Papers by Zhengyu Zhao
An Optimizable Suffix Is Worth A Thousand Templates: Efficient Black-box Jailbreaking without Affirmative Phrases via LLM as Optimizer (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing jailbreaking methods generate harmful and unethical content when subjected to jailbreaking attacks. |
| Approach: | They propose a black-box jailbreaking method with optimizable suffixes that translate jailbreaking objectives into natural language instructions. |
| Outcome: | The proposed method outperforms existing methods by 2.4 times in the ASR of three open-source LLMs and GPT-3.5-Turbo. |
Typographic Attacks in a Multi-Image Setting (2025.naacl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models are susceptible to typographic attacks, which are misclassifications caused by an attack text that is added to an image. |
| Approach: | They propose a multi-image setting for studying typographic attacks by leveraging the difficulty of the target image, the strength of the attack text, and text-image similarity. |
| Outcome: | The proposed approach improves success rates by 21% over random, non-specific methods on the CLIP model while maintaining stealth in a multi-image scenario. |
Deep Reinforcement Learning-based Dialogue Policy with Graph Convolutional Q-network (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for deep reinforcement learning lack the ability to learn the relationship between dialogue states and actions. |
| Approach: | They propose a graph-structured dialogue policy framework for task-oriented dialogue systems that uses bipartite graphs to construct two different bipartites and generate user-related and knowledge-related subgraphs. |
| Outcome: | The proposed framework significantly improves the effectiveness and stability of dialogue policies. |
Composite Backdoor Attacks Against Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated superior performance on various tasks, but untrustworthy third-party LLMs may covertly introduce vulnerabilities for downstream tasks. |
| Approach: | They propose a composite backdoor attack that scatters multiple trigger keys in different prompt components. |
| Outcome: | The proposed attack achieves 100% Attack Success Rate (ASR) with a False Triggered Rate (FTR) below 2.06% and negligible model accuracy degradation. |
What Did You Refer to? Evaluating Co-References in Dialogue (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing neural end-to-end dialogue models have limitations on exactly interpreting the linguistic structures in dialogue history context. |
| Approach: | They propose to directly measure the capability of neural end-to-end dialogue models on understanding the entity-oriented structures via question answering. |
| Outcome: | The proposed model can understand large-scale English and Chinese human human dialogues using a large-format dataset. |
Profile Consistency Identification for Open-domain Dialogue Agents (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on improving attribute consistency focus on incorporating attribute information in responses, but few efforts have identified the consistency relations between response and attribute profile. |
| Approach: | They propose a key-value structure information enriched BERT model to identify the profile consistency . they propose to incorporate attribute information into the generated responses . |
| Outcome: | The proposed model improves over strong baselines on downstream tasks. |
CoAct: Co-Active LLM Preference Learning with Human-AI Synergy (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to learn from preference-based feedback are expensive and scarce. |
| Approach: | They propose a framework that synergistically combines self-rewarding and active learning through human-AI collaboration. |
| Outcome: | The proposed framework outperforms existing methods on three reasoning benchmarks and achieves average improvements of +13.25% on GSM8K, +8.19% on MATH, and +13.16% on WebInstruct. |