Papers by Ziyu Yao
Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can solve reasoning and mathematical problems using the Chain-of-Thought technique, but require costly and long CoT data and fine-tuning. |
| Approach: | They propose a method that uses Sparse Autoencoders to extract interpretable features from vanilla CoT and use them to steer the LLM's internal states. |
| Outcome: | The proposed method uses Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT and steer the LLM's internal states during generation. |
From Outcome to Process: Optimizing MoE Load Balancing with MCTS (2026.findings-acl)
Copied to clipboard
Wenjun Ke, Hengyuan Xu, Ziyu Shang, Yao He, Jiahao Wang, Zijie Xu, Peng Wang, Yuhang Lou, Jiajun Liu
| Challenge: | Existing balancing strategies focus on constraining the final distribution of expert usage, but overlook the routing decisions made at each layer. |
| Approach: | They propose a three-stage framework that leverages process-level rewards to guide balanced expert routing. |
| Outcome: | Extensive experiments show that LayerMoE improves the performance of state-of-the-art LoRA-MoA baselines, yielding an average accuracy gain of 1.39%. |
Gentopia.AI: A Collaborative Platform for Tool-Augmented LLMs (2023.emnlp-demo)
Copied to clipboard
Binfeng Xu, Xukun Liu, Hua Shen, Zeyu Han, Yuhan Li, Murong Yue, Zhiyuan Peng, Yuchen Liu, Ziyu Yao, Dongkuan Xu
| Challenge: | Existing frameworks for Augmented Language Models lack flexibility, democratization, and holistic evaluation. |
| Approach: | They propose a lightweight and extensible framework for Augmented Language Models called Gentopia. |
| Outcome: | The proposed framework integrates language models, task formats, prompting modules, and plugins into a unified paradigm. |
Navigating the Shortcut Maze: A Comprehensive Analysis of Shortcut Learning in Text Classification by Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Language models (LMs) often rely on spurious correlations rather than causally relevant features to improve accuracy and generalizability. |
| Approach: | They propose a benchmark that categorizes shortcuts into occurrence, style, and concept . they aim to explore the nuanced ways shortcuts influence the performance of LMs . |
| Outcome: | The proposed benchmark categorizes shortcuts into occurrence, style, and concept . it systematically investigates models’ resilience and susceptibilities to sophisticated shortcuts . |
An Investigation of Neuron Activation as a Unified Lens to Explain Chain-of-Thought Eliciting Arithmetic Reasoning of LLMs (2024.acl-long)
Copied to clipboard
| Challenge: | Prior work has focused on ablating components in the CoT prompt, but the reason why these components are important to LLM reasoning is not explored. |
| Approach: | They investigate "neuron activation" as a lens to provide a unified explanation to previous work . they propose an approach to automatically identify neurons that imply arithmetic reasoning . |
| Outcome: | The proposed approach can explain the importance of components in a CoT prompt . it also automatically identifies neurons that imply arithmetic reasoning . |
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) can disentangle complex features into more interpretable components. |
| Approach: | They propose to use Sparse Autoencoders to disentangle LLM features into more interpretable components. |
| Outcome: | The proposed method disentangles complex features into more interpretable components. |
MailEx: Email Event and Argument Extraction (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing work on email event extraction only covers one specific aspect of email information and cannot connect with other relevant tasks. |
| Approach: | They propose a new taxonomy for performing event extraction from conversational email threads. |
| Outcome: | The proposed taxonomy covers 10 event types and 76 arguments in the email domain. |
Efficient but Vulnerable: Benchmarking and Defending LLM Batch Prompting Attack (2025.findings-acl)
Copied to clipboard
| Challenge: | a recent study reveals a significant security vulnerability in batch prompting . malicious users can inject attack instructions into a batch, leading to unwanted interference . |
| Approach: | They construct a batch prompting benchmark and test it against other LLMs to find out if batch prompts are vulnerable. |
| Outcome: | The proposed approach achieves 95% accuracy in detecting attacks. |
Why Do LLM-based Web Agents Fail? A Hierarchical Planning Perspective (2026.acl-long)
Copied to clipboard
| Challenge: | Existing evaluations focus primarily on end-to-end success, offering limited insight into where failures arise. |
| Approach: | They propose a hierarchical planning framework that analyzes web agents across three layers . they show that structured Planning Domain Definition Language (PDDL) plans produce more concise and goal-directed strategies than natural language (NL) plans . |
| Outcome: | The proposed framework analyzes web agents across three layers to improve reasoning, grounding, and recovery. |
All for One: LLMs Solve Mental Math at the Last Token With Information Transferred From Other Tokens (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) perform well on a multitude of computational tasks, yet their inner workings remain unclear. |
| Approach: | They propose two techniques to inhibit input-specific token computations in initial layers . they propose a transformer that allows for any token to immediately access all preceding tokens . |
| Outcome: | The proposed algorithms can perform on a variety of mental math tasks with high accuracy and transfer across models. |
MoVa: Towards Generalizable Classification of Human Morals and Values (2025.emnlp-main)
Copied to clipboard
Ziyu Chen, Junfei Sun, Chenxi Li, Tuan Dung Nguyen, Jing Yao, Xiaoyuan Yi, Xing Xie, Chenhao Tan, Lexing Xie
| Challenge: | Identifying human morals and values embedded in language is essential to empirical studies of communication. |
| Approach: | They propose a framework for generalizable classification of human morals and values . they recommend a classification strategy that scores all related concepts simultaneously . |
| Outcome: | The proposed method outperforms fine-tuned models across domains and frameworks. |
An Imitation Game for Learning Semantic Parsers from User Interaction (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for learning semantic parsers are expensive and tedious . despite the widespread applications, bootstrapping and fine-tuning is tedious a task . |
| Approach: | They propose an alternative method for learning semantic parsers directly from users . they propose an annotation-efficient imitation learning algorithm that iteratively collects new datasets . |
| Outcome: | The proposed method is cost-effective and shows promising performance on the text-to-SQL problem. |
Instances Need More Care: Rewriting Prompts for Instances with LLMs in the Loop Yields Better Zero-Shot Performance (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have revolutionized zero-shot task performance, mitigating the need for task-specific annotations while enhancing task generalizability. |
| Approach: | They propose an approach that optimizes the zero-shot prompts for individual task instances following an innovative manner of "LLMs in the loop" their results show that PRomPTed outperforms naive zero- shot approaches and a strong baseline which refines the task output instead of the input prompt. |
| Outcome: | The proposed approach outperforms naive approaches and a strong baseline which refines the task output instead of the input prompt. |
What’s Not Said Still Hurts: A Description-Based Evaluation Framework for Measuring Social Bias in LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks evaluate bias by term-based mode, but they fail to capture hidden biases in nuanced settings. |
| Approach: | They propose a dataset to assess bias at the semantic level that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios. |
| Outcome: | The proposed dataset shows that models reduce bias in response at term level, but reinforce bias in nuanced settings. |
UnifiedSKG: Unifying and Multi-Tasking Structured Knowledge Grounding with Text-to-Text Language Models (2022.emnlp-main)
Copied to clipboard
Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, Tao Yu
| Challenge: | Structured knowledge grounding (SKG) uses structured knowledge to complete user requests . since inputs and outputs of SKG tasks are heterogeneous, they have been studied separately . |
| Approach: | They propose a framework that unifies 21 SKG tasks into a text-to-text format . they use unifiedSKG to benchmark T5 with different sizes . |
| Outcome: | The proposed framework unifies 21 SKG tasks into a text-to-text format . it achieves state-of-the-art performance on almost all of the 21 tasks, the authors show . |
Instruction-Tuning LLMs for Event Extraction with Annotation Guidelines (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing applications of large language models to IE can be categorized into two lines: prompt engineering-based approaches and instruction-tuning open-weight LLMs. |
| Approach: | They propose to use annotation guidelines to teach large language models for event extraction . they use textual descriptions of event types and arguments to train the models . |
| Outcome: | The proposed approach improves cross-schema generalization and low-frequency event-type performance when there is a decent amount of training data. |
Learning to Simulate Natural Language Feedback for Interactive Semantic Parsing (2023.acl-long)
Copied to clipboard
| Challenge: | Existing work on interactive semantic parsing relies on human annotations to train a model . prior work relied on human-annotated feedback data, which is prohibitively expensive and not scalable . |
| Approach: | They propose a task of simulating NL feedback for interactive semantic parsing . they propose evaluators to assess the quality of the simulated feedback . |
| Outcome: | The proposed simulator can generate high-quality NL feedback to boost the error correction ability of a specific parser. |
Synthetic Question Value Estimation for Domain Adaptation of Question Answering (2022.acl-long)
Copied to clipboard
| Challenge: | Existing work adapts QA scores to select high-quality questions, but these scores do not improve QA performance on the target domain. |
| Approach: | They propose to synthesize QA pairs with a question generator on the target domain . they propose to train a Question Value Estimator that estimates usefulness of synthetic questions . |
| Outcome: | The proposed method improves the performance of the target domain QA model by using synthetic questions and only 15% of the human annotations on the targetdomain. |
Improving Generalization in Language Model-based Text-to-SQL Semantic Parsing: Two Simple Semantic Boundary-based Techniques (2023.acl-short)
Copied to clipboard
| Challenge: | Pre-trained language models (LMs)2 have been adopted for semantic parsing due to their promising performance and straightforward architectures. |
| Approach: | They propose to use token preprocessing to preserve semantic boundaries of tokens produced by LM tokenizers and special tokens to mark the boundaries of aligned components. |
| Outcome: | The proposed techniques improve the performance of pre-trained language models on two text-to-SQL semantic parsing datasets. |
Reinforced Dynamic Reasoning for Conversational Question Generation (P19-1)
Copied to clipboard
| Challenge: | Empirical results on the recently released CoQA dataset demonstrate the effectiveness of our method . large-scale highquality conversational question answering datasets such as CoQA and QuAC can help train models to answer sequential questions. |
| Approach: | They propose a task called Conversational Question Generation which generates a question based on a passage and a conversation history to generate the next question. |
| Outcome: | The proposed method is based on a question-answering style conversation dataset . it can be used to generate meaningful questions on QA and SQuAD datasets . |
Soul-Mix: Enhancing Multimodal Machine Translation with Manifold Mixup (2024.acl-long)
Copied to clipboard
| Challenge: | Multimodal machine translation (MMT) aims to improve the performance of machine translation with the help of visual information. |
| Approach: | They propose a multimodal machine translation mixup method that integrates visual information into conventional text-only neural machine translation systems. |
| Outcome: | The proposed method outperforms existing models on a multi-directional dataset with fewer parameters and achieves new state-of-the-art performance. |
Model-based Interactive Semantic Parsing: A Unified Framework and A Text-to-SQL Case Study (D19-1)
Copied to clipboard
| Challenge: | Existing semantic parsing technologies are not well-suited for use in real-world applications. |
| Approach: | They propose a model-based intelligent agent that generates a clarification question in natural language . they propose 'interactive semantic parsing' with a human user in the loop . |
| Outcome: | The proposed approach improves both parsing accuracy and user confidence . it is demonstrated on two text-to-SQL datasets with different state-of-the-art parsers . |