Papers by Yixin Wan
Where Fact Ends and Fairness Begins: Redefining AI Bias Evaluation through Cognitive Biases (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks conflate factual correctness and normative fairness . a model may generate responses that are factually accurate but socially unfair . |
| Approach: | They propose a benchmark to examine the boundary between fact and fair . they draw on representativeness bias, attribution bias and ingroup–outgroup bias to explain why models often misalign fact and faireness. |
| Outcome: | The proposed model is based on ten frontier models and is available on github . it is compared with a standard model that generates people of color in Nazi-era uniforms . |
V-ALPHASOCIAL: Benchmark and Self-Reflective Chain-of-Thought Generation for Visual Social Commonsense Reasoning (2025.findings-acl)
Copied to clipboard
Zongyu Lin, Zhikun Xu, Xiaohan Song, Yixin Wan, Xingcheng Yao, Tsung-Han Lin, Selina Song, Pranav Subbaraman, Ben Zhou, Kai-Wei Chang, Yizhou Sun
| Challenge: | Social commonsense reasoning is a multimodal task that requires both textual and visual cues. |
| Approach: | They propose a method that integrates visual cues into social commonsense reasoning tasks. |
| Outcome: | The proposed method improves social commonsense reasoning on a multimodal foundation model. |
LUME: LLM Unlearning with Multitask Evaluations (2025.findings-emnlp)
Copied to clipboard
Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, Rahul Gupta
| Challenge: | Unlearning aims to remove copyrighted, sensitive, or private content from large language models without a full retraining. |
| Approach: | They propose a multi-task unlearning benchmark LUME that unlearns short novels, biographies and public biographie . |
| Outcome: | The proposed benchmark unlearns short novels, biographies and public biographie . it also releases fine-tuned models with 1B and 7B parameter sizes as targets . |
MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing contextual safety benchmarks are mostly single-turn and miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. |
| Approach: | They propose a benchmark that evaluates contextual safety in multimodal large language models . they observe persistent trade-offs between contextual safety and utility . |
| Outcome: | The proposed model combines multi-turn and multi-switch scenarios to evaluate safety in multimodal large language models. |
TheoremQA: A Theorem-driven Question Answering Dataset (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent LLMs like GPT-4 and PaLM-2 have made tremendous progress in solving fundamental math problems like GSM8K by achieving over 90% accuracy. |
| Approach: | They propose to use theorem-driven question-answering dataset to evaluate AI models' ability to apply theoretic concepts to solving challenging science problems. |
| Outcome: | TheoremQA is curated by domain experts and contains 800 high-quality questions covering 350 theoremics from Math, Physics, EE&CS, and Finance. |
Benchmarking Large Language Models Under Data Contamination: A Survey from Static to Dynamic Evaluation (2025.emnlp-main)
Copied to clipboard
Simin Chen, Yiming Chen, Zexin Li, Yifan Jiang, Zhongwei Wan, Yixin He, Dezhi Ran, Tianle Gu, Haizhou Li, Tao Xie, Baishakhi Ray
| Challenge: | In the era of evaluating large language models, data contamination is an increasingly prominent concern . static benchmarking has been used for evaluation, but there are limitations of *dynamic* benchmarks . |
| Approach: | They propose a series of optimal design principles for *dynamic* benchmarking and analyze the limitations of existing *static* benchmarks. |
| Outcome: | The proposed benchmarks highlight a critical gap in the evaluation of LLMs. |
InsideOut: Measuring and Mitigating Insider–Outsider Bias in Interview Script Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Recent research has raised concerns about culture-related fairness issues in LLM-generated content. |
| Approach: | They propose to use 4,000 generation prompts and three evaluation metrics to quantify LLMs' **insider-outsider bias** . |
| Outcome: | The proposed method reduces bias in Llama model by 89.70% and mitigates bias on Qwen by 82.54% on cultural alignment gap metric. |
PIP: Parse-Instructed Prefix for Syntactically Controlled Paraphrase Generation (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing fine-tuning methods for this task are costly and require updating the parameters of the entire model to adapt to the newly included syntax information. |
| Approach: | They propose a method to instruct model’s encoder prefix to capture syntax-related knowledge by direct initiation and indirect optimization. |
| Outcome: | The proposed methods are 10 times more efficient and learnable than existing methods. |
“Kelly is a Warm Person, Joseph is a Role Model”: Gender Biases in LLM-Generated Reference Letters (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are an effective tool to assist individuals in writing documents. |
| Approach: | They examine gender biases in large language models (LLMs)-generated reference letters . they find that models are biased because they are hallucinated . |
| Outcome: | The proposed model-generated reference letters are evaluated on 2 popular LLMs- ChatGPT and Alpaca. |
MACAROON: Training Vision-Language Models To Be Your Engaged Partners (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large vision-language models (LVLMs) generate detailed responses even when questions are ambiguous or unanswerable, leading to hallucinations and bias issues. |
| Approach: | They propose a three-tiered hierarchy for questions of invalid, ambiguous, and personalizable nature to measure the proactive engagement capabilities of LVLMs. |
| Outcome: | The proposed model generates contrastive response pairs for unlabeled questions, achieving 0.84 AAR, while maintaining comparable performance on general tasks. |
Not Every Token Needs Forgetting: Selective Unlearning Balancing Forgetting and Utility in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Conventional unlearning approaches forget all tokens in a target document, including common tokens that carry general knowledge. |
| Approach: | They propose a method that identifies a critical subset of tokens within the forgetting set that is relevant to the unwanted information and unlearns only those tokens. |
| Outcome: | Experiments on two benchmarks and six baseline unlearning algorithms show that selective unlearning achieves effective unlearning on the targeted forget data. |
Re-evaluating Automatic LLM System Ranking for Alignment with Human Preference (2025.findings-naacl)
Copied to clipboard
| Challenge: | Evaluating and ranking the capabilities of different LLMs is crucial for understanding their performance and alignment with human preferences. |
| Approach: | They propose a system-level evaluation framework that ranks LLMs based on their alignment with human preferences. |
| Outcome: | The proposed framework aims to rank LLMs based on their performance and alignment with human preferences. |
The Factuality Tax of Diversity-Intervened Text-to-Image Generation: Benchmark and Fact-Augmented Intervention (2024.emnlp-main)
Copied to clipboard
| Challenge: | Prompt-based “diversity interventions” are commonly adopted to improve the diversity of Text-to-Image models depicting individuals with diverse racial or gender traits. |
| Approach: | They propose a benchmark to quantify the trade-off between using diversity interventions and preserving demographic factuality in T2I models. |
| Outcome: | The proposed model significantly improves the demographic factuality under diversity interventions while preserving diversity. |
The Male CEO and the Female Assistant: Evaluation and Mitigation of Gender Biases in Text-To-Image Generation of Dual Subjects (2025.acl-long)
Copied to clipboard
| Challenge: | Recent large-scale T2I models like DALLE-3 have made progress in reducing gender stereotypes when generating single-person images. |
| Approach: | They propose a framework that queries T2I models to depict two individuals with gender-stereotyped social identities to evaluate gender biases. |
| Outcome: | The proposed framework reduces gender stereotypes when generating images with more than one person. |
LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected? (2024.findings-naacl)
Copied to clipboard
Qihui Zhang, Chujie Gao, Dongping Chen, Yue Huang, Yixin Huang, Zhenyang Sun, Shilin Zhang, Weiye Li, Zhengyan Fu, Yao Wan, Lichao Sun
| Challenge: | Current research focuses on purely MGT detection without adequately addressing mixed scenarios including AI-revised Human-Written Text (HWT) and human-revealed MGT. |
| Approach: | They define mixtext, a form of mixed text involving both AI and human-generated content, and then use a MixSet dataset to assess their effectiveness. |
| Outcome: | The proposed detectors struggle to identify mixtext, particularly in dealing with subtle modifications and style adaptability. |
Improving the Adversarial Robustness of NLP Models by Information Bottleneck (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have shown that adversarial examples can be directly attributed to the presence of non-robust features. |
| Approach: | They propose to capture task-specific robust features while eliminating non-robust ones . they show that models can achieve significant improvement in robust accuracy . |
| Outcome: | The proposed method outperforms all defense methods on SST-2, AGNEWS and IMDB datasets while achieving no performance drop. |
Are Personalized Stochastic Parrots More Dangerous? Evaluating Persona Biases in Dialogue Systems (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models enable them to follow freeform instructions, including imitating generic or specific demographic personas in conversations. |
| Approach: | They propose to investigate persona biases by experimenting with UNIVERSALPERSONA, a model that incorporates both generic and specific personas. |
| Outcome: | The proposed model systematically measures persona biases in harmful expression and harmful agreement. |
VisRet: Visualization Improves Knowledge-Intensive Text-to-Image Retrieval (2026.acl-long)
Copied to clipboard
| Challenge: | Text-to-image retrieval is challenging because of cross-modal embeddings are bags of concepts, underrepresenting structured visual relationships. |
| Approach: | They propose a retrieval paradigm that embeds textual queries into the image modality via T2I generation and performs retrieval within the image mode to bypass weaknesses of cross-modal retrievers in recognizing subtle visual-spatial features. |
| Outcome: | The proposed retrieval paradigm outperforms previous approaches in visual-spatial retrieval benchmarks. |
Knowledge Control for Responsible Generative AI: Bridging Academia, Industry, and Society (2026.acl-tutorials)
Copied to clipboard
Zheyuan Liu, Yixin Wan, Kai-Wei Chang, Meng Jiang, Jieyu Zhao, Nouha Dziri, Yuning Mao, Jia-Chen Gu, Jindong Gu
| Challenge: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods. |
| Approach: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods. |
| Outcome: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods . key motivations and failure modes, harmful generation and stereotype reinforcement, are addressed . core methods such as machine unlearning, knowledge editing, and inference-time interventions are also included . |
White Men Lead, Black Women Help? Benchmarking and Mitigating Language Agency Social Biases in LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Social biases manifest in language agency, but there is no comprehensive benchmark for evaluating such biase in language models. |
| Approach: | They propose a benchmark to evaluate language agency biases in large language models . they propose 'Mitigation via Selective Rewrite' to selectively revise parts of generated texts . |
| Outcome: | The proposed language agency bias evaluation benchmark identifies gender, racial, and intersectional biases in 3 recent LLMs. |