Papers by Zhexin Zhang
Automatic Comment Generation for Chinese Student Narrative Essays (2022.emnlp-demos)
Copied to clipboard
| Challenge: | Existing studies focus on giving discrete scores for holistic quality or distinct traits, but real-world teachers usually provide detailed comments in natural language, which are more informative than single scores. |
| Approach: | They propose a model which generates comments for specified segments from given student narrative essays using a human-written Chinese dataset. |
| Outcome: | The proposed model outperforms baselines and has 91% success rate. |
ShieldLM: Empowering LLMs as Aligned, Customizable and Explainable Safety Detectors (2024.findings-emnlp)
Copied to clipboard
Zhexin Zhang, Yida Lu, Jingyuan Ma, Di Zhang, Rui Li, Pei Ke, Hao Sun, Lei Sha, Zhifang Sui, Hongning Wang, Minlie Huang
| Challenge: | Existing tools for detecting safety issues in LLMs are expensive and inefficient. |
| Approach: | They propose an LLM-based safety detector which annotates the safety of queries and provides explanations for its decisions. |
| Outcome: | The proposed detector outperforms baselines on four sets of query-response pairs and is effective as a safety evaluator for advanced LLMs. |
InstructSafety: A Unified Framework for Building Multidimensional and Explainable Safety Detector through Instruction Tuning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing safety detection systems have limitations in terms of their versatility and interpretability. |
| Approach: | They introduce a safety detection framework that unifies 7 common sub-tasks into a uniform formulation and process 39 human-annotated datasets for instruction tuning. |
| Outcome: | The proposed framework unifies 7 common sub-tasks into a uniform formulation and then runs on 39 human-annotated datasets to fine-tune it. |
LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety (2026.acl-long)
Copied to clipboard
Junxiao Yang, Haoran Liu, Jinzhe Tu, Jiale Cheng, Zhexin Zhang, Shiyao Cui, Jiaqi Weng, Jialing Tao, Hui Xue, Hongning Wang, Han Qiu, Minlie Huang
| Challenge: | Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resourced languages. |
| Approach: | They propose language-agnostic semantic alignment (LASA) which anchors safety alignment directly in semantic bottlenecks. |
| Outcome: | The proposed approach significantly improves safety across all languages: average attack success rate drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwend3 Instruct models (7B–32B). |
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. |
Guiding not Forcing: Enhancing the Transferability of Jailbreaking Attacks on LLMs via Removing Superfluous Constraints (2025.acl-long)
Copied to clipboard
| Challenge: | Jailbreaking attacks can effectively induce unsafe behaviors in Large Language Models (LLMs). |
| Approach: | They propose a conceptual framework to elucidate transferability of gradient-based jailbreaking methods . they identify superfluous constraints as significant barriers to improved transferability . |
| Outcome: | The proposed method increases the overall transfer attack success rate (T-ASR) across target and source models with varying safety levels from 18.4% to 50.3% while improving stability and controllability of jailbreak behaviors. |
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. |
Self-Supervised Sentence Polishing by Adding Engaging Modifiers (2023.acl-demo)
Copied to clipboard
| Challenge: | a typical way to polish sentences is to add engaging modifiers, which enhance the meaning of the sentence. |
| Approach: | They propose a task that requires polishing sentences while maintaining fluency . they remove engaging modifiers from public resources and fine-tune LongLM to reconstruct original sentences from corrupted ones. |
| Outcome: | The proposed model generates more engaging sentences with suitable modifiers than strong baselines while keeping fluency. |
Selecting Stickers in Open-Domain Dialogue through Multitask Learning (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to select appropriate stickers in open-domain dialogues have not been explored. |
| Approach: | They propose a multitask learning method consisting of three auxiliary tasks to combine multimodal information to enhance the understanding of dialogue history, emotion and semantic meaning of stickers. |
| Outcome: | The proposed model can combine multimodal information and achieve significantly higher accuracy over strong baselines. |
ETHICIST: Targeted Training Data Extraction Through Loss Smoothed Soft Prompting and Calibrated Confidence Estimation (2023.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that pre-trained language models memorize a considerable fraction of training data, leading to privacy risk of information leakage. |
| Approach: | They propose a method for targeted training data extraction using a smoothed soft prompting and calibrated confidence estimation. |
| Outcome: | The proposed method significantly improves the extraction performance on a recently proposed public benchmark. |
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 . |
LongSafety: Evaluating Long-Context Safety of Large Language Models (2025.acl-long)
Copied to clipboard
Yida Lu, Jiale Cheng, Zhexin Zhang, Shiyao Cui, Cunxiang Wang, Xiaotao Gu, Yuxiao Dong, Jie Tang, Hongning Wang, Minlie Huang
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating long sequences. |
| Approach: | They propose a benchmark to evaluate LLM safety in open-ended long-context tasks . they find that relevant context and extended input sequences can exacerbate safety risks . |
| Outcome: | The proposed benchmark identifies significant safety vulnerabilities in 16 LLMs . strong safety performance in short-context scenarios does not correlate with safety in long-contact tasks . |
Unveiling the Implicit Toxicity in Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies focus on probing toxic outputs that can be easily detected with existing toxicity classifiers, but LLMs can generate diverse implicit toxic output that are difficult to detect via simply zero-shot prompting. |
| Approach: | They propose a reinforcement learning based attacking method to induce the implicit toxic outputs in large language models by fine-tuning toxicity classifiers. |
| Outcome: | The proposed method generates implicit toxic outputs that are difficult to detect via zero-shot prompting on five widely-adopted toxicity classifiers. |
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. |
Persona-Guided Planning for Controlling the Protagonist’s Persona in Story Generation (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to control the protagonist's persona in story generation are implicitly and sparsely embodied in stories, so we propose a planning-based generation model called ConPer to explicitly model the relationship between personas and events. |
| Approach: | They propose a model to control the protagonist's persona in story generation by predicting one target sentence and planning the plot as a sequence of keywords with the guidance of the predicted persona-related events and commonsense knowledge. |
| Outcome: | The proposed model outperforms state-of-the-art models for generating more coherent and persona-controllable stories. |
New Terms, New Toxicity: Consensus-based Chinese Neologism Toxicity Detection via Search-Augmented LLMs (2026.acl-long)
Copied to clipboard
Shiyao Cui, QingLin Zhang, Di Wang, Yida Lu, Zhexin Zhang, Jinhua Gao, Jinglin Yang, Min He, Han Qiu, Minlie Huang
| Challenge: | Neologisms can foster new linguistic consensus by stabilizing shared meanings and usage in common communicative norms. |
| Approach: | They propose a taxonomy that captures the origins and consensus-verification criteria of toxic neologisms . they propose 'SeTox' framework that integrates real-time web context for naeologim detection . |
| Outcome: | The proposed framework outperforms large-scale models in detecting neologism toxicity. |
OpenMEVA: A Benchmark for Evaluating Open-ended Story Generation Metrics (2021.acl-long)
Copied to clipboard
Jian Guan, Zhexin Zhang, Zhuoer Feng, Zitao Liu, Wenbiao Ding, Xiaoxi Mao, Changjie Fan, Minlie Huang
| Challenge: | Existing automatic metrics are observed to correlate poorly with human evaluation. |
| Approach: | They propose to use OpenMEVA to evaluate open-ended story generation metrics. |
| Outcome: | The proposed test suite assesses the capabilities of open-ended story generation metrics on annotated stories and auto-constructed test examples. |