Papers by Youliang Yuan
Insight Over Sight: Exploring the Vision-Knowledge Conflicts in Multimodal LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to mitigating vision-knowledge conflict in Large Language Models (MLLMs) are not effective and can be further scaled. |
| Approach: | They propose a framework to generate inputs to simulate and evaluate vision-knowledge conflict in Multimodal Large Language Models (MLLMs) using original images and 1,122 high-quality question-answer pairs, they propose 'a diagnostic benchmark' |
| Outcome: | The proposed framework, benchmark, and analysis contribute to the understanding and mitigation of vision-knowledge conflicts in Multimodal Large Language Models (MLLMs). |
ToolSafety: A Comprehensive Dataset for Enhancing Safety in LLM-Based Agent Tool Invocations (2025.emnlp-main)
Copied to clipboard
| Challenge: | Current models exhibit notable vulnerabilities in maintaining safety during multi-step tool interactions and in indirect harm scenarios. |
| Approach: | They propose a safety fine-tuning dataset to fine- tune LLMs into assistants . they propose to use synthesized trajectories and realistic, context-aware sample generation . |
| Outcome: | The proposed model maintains safety in multi-step and indirect harm scenarios with little impact on helpfulness. |
Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training (2025.acl-long)
Copied to clipboard
Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Jiahao Xu, Tian Liang, Pinjia He, Zhaopeng Tu
| Challenge: | Large Language Models exhibit a level of intelligence that is both impressive and everevolving, but their ability to refuse generating unsafe content is a double-edged sword. |
| Approach: | They propose a method to tackle a refusal position bias within safety tuning data that compromises the models’ ability to appropriately refuse generating unsafe content. |
| Outcome: | The proposed method significantly improves model safety without compromising performance and surpasses baseline methods in defending against attacks. |
Chain-of-Jailbreak Attack for Image Generation Models via Step by Step Editing (2025.findings-acl)
Copied to clipboard
Wenxuan Wang, Kuiyi Gao, Youliang Yuan, Jen-tse Huang, Qiuzhi Liu, Shuai Wang, Wenxiang Jiao, Zhaopeng Tu
| Challenge: | Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows . however, considerable efforts are being made to prevent the generation of harmful content, such abusive, violent, or pornographic material. |
| Approach: | They propose a chain-of-jailbreak method which decomposes malicious queries into multiple sub-queries and iteratively edits images based on these sub-questions. |
| Outcome: | The proposed method can bypass safeguards of image generation models for over 60% cases, significantly outperforms other jailbreaking methods (14%) |
Can’t See the Forest for the Trees: Benchmarking Multimodal Safety Awareness for Multimodal LLMs (2025.acl-long)
Copied to clipboard
Wenxuan Wang, Xiaoyuan Liu, Kuiyi Gao, Jen-tse Huang, Youliang Yuan, Pinjia He, Shuai Wang, Zhaopeng Tu
| Challenge: | Multimodal Large Language Models (MLLMs) have expanded the capabilities of traditional language models by enabling interaction through both text and images. |
| Approach: | They propose a multimodal safety awareness benchmark to evaluate MLLMs across 29 safety scenarios with 1,500 carefully curated image-prompt pairs. |
| Outcome: | The proposed model is able to identify unsafe content and avoid over-sensitivity that can hinder helpfulness. |
SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing educational LLMs are vulnerable to pedagogical jailbreaks where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions. |
| Approach: | They propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries and identifies mastery gaps. |
| Outcome: | The proposed method improves safety under two pedagogical jailbreak scenarios while maintaining near-ceiling helpfulness under the same evaluation protocol. |
LogicAsker: Evaluating and Improving the Logical Reasoning Ability of Large Language Models (2024.emnlp-main)
Copied to clipboard
Yuxuan Wan, Wenxuan Wang, Yiliu Yang, Youliang Yuan, Jen-tse Huang, Pinjia He, Wenxiang Jiao, Michael Lyu
| Challenge: | LogicAsker examines and improves the reasoning abilities of large language models such as ChatGPT and GPT-4. |
| Approach: | They propose a set of atomic reasoning skills grounded in propositional and predicate logic to examine and improve the reasoning abilities of large language models such as ChatGPT and GPT-4. |
| Outcome: | The proposed approach improves reasoning abilities in large language models such as ChatGPT and GPT-4 by up to 5%. |
Identifying the Achilles’ Heel: An Iterative Method for Uncovering Factual Errors in Large Language Models (2026.findings-acl)
Copied to clipboard
Wenxuan Wang, Yuk-Kit Chan, Zixuan Ling, Shi Juluan, Youliang Yuan, Jen-tse Huang, Yifei Zhang, Wenxiang Jiao, Zhaopeng Tu, Michael R. Lyu
| Challenge: | Current methods for evaluating LLMs’ veracity are limited by the need for extensive human labor, test data contamination, or limited scope, hindering efficient and effective exposure of errors. |
| Approach: | They propose a framework that extracts fact triplets to generate diverse question types using rule-based natural language processing techniques. |
| Outcome: | The proposed framework can trigger factual errors in up to 55% of questions in large LLMs while maintaining coverage of questions. |
Does ChatGPT Know That It Does Not Know? Evaluating the Black-Box Calibration of ChatGPT (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent performance of ChatGPT in downstream tasks is questionable, but does it know that it does not know? |
| Approach: | They propose to use three types of proxy confidence to evaluate ChatGPT's black-box calibration ability. |
| Outcome: | The proposed model exhibits a positive correlation with accuracy in TruthfulQA and a negative correlation in the ModAr dataset. |
Learning to Ask: When LLM Agents Meet Unclear Instruction (2025.emnlp-main)
Copied to clipboard
Wenxuan Wang, Shi Juluan, Zixuan Ling, Yuk-Kit Chan, Chaozheng Wang, Cheryl Lee, Youliang Yuan, Jen-tse Huang, Wenxiang Jiao, Michael R. Lyu
| Challenge: | Despite their impressive capabilities, LLMs struggle with complex computations and delivering accurate, timely information. |
| Approach: | They propose a framework that prompts LLM agents to ask questions when they encounter obstacles due to unclear instructions and an automated evaluation tool called ToolEvaluator. |
| Outcome: | The proposed framework outperforms existing frameworks for tool learning in the Noisy ToolBench. |
Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards (2026.acl-long)
Copied to clipboard
Youliang Yuan, Qiuyang Mang, Jingbang Chen, Hong Wan, Xiaoyuan Liu, Junjielong Xu, Jen-tse Huang, Wenxuan Wang, Wenxiang Jiao, Pinjia He
| Challenge: | Existing models are susceptible to reward hacking, leading to a substantial overestimation of a model's reasoning ability. |
| Approach: | They propose a Rubric Reward Model that rewards the entire reasoning trajectory against problem-specific rubrics. |
| Outcome: | The proposed model outperforms outcome-only supervision on four math benchmarks and boosts Verified Pass@1024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. |
Difficult Task Yes but Simple Task No: Unveiling the Laziness in Multimodal LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing research indicates that even state-of-the-art MLLMs still suffer from some straightforward visual question-answering (VQA) problems. |
| Approach: | They propose to use a model-based benchmark to investigate model laziness to identify models that err when answering simple visual questions about an image. |
| Outcome: | The proposed model laziness is found to be widespread in current MLLMs, including GPT-4o, Gemini-1.5-pro, Claude 3, LLaVA-1.5, LLva-1.6, and QWen-VL. |
VisBias: Measuring Explicit and Implicit Social Biases in Vision Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Identifying and addressing potential social biases is essential to prevent harm to users. |
| Approach: | They examine explicit and implicit biases exhibited by Vision-Language Models . they pose questions related to gender and racial differences to test their models . |
| Outcome: | The proposed models are used in image description tasks, form completion tasks and medical applications. |
All Languages Matter: On the Multilingual Safety of LLMs (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing safety benchmarks only concern the safety in one language, e.g. the majority language in the pretraining data such as English. |
| Approach: | They propose a prompting method to improve multilingual safety of ChatGPT by enhancing cross-lingual generalization of safety alignment. |
| Outcome: | The proposed method can significantly reduce the ratio of unsafe responses by 42% for non-English queries. |