Papers by Jieyu Zhao
What’s Missing in Vision-Language Models? Probing Their Struggles with Causal Order Reasoning (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing benchmarks often include a mixture of reasoning questions, making it difficult to truly assess VLMs’ causal reasoning abilities. |
| Approach: | They propose two new benchmarks specifically designed to isolate and rigorously evaluate VLMs’ causal reasoning abilities. |
| Outcome: | The proposed benchmarks show that vision-language models perform poorly on causal reasoning tasks, often only marginally surpassing random guessing. |
Mitigating Gender Bias Amplification in Distribution by Posterior Regularization (2020.acl-main)
Copied to clipboard
| Challenge: | Recent studies show that data-driven machine learning models carry societal biases in the dataset they trained on. |
| Approach: | They propose to calibrate top predictions of a model by injecting corpus-level constraints to ensure that the gender disparity is not amplified. |
| Outcome: | The proposed method can almost remove bias amplification in the distribution with little loss of performance. |
Can LLMs Express Personality Across Cultures? Introducing CulturalPersonas for Evaluating Trait Alignment (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have explored personality evaluation of LLMs, but they largely overlook the interplay between culture and personality. |
| Approach: | They propose a large-scale benchmark for evaluating LLMs’ personality expression in culturally grounded, behaviorally rich contexts. |
| Outcome: | The proposed benchmark improves alignment with country-specific human personality distributions and elicits more expressive, culturally coherent outputs compared to existing benchmarks. |
Mitigating Gender Bias in Natural Language Processing: Literature Review (P19-1)
Copied to clipboard
Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, William Yang Wang
| Challenge: | NLP models propagate and may even amplify gender bias found in text corpora . methods to mitigate gender bias in NLP are relatively nascent . |
| Approach: | They propose to analyze gender bias based on four forms of representation bias and discuss the advantages and drawbacks of existing gender debiasing methods. |
| Outcome: | The proposed methods are based on four forms of representation bias and have advantages and drawbacks. |
Gender Bias in Contextualized Word Embeddings (N19-1)
Copied to clipboard
| Challenge: | Existing studies show that training word embeddings in large corpora could lead to encoding societal biases present in these human-produced data. |
| Approach: | They conduct several intrinsic analyses to quantify, analyze and mitigate gender bias exhibited in ELMo’s contextualized word vectors. |
| Outcome: | The proposed method mitigates gender bias on WinoBias probing corpus and demonstrates that it can be implemented in other systems. |
Do Not Design, Learn: A Trainable Scoring Function for Uncertainty Estimation in Generative LLMs (2025.findings-naacl)
Copied to clipboard
Duygu Nur Yaldiz, Yavuz Faruk Bakman, Baturalp Buyukates, Chenyang Tao, Anil Ramakrishna, Dimitrios Dimitriadis, Jieyu Zhao, Salman Avestimehr
| Challenge: | Existing methods for probability-based UE are limited by their inability to handle biased probabilities and complex semantic dependencies between tokens. |
| Approach: | They propose a learning-based scoring function that captures complex dependencies between tokens and probabilities and produces more reliable responses. |
| Outcome: | The proposed function outperforms existing scoring functions in question-answering and arithmetical reasoning tasks with different datasets. |
Towards Understanding Gender Bias in Relation Extraction (2020.acl-main)
Copied to clipboard
Andrew Gaut, Tony Sun, Shirlyn Tang, Yuxin Huang, Jing Qian, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, William Yang Wang
| Challenge: | Existing bias mitigation techniques have a negative effect on NRE, a study finds . |
| Approach: | They create a dataset to analyze gender bias in relation extraction systems . they find that existing bias mitigation techniques have a negative effect on NRE . |
| Outcome: | The proposed dataset analyzes gender bias in relation extraction systems using a 10% human annotated test set. |
MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval (2026.acl-demo)
Copied to clipboard
| Challenge: | MED-COPILOT is an interactive research prototype for evidence-aware clinical reasoning . large language models (LLMs) are prone to hallucinations and lack verifiable evidence grounding . |
| Approach: | They propose a system that integrates GraphRAG and semantic-keyword similar-patient retrieval to support transparent clinical reasoning. |
| Outcome: | The proposed system outperforms baseline and standard RAGs on clinical note completion and medical question answering. |
Learning Gender-Neutral Word Embeddings (D18-1)
Copied to clipboard
| Challenge: | Word embeddings trained on human-generated corpora inherit strong gender stereotypes . prior studies show such embeddables exhibit social biases, such as gender stereotype . |
| Approach: | They propose a method to preserve gender information in certain dimensions of word vectors . they propose GN-GloVe, which is a gender-neutral variant of the word embedding model . |
| Outcome: | The proposed method preserves gender information in certain dimensions of word vectors while compelling other dimensions to be free of gender influence. |
Self-contradictory reasoning evaluation and detection (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown impressive reasoning ability, but many downstream reasoning tasks focus on performance-wise evaluation. |
| Approach: | They define and assess the Self-Contra rate across three datasets and delve into finer-grained categories of Self-contra reasoning. |
| Outcome: | The proposed model can detect self-contra reasoning with a 52.2% F1 score, much lower than for humans. |
InterIntent: Investigating Social Intelligence of LLMs via Intention Understanding in an Interactive Game Context (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated the potential to mimic human social intelligence, but most studies focus on static self-report or performance-based tests. |
| Approach: | They propose a framework to assess LLMs' ability to understand and manage intentions by mapping their ability to infer the intentions of others in a game setting. |
| Outcome: | The proposed framework assesses LLMs' ability to understand and manage intentions in a game setting. |
On Measures of Biases and Harms in NLP (2022.findings-aacl)
Copied to clipboard
Sunipa Dev, Emily Sheng, Jieyu Zhao, Aubrie Amstutz, Jiao Sun, Yu Hou, Mattie Sanseverino, Jiin Kim, Akihiro Nishi, Nanyun Peng, Kai-Wei Chang
| Challenge: | Recent studies show that natural language processing (NLP) technologies propagate societal biases about demographic groups associated with attributes such as gender, race, and nationality. |
| Approach: | They propose a framework for harms and questions to help practitioners understand biases . they propose measurable measures to detect and mitigate biased groups . |
| Outcome: | The proposed framework provides a framework for harms and questions for practitioners to answer to guide the development of bias measures. |
Fair Abstractive Summarization of Diverse Perspectives (2024.naacl-long)
Copied to clipboard
Yusen Zhang, Nan Zhang, Yixin Liu, Alexander Fabbri, Junru Liu, Ryo Kamoi, Xiaoxin Lu, Caiming Xiong, Jieyu Zhao, Dragomir Radev, Kathleen McKeown, Rui Zhang
| Challenge: | Existing work on summarization metrics and large language models has not explored fair abstractive summarizing. |
| Approach: | They propose four reference-free automatic metrics to measure the differences between target and source perspectives. |
| Outcome: | The proposed methods alleviate fair abstractive summarization on user-generated data. |
ProMediate: A Simulation Testbed for Evaluating Proactive Mediation in Multi-Party Negotiation (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly integrated into agentic frameworks to assist individual users in completing diverse tasks. |
| Approach: | They propose a simulation environment with a plug-and-play proactive AI mediator . they use a socio-cognitive evaluation framework to measure consensus changes, intervention latency, mediator effectiveness and intelligence. |
| Outcome: | The proposed model outperforms a generic baseline in multi-party negotiation scenarios while being 77% faster in response. |
AI Sees Your Location—But With A Bias Toward The Wealthy World (2025.emnlp-main)
Copied to clipboard
| Challenge: | Visual Language Models (VLMs) have shown remarkable performance across various tasks, particularly in recognizing geographic information from images. |
| Approach: | They propose to use 1,200 images paired with detailed geographic metadata to evaluate VLMs' performance. |
| Outcome: | The models achieve 53.8% accuracy in city prediction, but exhibit significant biases in regional tasks. |
Images Speak Louder than Words: Understanding and Mitigating Bias in Vision-Language Model from a Causal Mediation Perspective (2024.emnlp-main)
Copied to clipboard
| Challenge: | Current methods to learn biases from the perspective of model components are limited by their complexity and performance. |
| Approach: | They propose a framework that incorporates causal mediation analysis to measure and map the pathways of bias generation and propagation within vision-language and multimodal tasks. |
| Outcome: | The proposed framework is applicable to a wide range of vision-language and multimodal tasks and reduces bias by 22.03% and 9.04% in the MSCOCO and PASCAL-SENTENCE datasets. |
Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods (N18-2)
Copied to clipboard
| Challenge: | Existing methods for co-reference resolution focus on gender bias. |
| Approach: | They propose a new benchmark for co-reference resolution focused on gender bias, WinoBias. |
| Outcome: | The proposed system removes the bias without significantly affecting performance on existing datasets. |
Investigating Ensemble Methods for Model Robustness Improvement of Text Classifiers (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to reduce model's reliance on bias features ignore the learnability of these features. |
| Approach: | They propose to reduce models' reliance on bias features by first training models with fixed low-capacity models which ignore the learnability of the bias features. |
| Outcome: | The proposed models can perform better on out-of-distribution datasets than baseline models with a more sophisticated model design. |
“You Gotta be a Doctor, Lin” : An Investigation of Name-Based Bias of Large Language Models in Employment Recommendations (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated racial and gender biases in various applications. |
| Approach: | They use Large Language Models to simulate hiring decisions and salary recommendations for candidates with 320 first names that strongly signal their race and gender, across over 750,000 prompts. |
| Outcome: | The proposed models favor candidates with White female-sounding names over other demographic groups across 40 occupations. |
Gender Bias in Multilingual Embeddings and Cross-Lingual Transfer (2020.acl-main)
Copied to clipboard
| Challenge: | Multilingual word embeddings embed words from many languages into a single semantic space such that words with similar meanings are close to each other regardless of the language. |
| Approach: | They propose to use multilingual word embeddings to align embeddable words from multiple languages into a single semantic space so that words with similar meanings are close to each other regardless of the language. |
| Outcome: | The proposed model can be used to learn gender bias in multilingual representations and to improve transfer learning. |
Double Perturbation: On the Robustness of Robustness and Counterfactual Bias Evaluation (2021.naacl-main)
Copied to clipboard
| Challenge: | Recent studies show that NLP models are vulnerable to adversarial perturbations such as synonym substitutions or syntax-guided paraphrasing. |
| Approach: | They propose a “double perturbation” framework to uncover model weaknesses beyond the test dataset. |
| Outcome: | The proposed attack achieves high success rates on both original and robustly trained CNNs and Transformers. |
LOGAN: Local Group Bias Detection by Clustering (2020.emnlp-main)
Copied to clipboard
| Challenge: | a number of machine learning models inherit and amplify the societal biases in data. |
| Approach: | a new bias detection technique based on clustering is proposed to detect local biases in data . authors propose to use LOGAN to analyze local bias in data. |
| Outcome: | The proposed technique detects bias in a local region and allows better analysis of model predictions. |
Examining Gender Bias in Languages with Grammatical Gender (D19-1)
Copied to clipboard
| Challenge: | Existing studies on gender bias in word embeddings focus on English . however, these studies cannot be extended to languages with morphological agreement on gender . |
| Approach: | They propose new metrics to evaluate gender bias in word embeddings of English and Spanish . they extend existing approaches to mitigate gender bias while preserving original embeddables . |
| Outcome: | The proposed methods reduce gender bias while preserving the original embeddings. |
Safer-Instruct: Aligning Language Models with Automated Preference Data (2024.naacl-long)
Copied to clipboard
| Challenge: | annotating preference data by humans is resource-intensive and creativity-demanding . existing methods face limitations in data diversity and quality . |
| Approach: | They propose a pipeline for annotating large-scale preference data without human annotators. |
| Outcome: | The proposed pipeline outperforms models fine-tuned on human-annotated safety preference data while maintaining a competitive edge in downstream tasks. |
Cross-Lingual Pitfalls: Automatic Probing Cross-Lingual Weakness of Multilingual Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models have achieved remarkable success in Natural Language Processing, yet their cross-lingual consistency remains a significant challenge. |
| Approach: | They propose a method to identify cross-lingual weaknesses in Large Language Models . they construct bilingual question pairs that expose performance discrepancies between English and target languages . |
| Outcome: | The proposed method uncovers over 50% accuracy drops in target languages across models. |
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. |
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. |
Ethical-Advice Taker: Do Language Models Understand Natural Language Interventions? (2021.findings-acl)
Copied to clipboard
| Challenge: | a new task is proposed to modify a model's unethical behavior by communicating context-specific principles of ethics and equity to it. |
| Approach: | They propose a task to amend a questionanswering model's unethical behavior . they use social stereotypes to augment social stereotype models with ethical interventions . |
| Outcome: | The proposed task improves model behavior but fails to generalize to new domains . it is based on a zero-shot evaluation of social stereotypes in reading comprehension systems . |
“The Boating Store Had Its Best Sail Ever”: Pronunciation-attentive Contextualized Pun Recognition (2020.acl-main)
Copied to clipboard
| Challenge: | Identifying and modeling puns is challenging as they involve implicit semantic or phonological tricks. |
| Approach: | They propose a method to detect puns in a sentence and then locate them in it . they propose to capture phonetic associations between the context and phonetic symbols . |
| Outcome: | The proposed method outperforms state-of-the-art methods in pun detection and location tasks. |
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 . |
GRAVITY: A Framework for Personalized Text Generation via Profile-Grounded Synthetic Preferences (2026.eacl-long)
Copied to clipboard
| Challenge: | Personalization in LLMs often relies on costly human feedback or interaction logs, limiting scalability and neglecting deeper user attributes. |
| Approach: | They propose a framework for generating synthetic, profile-grounded preference data that captures users’ interests, values, beliefs, and personality traits. |
| Outcome: | The proposed framework improves on book descriptions for 400 Amazon users across multiple cultures, with user studies showing that outputs are preferred over 86% of the time. |
A Rose by Any Other Name would not Smell as Sweet: Social Bias in Names Mistranslation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Using demographics, we hypothesize that the ability of translation systems to correctly translate female-associated names is significantly lower than male-associated name. |
| Approach: | They propose a translation evaluation procedure based on round-trip translation of names that are demographically aligned and analyze the effect of name demographics on translation quality using generalized linear mixed effects models. |
| Outcome: | The proposed evaluation procedure is based on round-trip translation of names from a dataset of names that are demographically aligned and shows that the ability of translation systems to translate female-associated names is significantly lower than male-associated name. |
Analyzing Uncertainty of LLM-as-a-Judge: Interval Evaluations with Conformal Prediction (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are powerful automatic evaluators for natural language generation (NLG) tasks, but their uncertainty may limit their deployment in many applications. |
| Approach: | They propose a conformal prediction framework that provides a prediction interval with coverage guarantees and a midpoint-based score as a low-bias alternative to raw model score and weighted average. |
| Outcome: | The proposed framework provides a prediction interval with coverage guarantees and a midpoint-based score as a low-bias alternative to raw model score and weighted average. |
SODAPOP: Open-Ended Discovery of Social Biases in Social Commonsense Reasoning Models (2023.eacl-main)
Copied to clipboard
| Challenge: | Existing diagnostic tests for detecting social biases in NLP models only detect stereotypic associations pre-specified by the designer. |
| Approach: | They propose an approach for automatic social bias discovery in social commonsense question-answering by substituting names associated with different demographic groups and generating many distractor answers from a masked language model. |
| Outcome: | The proposed approach uncovers model’s stereotypic associations between demographic groups and an open set of words. |
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback (2026.acl-long)
Copied to clipboard
Taiwei Shi, Zhuoer Wang, Longqi Yang, Ying-Chun Lin, Zexue He, Mengting Wan, Pei Zhou, Sujay Kumar Jauhar, Sihao Chen, Shan Xia, Hongfei Zhang, Jieyu Zhao, Xiaofeng Xu, Xia Song, Jennifer Neville
| Challenge: | Traditional alignment methods rely on human annotations and are subjective and misalignment with real-world user preferences. |
| Approach: | They propose a framework that leverages in-situ user feedback during conversations with LLMs to create preference datasets automatically. |
| Outcome: | The proposed framework identifies and classifies user feedback to LLM responses between conversation turns and creates examples of preferred and dispreferred responses according to user preferences. |
The Hallucination Tax of Reinforcement Finetuning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose to incorporate 10% SUM into RFT to restore appropriate refusal behavior with minimal accuracy trade-offs on solvable tasks. |
| Outcome: | The proposed approach reduces model refusal rates by more than 80%, which significantly increases model’s tendency to hallucinate. |