Papers by Zezhong Wang
DPDLLM: A Black-box Framework for Detecting Pre-training Data from Large Language Models (2024.findings-acl)
Copied to clipboard
Baohang Zhou, Zezhong Wang, Lingzhi Wang, Hongru Wang, Ying Zhang, Kehui Song, Xuhui Sui, Kam-Fai Wong
| Challenge: | Existing methods to detect pretraining data from large language models are unrealistic to them. |
| Approach: | They propose to detect pre-training data from LLM in a black-box way by using GPT-2 as reference model and feed it with sequence probabilities to detect whether it was used to train it. |
| Outcome: | The proposed framework outperforms existing methods on the benchmark datasets and shows that it is effective on different popular LLMs. |
Guaranteeing Knowledge Integration with Joint Decoding for Retrieval-Augmented Generation (2026.acl-long)
Copied to clipboard
Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Yuxi Zhang, Huimin Wang, Yutian Zhao, Yefeng Zheng, Binyang Li, Kam-Fai Wong, Xian Wu
| Challenge: | Retrieval-Augmented Generation (RAG) provides access to external knowledge, but current research focuses on retrieval quality and 'integration bottleneck' . |
| Approach: | They propose a framework that explicitly decouples reasoning from evidence integration by generating an 'Inner-Answer' and a 'Refer-Aswer" they propose 'a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Andswer with the factual precision of the Refer-Adswer at the token level' |
| Outcome: | The proposed framework improves accuracy by 12.1% and reduces hallucinations by 16.3% on five QA benchmarks. |
MlingConf: A Comprehensive Study of Multilingual Confidence Estimation on Large Language Models (2025.findings-acl)
Copied to clipboard
Boyang Xue, Hongru Wang, Rui Wang, Sheng Wang, Zezhong Wang, Yiming Du, Bin Liang, Wenxuan Zhang, Kam-Fai Wong
| Challenge: | Existing studies on LLM confidence estimations in languages other than English have been limited to English. |
| Approach: | They propose to use question-related language to prompt LLMs to assess their confidence in large language models. |
| Outcome: | The proposed model improves on question-related language prompts for LS tasks, while English exhibits notable linguistic dominance in confidence estimations. |
Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors (2026.acl-long)
Copied to clipboard
Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao, Kam-Fai Wong
| Challenge: | Large language models (LLMs) can call tools effectively, but they remain brittle in multi-turn execution. |
| Approach: | They propose a framework that converts execution errors into on-policy corrective supervision within the RL training loop. |
| Outcome: | The proposed framework improves the error recovery rate of Qwen3-8B by 5.7% absolute and overall accuracy by 4.0% on BFCL v4 Multi-Turn. |
MemeReaCon: Probing Contextual Meme Understanding in Large Vision-Language Models (2025.emnlp-main)
Copied to clipboard
Zhengyi Zhao, Shubo Zhang, Yuxi Zhang, Yanxi Zhao, Yifan Zhang, Zezhong Wang, Huimin Wang, Yutian Zhao, Bin Liang, Yefeng Zheng, Binyang Li, Kam-Fai Wong, Xian Wu
| Challenge: | Current approaches focus on isolated meme analysis, either for harmful content detection or standalone interpretation, overlooking a fundamental challenge: the same meme can express different intents depending on its conversational context. |
| Approach: | They propose a benchmark to evaluate how large vision language models understand memes in their original context. |
| Outcome: | The proposed benchmark evaluates how large vision language models understand meme intent in their original context. |
Cue-CoT: Chain-of-thought Prompting for Responding to In-depth Dialogue Questions with LLMs (2023.findings-emnlp)
Copied to clipboard
| 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. |
Stepwise Reasoning Checkpoint Analysis: A Test Time Scaling Method to Enhance LLMs’ Reasoning (2025.emnlp-main)
Copied to clipboard
Zezhong Wang, Xingshan Zeng, Weiwen Liu, Yufei Wang, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu, Kam-Fai Wong
| Challenge: | Existing methods that use Chain-of-Thought suffer from path homogenization and inefficient use of intermediate results. |
| Approach: | They propose a framework that introduces checkpoints between reasoning steps to reduce path homogenization and create fault-tolerant mechanisms. |
| Outcome: | The proposed framework reduces path homogenization and creates fault-tolerant mechanism by utilizing high-quality intermediate results. |
T2: An Adaptive Test-Time Scaling Strategy for Contextual Question Answering (2025.emnlp-main)
Copied to clipboard
Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Huimin Wang, Yutian Zhao, Bin Liang, Yefeng Zheng, Binyang Li, Kam-Fai Wong, Xian Wu
| Challenge: | Existing efficient test-time scaling methods introduce budget constraints or early stop mechanisms to avoid overthinking for straightforward questions but add human bias to the reasoning process. |
| Approach: | They propose a framework that dynamically adapts reasoning depth based on question complexity. |
| Outcome: | Experimental results show that the proposed framework achieves higher accuracy than baseline methods and reduces computational overhead by up to 25.2%. |
SELF-GUARD: Empower the LLM to Safeguard Itself (2024.naacl-long)
Copied to clipboard
| Challenge: | Recent studies have investigated methods to improve the safety of large language models (LLMs) safety training involves fine-tuning the LLM with adversarial samples, which activate the LRM’s capabilities against jailbreak. |
| Approach: | They propose a safety training approach that integrates safety training and safeguards to train the LLM to perform harmfulness detection on its own outputs. |
| Outcome: | The proposed method reduces harmful output and adds a [harmful] or [harmless] tag to the end of the LLM's response. |
ToolFlow: Boosting LLM Tool-Calling Through Natural and Coherent Dialogue Synthesis (2025.naacl-long)
Copied to clipboard
Zezhong Wang, Xingshan Zeng, Weiwen Liu, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu, Kam-Fai Wong
| Challenge: | Large Language Models (LLMs) can be enhanced by using supervised fine-tuning . however, access to fine-timing data can be limited. |
| Approach: | They propose a Graph-based Sampling strategy and a Planned-generation strategy to enhance the coherence between dialogues by using 8,000 synthetic dialogues. |
| Outcome: | The proposed model achieves tool-calling performance comparable to or surpassing GPT-4 while maintaining strong general capabilities. |
Chain-of-Probe: Examining the Necessity and Accuracy of CoT Step-by-Step (2025.findings-naacl)
Copied to clipboard
Zezhong Wang, Xingshan Zeng, Weiwen Liu, Yufei Wang, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu, Kam-Fai Wong
| Challenge: | Current research found the issue of Early Answering in large language models where the models already have an answer before generating the Chain-of-Thought (CoT). |
| Approach: | They propose a method to probe changes in confidence during the model’s reasoning and prioritize answers with correct reasoning among multiple candidates. |
| Outcome: | The proposed method reveals that in a significant number of question-answer cases, CoT appears to be unnecessary and this necessity correlates with the simplicity of the task, defined by the reasoning steps required. |
ReadPrompt: A Readable Prompting Method for Reliable Knowledge Probing (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to prob pre-trained language models (PLMs) lack readability and credibility. |
| Approach: | They propose a method to identify meaningful sentences to serve as prompts to assess the knowledge encoded within pre-trained language models (PLMs). |
| Outcome: | The proposed method achieves state-of-the-art on the current knowledge probing benchmark. |
JoTR: A Joint Transformer and Reinforcement Learning Framework for Dialogue Policy Learning (2024.lrec-main)
Copied to clipboard
Wai-Chung Kwan, Huimin Wang, Hongru Wang, Zezhong Wang, Bin Liang, Xian Wu, Yefeng Zheng, Kam-Fai Wong
| Challenge: | Dialogue policy learning (DPL) aims to determine an abstract representation (also known as action) to guide what the response should be. |
| Approach: | They propose a joint Transformer-based model that generates a token-grained policy that allows more dynamic dialogue action generation without the need for predefined action candidates. |
| Outcome: | The proposed model outperforms existing models showing improvements of 9% and 13% in success rate and 34% and 37% in diversity of dialogue actions across two benchmark dialogue modeling tasks. |
“I Know Who You Are”: Character-Based Features for Conversational Humor Recognition in Chinese (2022.findings-emnlp)
Copied to clipboard
| Challenge: | a recent study has focused on how to recognize punchlines from dialogues, but has neglected character information. |
| Approach: | They propose a character-fusion conversational humor recognition model that uses character information to recognize punchlines from dialogue. |
| Outcome: | The proposed model improves performance on Chinese sitcoms corpus and punchline identification. |
Towards Robust Personalized Dialogue Generation via Order-Insensitive Representation Regularization (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent studies have focused on developing persona consistent dialogue models . order sensitivity affects the quality and consistency of generated response . |
| Approach: | They propose a model-agnostic framework to improve persona consistent dialogue response generation by concatenating persona texts and dialogue history as a single input sequence. |
| Outcome: | The proposed framework outperforms existing models on the Persona-Chat dataset and shows that it is more robust under different persona orders and more consistent with the persona profile. |
Empower Large Language Model to Perform Better on Industrial Domain-Specific Question Answering (2023.emnlp-industry)
Copied to clipboard
Fangkai Yang, Pu Zhao, Zezhong Wang, Lu Wang, Bo Qiao, Jue Zhang, Mohit Garg, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang
| Challenge: | Large Language Models (LLMs) have gained popularity but lack specific domain knowledge in domain-specific tasks. |
| Approach: | They propose a model interaction paradigm that empowers LLM to achieve better performance on domain-specific tasks where it is not proficient. |
| Outcome: | The proposed approach outperforms the commonly used LLM with retrieval methods in domain-specific tasks. |