Papers by Zhaohan Zhang
CoCo: Coherence-Enhanced Machine-Generated Text Detection Under Low Resource With Contrastive Learning (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent proposed methods fail to consider the linguistic structure of texts and lack the ability to handle the low-resource problem. |
| Approach: | They propose a coherence-based contrastive learning model named CoCo to detect MGTs under low-resource scenario. |
| Outcome: | The proposed model outperforms state-of-the-art methods on two datasets and two self-constructed datasets. |
Iron Sharpens Iron: Defending Against Attacks in Machine-Generated Text Detection with Adversarial Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing MGT detectors are vulnerable to simple perturbations and adversarial attacks. |
| Approach: | They propose an adversarial framework for training a robust machine-generated text detector called GREedy Adversary PromoTed DefendER. |
| Outcome: | The proposed framework reduces the Attack Success Rate (ASR) by 0.67% compared with SOTA defense methods. |
Confidence Should Be Calibrated More Than One Turn Deep (2026.acl-long)
Copied to clipboard
| Challenge: | Existing work on confidence estimation and calibration focuses on single-turn settings . existing work on multi-turn calibration ignores the risks and potential of multi-turned conversations . |
| Approach: | They propose a multi-turn calibration task that reframes calibration from a static property into a dynamic challenge central to reliable multi- turn conversations. |
| Outcome: | The proposed model minimizes ECE@T and leverages ConfChat to improve confidence . the proposed model preserves and even enhances model performance in multi-turn interactions. |
GrACE: A Generative Approach to Better Confidence Elicitation and Efficient Test-Time Scaling in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for assessing the reliability of Large Language Models (LLMs) by confidence elicitation require expensive computational overhead or suffer from poor calibration, making them unreliable for real-world deployment. |
| Approach: | They propose a Generative Approach to Confidence Elicitation that enables reliable confidence elicitation for Large Language Models. |
| Outcome: | The proposed method achieves the best discriminative capacity and calibration on open-ended tasks without resorting to additional sampling or an auxiliary model. |
PromptFix: Few-shot Backdoor Removal via Adversarial Prompt Tuning (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that pre-trained language models can be backdoored such that model behavior is manipulated when trigger tokens are presented. |
| Approach: | They propose a backdoor mitigation strategy for NLP models via adversarial prompt-tuning in few-shot settings that uses two extra sets of soft tokens which approximate the trigger and counteract it respectively. |
| Outcome: | The proposed method keeps model parameters intact and only utilizes two extra sets of soft tokens which approximate the trigger and counteract it respectively. |
Can Reasoning Path still be Effective as Input? Bridging Post-Reasoning to Chain-of-Thought Compression (2026.acl-long)
Copied to clipboard
Chengzhengxu Li, Xiaoming Liu, Zhaohan Zhang, Shengchao Liu, Guoxin Ma, Yu Lan, Cong Wang, Chao Shen
| Challenge: | Existing work on reducing CoT generation in reasoning impairs the necessary information for deriving the correct answer. |
| Approach: | They propose a reasoning paradigm that takes CoT as a part of context to simplify the reasoning task for Large Language Models (LLMs). |
| Outcome: | The proposed framework reduces the generation length of LLMs, but its effectiveness hinges on the efficiency and reliability of the contextual CoT generation. |
Does DetectGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned Contrastive Learning Detector would be Better (2024.acl-long)
Copied to clipboard
Shengchao Liu, Xiaoming Liu, Yichen Wang, Zehua Cheng, Chengzhengxu Li, Zhaohan Zhang, Yu Lan, Chao Shen
| Challenge: | Existing methods to detect MGT from human-written texts are inadequate . existing methods are fine-tuned and zero-shot metric-based, but they can be more accurate. |
| Approach: | They propose a novel fine-tuned detector that can detect MGT from human-written texts by contrastive learning on selective perturbation. |
| Outcome: | The proposed method outperforms the state-of-the-art by 1.20% on four public datasets. |
HACo-Det: A Study Towards Fine-Grained Machine-Generated Text Detection under Human-AI Coauthoring (2025.acl-long)
Copied to clipboard
| Challenge: | Existing literature focuses on binary, document-level detection, neglecting texts composed jointly by human and LLM contributions. |
| Approach: | They propose to use a dataset to generate human-AI coauthored texts via an automatic pipeline with word-level attribution labels. |
| Outcome: | The proposed method can detect human-AI coauthored texts with a numeric AI ratio. |
Get Confused Cautiously: Textual Sequence Memorization Erasure with Selective Entropy Maximization (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods for erasure of memorized text fail to unlearn large numbers of memorizable samples without jeopardizing model utility. |
| Approach: | They propose a method that allows LLMs to memorize and recite some training sequences verbatim . they propose an entropy-based loss method that is shown to be more stable . |
| Outcome: | The proposed method improves model utility and accuracy while preserving model ability in language generation and understanding. |
StablePT : Towards Stable Prompting for Few-shot Learning via Input Separation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies on prompt tuning have shown that language models can be effective few-shot learners with prompting. |
| Approach: | They propose to treat the hard prompt and soft prompt as separate inputs to mitigate noise brought by prompt initialization. |
| Outcome: | Experimental results show that the proposed method outperforms state-of-the-art methods by 6.97% in accuracy and reduces the standard deviation by 1.92 on average. |