

Pan More Gold from the Sand: Refining Open-domain Dialogue Training with Noisy Self-Retrieval Generation (2022.coling-1)

Copied to clipboard

Yihe Wang, Yitong Li, Yasheng Wang, Fei Mi, Pingyi Zhou, Xin Wang, Jin Liu, Xin Jiang, Qun Liu

CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference (2024.emnlp-main)

Challenge:	Existing methods for generating open-domain dialogue systems underutilize training data.
Approach:	They propose a retrieval-generation training framework that takes advantage of heterogeneous training data by considering them as "evidence" they use BERTScore retrieval framework which gives better qualities of the training data, they show .
Outcome:	The proposed method performs well on zero-shot experiments and is more robust to real-world data.

Copied to clipboard

Erxin Yu, Jing Li, Ming Liao, Siqi Wang, Gao Zuchen, Fei Mi, Lanqing Hong

Robustness in NLP

Challenge:	Existing studies have not noticed the safety risks of large language models . authors evaluated 1,400 questions in multi-turn dialogue coreference .
Approach:	They are the first to evaluate LLM safety in multi-turn dialogue coreference . they created a dataset of 1,400 questions and tested five open-source models .
Outcome:	The study shows that model safety decreases in multi-turn dialogue coreference scenarios . the highest success rate was with the LLaMA2-Chat-7b model, while the lowest was with mistral-7B-Instruct model .

Improved Policy Optimization for Mixture-of-Experts Models: Importance Sampling and Rewarding from an Expert-Centric Perspective (2026.findings-acl)

Copied to clipboard

Yining Qian, Jinpeng Li, Fei Mi, Lifeng Shang, Xiang Zhang

Challenge:	Existing approaches to reinforcement learning (RL) suffer from training instability . existing approaches often ignore token-specific discrepancies in expert assignments .
Approach:	They propose to introduce expert-level importance sampling to reduce complexity of RL . they propose to leverage expert-centric granularity to ensure a rigorous alignment between reward signals and policy updates.
Outcome:	The proposed method outperforms strong baselines across reasoning tasks.

Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue (2023.acl-short)

Copied to clipboard

Bin Sun, Yitong Li, Fei Mi, Fanhu Bie, Yiwei Li, Kan Li

MoralDial: A Framework to Train and Evaluate Moral Dialogue Systems via Moral Discussions (2023.acl-long)

Challenge:	Existing knowledge-grounded dialogue generation models face the hallucination problem . Existing models generate inappropriate knowledge and generate inconsistent responses .
Approach:	They propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework to enhance existing knowledge dialogue models by polarizing optimization objectives and weak knowledge generation ability.
Outcome:	The proposed framework expands existing training sets and smooths the optimization objective that enables models to generate ground-truth with or without gold knowledge.

Copied to clipboard

Hao Sun, Zhexin Zhang, Fei Mi, Yasheng Wang, Wei Liu, Jianwei Cui, Bin Wang, Qun Liu, Minlie Huang

FollowBench: A Multi-level Fine-grained Constraints Following Benchmark for Large Language Models (2024.acl-long)

Challenge:	A moral dialogue system aligned with users’ values could enhance conversation engagement and user connections.
Approach:	They propose a framework to train and evaluate moral dialogue systems based on communication mechanisms of morality and a method to construct moral discussions between simulated users and the dialogue system.
Outcome:	The proposed framework can train and evaluate moral dialogue systems based on simulated users and their values .

Copied to clipboard

Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, Wei Wang

Towards Identifying Social Bias in Dialog Systems: Framework, Dataset, and Benchmark (2022.findings-emnlp)

Challenge:	Existing benchmarks focus on evaluating pure response quality, rather than assessing whether the response follows constraints stated in the instruction.
Approach:	They propose a Multi-level Fine-grained Constraints Following Benchmark for Large Language Models that adds a single constraint to the initial instruction at each increased level.
Outcome:	The proposed model can follow instructions with more constraints, and is deemed to have better instruction-following ability.

Copied to clipboard

Jingyan Zhou, Jiawen Deng, Fei Mi, Yitong Li, Yasheng Wang, Minlie Huang, Xin Jiang, Qun Liu, Helen Meng

Defending Large Language Models Against Jailbreaking Attacks Through Goal Prioritization (2024.acl-long)

Challenge:	a number of safety concerns hinder the deployment of open-domain dialog systems, such as offensive languages and toxic behaviors, such social bias is difficult to detect.
Approach:	They propose a Dial-Bias Framework for analyzing social bias in conversations . they introduce a Chinese social bias dialog dataset and conduct in-depth ablation studies .
Outcome:	The proposed framework is the first annotated Chinese social bias dialog dataset . the proposed framework also provides a fine-grained dialog bias measurement benchmark .

Copied to clipboard

Zhexin Zhang, Junxiao Yang, Pei Ke, Fei Mi, Hongning Wang, Minlie Huang

Robustness in NLP

Challenge:	Despite the success of jailbreaking attacks, there is a lack of effort in defending against them.
Approach:	They propose to integrate goal prioritization at both training and inference stages to counteract this conflict between the goals of being helpful and ensuring safety.
Outcome:	The proposed approach reduces the Attack Success Rate (ASR) of jailbreaking from 66.4% to 3.6% for ChatGPT and reduces it from 71.0% to 6.6% for Llama2-13B.

Compilable Neural Code Generation with Compiler Feedback (2022.findings-acl)

Copied to clipboard

Xin Wang, Yasheng Wang, Yao Wan, Fei Mi, Yitong Li, Pingyi Zhou, Jin Liu, Hao Wu, Xin Jiang, Qun Liu

Programming Languages in NLP

Challenge:	Existing deep-learning approaches model code generation as text generation, but few of them account for compilability of the generated programs.
Approach:	They propose a three-stage pipeline utilizing compiler feedback for compilable code generation to improve compilability.
Outcome:	The proposed pipeline improves compilability of generated programs by combining compiler feedback, language model fine-tuning, and compilable discrimination.

A Synthetic Data Generation Framework for Grounded Dialogues (2023.acl-long)

Copied to clipboard

Jianzhu Bao, Rui Wang, Yasheng Wang, Aixin Sun, Yitong Li, Fei Mi, Ruifeng Xu

Continual Prompt Tuning for Dialog State Tracking (2022.acl-long)

Challenge:	Existing approaches to train grounded dialogues require large amounts of data.
Approach:	They propose a synthetic data generation framework for grounded dialogues that takes knowledge data and heuristics to determine a dialogue flow and incrementally turn it into a dialog.
Outcome:	The proposed framework significantly boosts model performance in training data and low-resource scenarios.

Copied to clipboard

Qi Zhu, Bing Li, Fei Mi, Xiaoyan Zhu, Minlie Huang

Responsible & Trustworthy NLP

Challenge:	Existing methods to train a model on a sequence of tasks are not efficient enough to mitigate catastrophic forgetting.
Approach:	They propose a parameter-efficient framework that prevents forgetting and enables knowledge transfer between tasks by learning and freezing a pre-trained model.
Outcome:	The proposed framework avoids forgetting and enables knowledge transfer between tasks.

UAlign: Leveraging Uncertainty Estimations for Factuality Alignment on Large Language Models (2025.acl-long)

Copied to clipboard

Boyang Xue, Fei Mi, Qi Zhu, Hongru Wang, Rui Wang, Sheng Wang, Erxin Yu, Xuming Hu, Kam-Fai Wong

Semantic Text Processing

Challenge:	Large Language Models (LLMs) often struggle to accurately express factual knowledge, especially in cases where the knowledge boundaries are ambiguous.
Approach:	They propose a framework that leverages Uncertainty estimations to represent knowledge boundaries and incorporates these representations into prompts for LLMs to Align with factual knowledge.
Outcome:	The proposed framework significantly improves the LLMs’ capacities to confidently answer known questions and refuse unknown questions on both in-domain and out-of-domain tasks.

Retrieval-free Knowledge Injection through Multi-Document Traversal for Dialogue Models (2023.acl-long)

Copied to clipboard

Rui Wang, Jianzhu Bao, Fei Mi, Yi Chen, Hongru Wang, Yasheng Wang, Yitong Li, Lifeng Shang, Kam-Fai Wong, Ruifeng Xu

Mixture of insighTful Experts (MoTE): The Synergy of Reasoning Chains and Expert Mixtures in Self-Alignment (2025.acl-long)

Challenge:	Existing research on retrieval-augmented and retrieval free dialogue models focuses on retrieving knowledge from external sources and rely on finely annotated retrieval training data and knowledge-grounded responses.
Approach:	They propose a retrieval-free approach by turning knowledge documents into simulated multi-turn dialogues using a Multi-Document Traversal algorithm.
Outcome:	The proposed approach outperforms retrieval-augmented models while being cheaper and faster at domain transfer.

Copied to clipboard

Zhili Liu, Yunhao Gou, Kai Chen, Lanqing Hong, Jiahui Gao, Fei Mi, Yu Zhang, Zhenguo Li, Xin Jiang, Qun Liu, James Kwok

Masking as an Efficient Alternative to Finetuning for Pretrained Language Models (2020.emnlp-main)

Challenge:	Recent studies show that reasoning abilities contribute significantly to model safety, while integrating Mixture-of-Experts (MoE) architectures can further enhance alignment.
Approach:	They propose a framework that synergistically combines reasoning chains and expert mixtures to improve self-alignment.
Outcome:	The proposed framework improves model safety, jailbreak resistance, and over-refusal capabilities, achieving performance comparable to OpenAI’s state-of-the-art o1 model.

Copied to clipboard

Mengjie Zhao, Tao Lin, Fei Mi, Martin Jaggi, Hinrich Schütze

Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning (2026.acl-long)

Challenge:	Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that our binary masked language models encode information necessary for solving downstream tasks.
Approach:	They propose an efficient method of utilizing pretrained language models where selective binary masks are learned instead of finetuning.
Outcome:	Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that the proposed method yields comparable performance to finetuning, but has a much smaller memory footprint when multiple tasks need to be solved.

Copied to clipboard

Bowen Ding, Yuhan Chen, Jiayang Lyu, Jiyao Yuan, Qi Zhu, Shuangshuang Tian, Dantong Zhu, Futing Wang, Heyuan Deng, Fei Mi, Lifeng Shang, Tao Lin

COLD: A Benchmark for Chinese Offensive Language Detection (2022.emnlp-main)

Challenge:	evolving generic Large Language Models into specialized Large Reasoning Models requires effective post-training.
Approach:	They propose a plasticity-ceiling framework to harness expert trajectories . they establish the Sequential SFT-then-RL pipeline as the superior standard .
Outcome:	The proposed framework overcomes stability and premature convergence deficits in synchronized approaches.

Copied to clipboard

Jiawen Deng, Jingyan Zhou, Hao Sun, Chujie Zheng, Fei Mi, Helen Meng, Minlie Huang

Ethical NLP

Challenge:	Offensive language detection is crucial for maintaining a civilized social media platform and deploying pre-trained language models.
Approach:	They propose a benchmark benchmark for Chinese offensive language analysis including a Chinese Offensive Language Dataset and a baseline detector which is trained on the dataset.
Outcome:	The proposed benchmark contributes to Chinese offensive language detection which is challenging for existing resources.

Role Prompting Guided Domain Adaptation with General Capability Preserve for Large Language Models (2024.findings-naacl)

Copied to clipboard

Rui Wang, Fei Mi, Yi Chen, Boyang Xue, Hongru Wang, Qi Zhu, Kam-Fai Wong, Ruifeng Xu

Semantic Text Processing

Challenge:	Large Language Models (LLMs) suffer catastrophic forgetting when tailored to specific domains . authors present a novel approach to manage multi-domain LLM adaptation .
Approach:	They propose a strategy to manage multi-domain LLM adaptation using self-distillation and role integration.
Outcome:	The proposed model alleviates catastrophic forgetting and inter-domain confusion while maintaining robust general capabilities.

ReSee: Responding through Seeing Fine-grained Visual Knowledge in Open-domain Dialogue (2023.emnlp-main)

Copied to clipboard

Haoqin Tu, Yitong Li, Fei Mi, Zhongliang Yang

Multimodality

Challenge:	Existing multimodal dialogue systems are limited by the scale and quality of available datasets or the coarse concept of visual knowledge.
Approach:	They propose to explicitly split visual knowledge into finer granularity and turn-level . they propose a framework to add visual representation into vanilla dialogue models .
Outcome:	The proposed framework outperforms state-of-the-art methods on automatic and human evaluations.

Dynamic Stochastic Decoding Strategy for Open-Domain Dialogue Generation (2024.findings-acl)

Copied to clipboard

Yiwei Li, Fei Mi, Yitong Li, Yasheng Wang, Bin Sun, Shaoxiong Feng, Kan Li

AEG: Argumentative Essay Generation via A Dual-Decoder Model with Content Planning (2022.emnlp-main)

Challenge:	Stochastic sampling strategies are not widely used in open-domain dialogue systems.
Approach:	They propose a dynamic decoding strategy which can adjust the decoding space w.r.t. different contexts.
Outcome:	The proposed decoding strategy can improve the performance of pre-trained models when coupled with four well-used stochastic decoding algorithms.

Copied to clipboard

Jianzhu Bao, Yasheng Wang, Yitong Li, Fei Mi, Ruifeng Xu