Papers by Yuji Zhang
SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing safety mechanisms for large language models (LLMs) are inadequate to fully leverage their internal cognitive processes. |
| Approach: | They propose a framework that regulates unsafe outputs by utilizing the prober-based internal state monitor that actively detects harmful intentions. |
| Outcome: | The proposed framework reduces harmful outputs by approximately 80% while maintaining strong utility. |
#HowYouTagTweets: Learning User Hashtagging Preferences via Personalized Topic Attention (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods based on latent topics cannot capture user interests and thus can't be used to predict how likely a user will post with a hashtag. |
| Approach: | They propose a personalized topic attention model that captures salient contents to personalize hashtag contexts by predicting how likely a user will post with a hashtag. |
| Outcome: | The proposed model significantly outperforms the state-of-the-art recommendation approach without exploiting latent topics. |
EVEDIT: Event-based Knowledge Editing for Deterministic Knowledge Propagation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing knowledge editing approaches only operate on (subject, relation, object) triple . current methods are limited to (substance, relation) triple, causing low confidence in their answers. |
| Approach: | They propose a task of event-based knowledge editing that pairs facts with event descriptions to improve model confidence. |
| Outcome: | The proposed method improves model confidence by 55.6% while maintaining the naturalness of generation. |
EMCompress: Video-LLMs with Endomorphic Multimodal Compression (2026.findings-acl)
Copied to clipboard
| Challenge: | Static, sparse frame sampling either dilutes evidence across task-irrelevant segments at significant cost or misses fine-grained temporal semantics altogether. |
| Approach: | They propose a novel task that compresses multimodal input while preserving answer invariance across reasonable downstream models. |
| Outcome: | The proposed task surpasses prior methods by 0.40 F-1 with competitive query rewriting. |
EscapeBench: Towards Advancing Creative Intelligence of Language Model Agents (2025.acl-long)
Copied to clipboard
Cheng Qian, Peixuan Han, Qinyu Luo, Bingxiang He, Xiusi Chen, Yuji Zhang, Hongyi Du, Jiarui Yao, Xiaocheng Yang, Denghui Zhang, Yunzhu Li, Heng Ji
| Challenge: | Existing language model agents excel in planning and reasoning, but lack creativity in unfamiliar environments. |
| Approach: | They propose a benchmark suite of room escape game environments to challenge agents with creative reasoning, unconventional tool use and iterative problem-solving to uncover implicit goals. |
| Outcome: | The proposed framework can perform with 40% fewer steps and hints and performs robustly across difficulty levels. |
Current Agents Fail to Leverage World Model as Tool for Foresight (2026.acl-long)
Copied to clipboard
Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Gokhan Tur, Dilek Hakkani-Tür, Yunzhu Li, Heng Ji
| Challenge: | Generative world models could be used to enhance agents' cognition . agents are expected to operate in settings where tasks unfold over long horizons and involve intricate chains of interdependent decisions. |
| Approach: | They propose to use vision-language models as external simulators to enhance cognition . they find that agents rarely invoke simulation and misuse predicted rollouts . |
| Outcome: | The proposed model could be used to predict future states rather than short-horizon reasoning . the model could also be used for real-world planning and robotics . |
VIBE: Topic-Driven Temporal Adaptation for Twitter Classification (2023.emnlp-main)
Copied to clipboard
| Challenge: | Language features are evolving in real-world social media, resulting in deteriorating performance of text classification. |
| Approach: | They propose a model that allows models to adapt to shifted data via latent topic evolution . they use two information bottleneck regularizers to distinguish past and future topics . |
| Outcome: | The proposed model outperforms state-of-the-art models on Twitter on three tasks with 3% of data. |
ModelingAgent: Bridging LLMs and Mathematical Modeling for Real-World Challenges (2025.findings-emnlp)
Copied to clipboard
Cheng Qian, Hongyi Du, Hongru Wang, Xiusi Chen, Yuji Zhang, Avirup Sil, ChengXiang Zhai, Kathleen McKeown, Heng Ji
| Challenge: | Existing benchmarks for large language models fail to reflect real-world complexity . existing benchmarks often fail to capture real-life problems . |
| Approach: | They propose a benchmark that features real-world-inspired, open-ended problems from competitions . they propose 'ModelingBench' that supports multiple valid solutions . |
| Outcome: | The proposed framework outperforms baselines and produces well-grounded, creative solutions. |
Hashtags, Emotions, and Comments: A Large-Scale Dataset to Understand Fine-Grained Social Emotions to Online Topics (2020.emnlp-main)
Copied to clipboard
| Challenge: | A large-scale dataset is collected from Chinese microblog Sina Weibo with over 13 thousand trending topics, emotion votes in 24 fine-grained types from massive participants, and user comments to allow context understanding. |
| Approach: | They use a large-scale dataset from Chinese microblog Sina Weibo to examine readers' responses to online discussion topics. |
| Outcome: | The proposed model outperforms the human model in predicting social emotions in a multilabel classification setting. |
The Law of Knowledge Overshadowing: Towards Understanding, Predicting and Preventing LLM Hallucination (2025.findings-acl)
Copied to clipboard
Yuji Zhang, Sha Li, Cheng Qian, Jiateng Liu, Pengfei Yu, Chi Han, Yi R. Fung, Kathleen McKeown, ChengXiang Zhai, Manling Li, Heng Ji
| Challenge: | Hallucination is a persistent challenge in large language models where even with rigorous quality control, models often generate distorted facts. |
| Approach: | They propose a new framework to quantify factual hallucinations by modeling knowledge overshadowing. |
| Outcome: | The proposed framework improves model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). |
Engage the Public: Poll Question Generation for Social Media Posts (2021.acl-long)
Copied to clipboard
| Challenge: | a novel application to generate poll questions for social media posts offers an easy way to hear the public's voice . for the silent majority, they tend to read others' messages instead of voicing their opinions with words . |
| Approach: | They propose to encode user comments and discover latent topics therein as contexts to generate poll questions for social media posts. |
| Outcome: | The proposed model outperforms popular models without exploiting topics from comments . human evaluations show it can generate high-quality polls useful to draw user engagements . |