Papers by Mi Zhang
Friend-training: Learning from Models of Different but Related Tasks (2023.eacl-main)
Copied to clipboard
| Challenge: | Current self-training methods focus on improving model performance on a single task. |
| Approach: | They propose a cross-task self-training framework where models trained to do different tasks are used in iterative training, pseudo-labeling, and retraining processes to help each other for better selection of pseudo-labeled labels. |
| Outcome: | The proposed framework achieves the best performance compared to baselines on two dialogue understanding tasks. |
Recall with Reasoning: Chain-of-Thought Distillation for Mamba’s Long-Context Memory and Extrapolation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing long-context memory methods such as Mamba struggle with long-constituency when the length of the processed text exceeds the model's training length. |
| Approach: | They propose a method that uses chain-of-thought summarization to teach Mamba to actively recall and reason over long contexts. |
| Outcome: | Experiments on LONGMEMEVAL and HELMET show that RwR outperforms existing long-term memory methods while preserving short-context capabilities. |
Inconsistent dialogue responses and how to recover from them (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to assess and bolster utterance consistency of chat systems have been shown difficult to detect. |
| Approach: | They propose to use annotators to write dialogue responses and recovery utterances to assess and bolster utteration consistency of chat systems. |
| Outcome: | The proposed dataset significantly improves the detection and resolution of inconsistencies in chat conversations. |
Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation (2024.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to addressing factual inaccuracies require high-quality human factuality annotations to mitigate these hallucinations. |
| Approach: | They propose to leverage the self-evaluation capability of an LLM to provide training signals that steer the model towards factuality. |
| Outcome: | The proposed approach significantly improves factual accuracy over LLMs across three key knowledge-intensive tasks on TruthfulQA and BioGEN. |
WebRollback: Enhancing Web Agents with Explicit Rollback Mechanisms (2026.eacl-short)
Copied to clipboard
| Challenge: | Recent studies have adopted a greedy one-way search strategy to deal with dynamic web environments. |
| Approach: | They propose to integrate a rollback mechanism into web agents to allow them to revert back to a previous state in navigation trajectory. |
| Outcome: | The proposed method is able to revert back to a previous state in its navigation trajectory, allowing the models to directly control the search process. |
D-Artemis: A Deliberative Cognitive Framework for Mobile GUI Multi-Agents (2026.findings-acl)
Copied to clipboard
Hongze Mi, Yibo Feng, WenJie Lu, Yuqi Wang, Jinyuan Li, Song Cao, He Cui, Tengfei Tian, Xuelin Zhang, Haotian Luo, Di Sun, Jun Fang, Hua Chai, Naiqiang Tan, Gang Pan
| Challenge: | Graphical User Interface (GUI) agents aim to automate a wide spectrum of human tasks by emulating user interaction. |
| Approach: | They propose a deliberative framework that leverages a fine-grained tip retrieval mechanism to inform its decision-making process. |
| Outcome: | The proposed framework achieves SOTA among open-source general models on AndroidWorld and ScreenSpot-V2 . it leverages a fine-grained, app-specific tip retrieval mechanism to inform its decision-making process . |
How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study (2026.acl-long)
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
| 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. |
Crossing the Reward Bridge: Expanding Reinforcement Learning with Verifiable Rewards Across Diverse Domains (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has been effective on structured tasks, but its reliance on simple, rule-based verifiers creates a bottleneck. |
| Approach: | They propose a framework that uses a generative verifier to provide soft, probabilistic rewards. |
| Outcome: | The proposed framework outperforms existing models up to 10x their size and can be scalable and effective. |
ReasoningGuard: Safeguarding Large Reasoning Models with Inference-time Safety Aha Moments (2026.acl-long)
Copied to clipboard
| Challenge: | Existing defenses for Large Reasoning Models (LRMs) depend on costly fine-tuning and additional expert knowledge, which limits their scalability. |
| Approach: | They propose an inference-time safeguard for Large Reasoning Models that injects safety aha moments into the reasoning process to guide the model towards harmless yet helpful reasoning. |
| Outcome: | The proposed safeguard outperforms nine existing safeguards while avoiding common exaggerated safety issues. |
Constructing Highly Inductive Contexts for Dialogue Safety through Controllable Reverse Generation (2022.findings-emnlp)
Copied to clipboard
| 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. |
LatticeGen: Hiding Generated Text in a Lattice for Privacy-Aware Large Language Model Generation on Cloud (2024.findings-naacl)
Copied to clipboard
Mengke Zhang, Tianxing He, Tianle Wang, Lu Mi, Niloofar Mireshghallah, Binyi Chen, Hao Wang, Yulia Tsvetkov
| Challenge: | Currently, the server controls the generated text, but users can't keep it private . prompted generation is a common interaction paradigm for large language models on cloud . |
| Approach: | They propose a protocol where the server handles most of the computation while the client controls the sampling operation. |
| Outcome: | The proposed protocol protects both prompt and generation under strong attacks. |
Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data (2026.acl-short)
Copied to clipboard
| Challenge: | Strong base models saturate benchmarks, resulting in weaker performance, a paradox . a new approach to Reinforcement Learning (RL) is needed to improve performance . |
| Approach: | They propose a method that uses constrained uniform top-k sampling to flatten the local optimization landscape by sampling uniformly from constrained high-confidence candidates. |
| Outcome: | Experiments show that the proposed approach prevents policy degeneration and boosts out-of-domain generalization. |
Low-Bit Quantization Favors Undertrained LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Larger models or those trained on fewer tokens exhibit less quantization-induced degradation (QiD), while smaller, well-trained models face significant performance losses. |
| Approach: | They propose to use QiD to measure an LLM’s training levels and determine the number of training tokens required for fully training LLMs of various sizes. |
| Outcome: | The proposed scaling laws can predict the quantization performance of different-sized LLMs trained with tokens. |
WebAggregator: Enhancing Compositional Reasoning Capabilities of Deep Research Agent Foundation Models (2026.acl-long)
Copied to clipboard
Rui Wang, Ce Zhang, Jun-Yu Ma, Jianshu Zhang, Hongru Wang, Yi Chen, Boyang Xue, Tianqing Fang, Zhisong Zhang, Hongming Zhang, Haitao Mi, Dong Yu, Kam-Fai Wong
| Challenge: | Existing agentic systems are retrieval-heavy but reasoning-light . current systems lack compositional reasoning, a key component of deep research . |
| Approach: | They propose a data synthesis pipeline WebAggregator to shift agentic paradigm . they use Proactive Explorer to collect interconnected knowledge and Compositional Logic Proposer to weave knowledge into complex questions . |
| Outcome: | The proposed pipeline surpasses GPT-4.1 and matches Claude-3.7-Sonnet on GAIA, WebWalkerQA, and XBench. |
OpenFact: Factuality Enhanced Open Knowledge Extraction (2023.tacl-1)
Copied to clipboard
Linfeng Song, Ante Wang, Xiaoman Pan, Hongming Zhang, Dian Yu, Lifeng Jin, Haitao Mi, Jinsong Su, Yue Zhang, Dong Yu
| Challenge: | Existing OIE systems organize knowledge into subject-relation-object (SRO) triplets, and they use templates to extract such knowledge triplet. |
| Approach: | They propose a framework to handle expressiveness and groundedness in OpenFact . they propose to use templates, extra constraints, and adopt human efforts to ensure that most triplets contain enough details. |
| Outcome: | The proposed framework improves expressiveness and groundedness of OpenFact . it is more accurate and denser than OPIEC-Linked, which is grounded to Wikidata . |
SafeConv: Explaining and Correcting Conversational Unsafe Behavior (2023.acl-long)
Copied to clipboard
| Challenge: | Existing datasets do not provide enough annotation to explain unsafe behavior . current chatbots generate toxic and offensive responses, which can be dangerous . |
| Approach: | They construct a dataset called SafeConv that provides comprehensive annotations for chatbots . they compare safe alternatives to rewrite unsafe responses . |
| Outcome: | The proposed model can explain unsafe behavior and detoxify chatbots, the authors show . the proposed model is able to detect unsafe utterances, extract unsafe spans, and convert unsafe responses to safe versions. |
SlowBERT: Slow-down Attacks on Input-adaptive Multi-exit BERT (2023.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models such as Google’s BERT have been gaining significant improvements to various down-stream applications, but the enormous training and inference cost severely hinders its practice on real-time applications and hardwareconstrained edge devices. |
| Approach: | They propose a slow-down attack on input-adaptive multi-exit BERT where the adversary imperceptibly modifies the input texts to drastically increase the inference cost. |
| Outcome: | The proposed attack on input-adaptive multi-exit BERT dramatically increases the average inference cost by 4.57, which would hurt the service quality of multi-extit BRT in practice, e.g., increasing the real-time cloud services’ response times for online users. |
RespiraMFM: A Multimodal Foundation Model with Contrastive Audio-Language Alignment for Respiratory Disease Identification (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models for respiratory diseases rely on audio inputs, but they lack generalizability and diagnostic precision. |
| Approach: | They propose a multimodal foundation model that integrates respiratory sounds with medical history and symptoms to enhance diagnostic accuracy and disease detection capabilities. |
| Outcome: | The proposed model improves AUROC and zero-shot tasks across five respiratory diseases using real-world datasets. |
Improved Policy Optimization for Mixture-of-Experts Models: Importance Sampling and Rewarding from an Expert-Centric Perspective (2026.findings-acl)
Copied to clipboard
| 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. |
RobuT: A Systematic Study of Table QA Robustness Against Human-Annotated Adversarial Perturbations (2023.acl-long)
Copied to clipboard
Yilun Zhao, Chen Zhao, Linyong Nan, Zhenting Qi, Wenlin Zhang, Xiangru Tang, Boyu Mi, Dragomir Radev
| Challenge: | Existing Table QA models are vulnerable to task-specific perturbations, such as replacing key question entities or shuffling table columns. |
| Approach: | They propose to use large language models to generate adversarial examples to enhance training, which significantly improves the robustness of Table QA models. |
| Outcome: | The proposed model significantly improves on existing Table QA models against human-annotated adversarial perturbations. |
Self-Tuning: Instructing LLMs to Effectively Acquire New Knowledge through Self-Teaching (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to keeping large language models current involve continued pre-training on new documents. |
| Approach: | They propose a learning framework that augments documents with knowledge-intensive tasks created in a self-supervised manner, focusing on memorization, comprehension, and self-reflection. |
| Outcome: | The proposed learning framework improves an LLM’s ability to acquire new knowledge from unseen raw documents through self-teaching. |
Convolution over Hierarchical Syntactic and Lexical Graphs for Aspect Level Sentiment Analysis (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for aspect-level sentiment classification ignore corpus level word co-occurrence information . a novel architecture convolutes over hierarchical syntactic and lexical graphs . |
| Approach: | They propose a novel architecture which convolutes over hierarchical syntactic and lexical graphs . they employ a global lexical graph to encode corpus level word co-occurrence information . |
| Outcome: | The proposed architecture outperforms the state-of-the-art methods on five bench- mark datasets. |
Creating a Lens of Chinese Culture: A Multimodal Dataset for Chinese Pun Rebus Art Understanding (2025.findings-acl)
Copied to clipboard
Tuo Zhang, Tiantian Feng, Yibin Ni, Mengqin Cao, Ruying Liu, Kiana Avestimehr, Katharine Butler, Yanjun Weng, Mi Zhang, Shrikanth Narayanan, Salman Avestimehr
| Challenge: | a new study examines the performance of large vision-language models in understanding art . the Pun Rebus Art Dataset is a multimodal dataset for art understanding rooted in traditional Chinese culture . |
| Approach: | They propose a multimodal dataset for art understanding deeply rooted in traditional Chinese culture . they aim to facilitate the development of VLMs that can better understand culturally specific content . |
| Outcome: | The proposed dataset shows that state-of-the-art VLMs struggle with these tasks . the data will facilitate the development of VLM models that can better understand culturally specific content . |
MEIT: Multimodal Electrocardiogram Instruction Tuning on Large Language Models for Report Generation (2025.findings-acl)
Copied to clipboard
Zhongwei Wan, Che Liu, Xin Wang, Chaofan Tao, Hui Shen, Jing Xiong, Rossella Arcucci, Huaxiu Yao, Mi Zhang
| Challenge: | Recent studies have focused on classifying cardiac conditions using ECG data but have overlooked ECG report generation, which is time-consuming and requires clinical expertise. |
| Approach: | They propose a Multimodal ECG Instruction Tuning framework that extends the capability of large language models (LLMs) for the task. |
| Outcome: | The proposed framework outperforms open-source LLMs and LLM backbones across two large-scale ECG datasets. |
WebCoT: Enhancing Web Agent Reasoning by Reconstructing Chain-of-Thought in Reflection, Branching, and Rollback (2025.findings-emnlp)
Copied to clipboard
Minda Hu, Tianqing Fang, Jianshu Zhang, Jun-Yu Ma, Zhisong Zhang, Jingyan Zhou, Hongming Zhang, Haitao Mi, Dong Yu, Irwin King
| Challenge: | Web agents powered by Large Language Models lack the ability to perform in uncertain web environments. |
| Approach: | They propose to reconstruct web agents' reasoning skills into chain-of-thought rationales by fine-tuning their LLM backbone into a web-based model. |
| Outcome: | The proposed approach significantly improves the agent self-improving benchmark OpenWebVoyager, demonstrating that it can be used to improve the agent's reasoning skills. |
Argus: Benchmarking and Enhancing Vision-Language Models for 3D Radiology Report Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing work on 3D radiograph report generation focuses on 2D images, but 3D medical images provide more comprehensive diagnostic information. |
| Approach: | They propose a comprehensive training recipe for building high-performing VLMs for 3DRRG using a publicly available 3D CT-report dataset. |
| Outcome: | The proposed model achieves superior performance across different model sizes and input 3D medical image resolutions. |
WebEvolver: Enhancing Web Agent Self-Improvement with Co-evolving World Model (2025.emnlp-main)
Copied to clipboard
| Challenge: | Agent self-improvement, where agents train their underlying Large Language Model (LLM) on self-sampled trajectories, shows promising results but often stagnates in web environments due to limited exploration and under-utilization of pretrained web knowledge. |
| Approach: | They propose a co-evolving Large Language Model (LLM) that predicts the next observation based on current observation and action within the web environment. |
| Outcome: | The proposed framework shows that agents can perform better in real-world web environments without using any distillation from more powerful close-sourced models. |
MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference (2025.naacl-long)
Copied to clipboard
| Challenge: | Long-context Multimodal Large Language Models (MLLMs) require substantial computational resources as their multimodal Key-Value (KV) cache grows with increasing input lengths, challenging memory and time efficiency. |
| Approach: | They propose a dynamic multimodal KV cache allocation strategy that dynamically allocating KV size based on attention entropy to better adapt to multimodal interactions. |
| Outcome: | The proposed model achieves up to 72% KV cache memory reduction and 2.82 faster decoding speeds while maintaining or enhancing performance on various multimodal tasks in a long context. |
Learning a Grammar Inducer from Massive Uncurated Instructional Videos (2022.emnlp-main)
Copied to clipboard
| Challenge: | aims to find more accurate syntactic grammars for accompanying text using video data. |
| Approach: | They build a video-aided grammar induction model that can learn video-span correlation without manual features. |
| Outcome: | The proposed model can learn video-span correlation without manual features adopted by previous systems. |
MoralDial: A Framework to Train and Evaluate Moral Dialogue Systems via Moral Discussions (2023.acl-long)
Copied to clipboard
| 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 . |
Measure Twice, Click Once: Co-evolving Proposer and Visual Critic via Reinforcement Learning for GUI Grounding (2026.acl-long)
Copied to clipboard
| Challenge: | Graphical User Interface (GUI) grounding requires mapping natural language instructions to precise pixel coordinates due to visually homogeneous elements and dense layouts. |
| Approach: | They propose to replace static consistency strategies with a learnable selection mechanism that selects the optimal target by critiquing its own proposals rendered on the screenshot. |
| Outcome: | The proposed model significantly improves both grounding and critiquing capabilities over 6 benchmarks. |
Defending Large Language Models Against Jailbreaking Attacks Through Goal Prioritization (2024.acl-long)
Copied to clipboard
| 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. |
Mixture of insighTful Experts (MoTE): The Synergy of Reasoning Chains and Expert Mixtures in Self-Alignment (2025.acl-long)
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
| 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. |
SVD-LLM V2: Optimizing Singular Value Truncation for Large Language Model Compression (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for compressing Large Language Models suffer from significant truncation losses. |
| Approach: | They propose a novel method that optimizes singular value truncation in SVD compression . they use dynamic compression ratio allocation to balance the large tuncation loss . |
| Outcome: | The proposed method outperforms current state-of-the-art methods on ten datasets and five models on various scales. |