Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update

Huanhuan Ma, Henry Peng Zou, Chengze Li, Enze Ma, Yunyue Su, Philip S. Yu

Findings of the Association for Computational Linguistics: EMNLP 2026

We separate unsupported yielding (sycophancy) from rational updating (evidence-driven correction), and show that DPO, SFT, and activation steering all suppress sycophancy at a measurable cost to the ability to update — a trade-off we localize to a shared internal substrate.

[arXiv] [Code]

Beyond BFI: The CSI for Enhanced Reliability and Validity in Evaluating LLM Personality Traits

Huanhuan Ma, Haisong Gong, Xiaoyuan Yi, Xing Xie, Philip S. Yu, Dongkuan Xu

Preprint, 2025

We propose Core Sentiment Inventory (CSI), an implicit-association-test-inspired behavioral evaluation framework for more reliable assessment of LLM traits beyond self-report metrics.

[Paper] [Code]

Scaling LLM Agent Learning with Data Synthesis: A Comprehensive Survey

Hanrong Zhang, Yankai Chen, Shicheng Fan, Dehai Min, Shaowen Chen, Huanhuan Ma, et al.

Preprint, 2026

We survey how task specifications, trajectories, feedback signals, and environments can be synthesized to support reliable and scalable learning for LLM agents.

[Paper]

InterruptBench: Evaluating LLM Agents Under User Interruptions in Long-Horizon Web Tasks

Henry Peng Zou, Chunyu Miao, Wei-Chieh Huang, Yankai Chen, Yue Zhou, Hanrong Zhang, Yaozu Wu, Liancheng Fang, Zhengyao Gu, Zhen Zhang, Kening Zheng, Fangxin Wang, Yi Nian, Shanghao Li, Wenzhe Fan, Langzhou He, Shicheng Fan, Huanhuan Ma, Dehai Min, Weizhi Zhang, Xue Liu, Philip S. Yu

Preprint, 2026

We formalize three realistic interruption types — addition, revision, and retraction — and build a benchmark from WebArena-Lite to test whether agents can adapt to mid-task intent changes in long-horizon web navigation.

[Paper]

Infinity-MM: Scaling Multimodal Performance with Large-Scale and High-Quality Instruction Data

Shuhao Gu, Jialing Zhang, et al., Huanhuan Ma, et al.

arXiv preprint, revised 2025

We introduce a large-scale, high-quality multimodal instruction dataset and a targeted synthetic-data pipeline for scaling vision-language models.

[Paper] [Data]

EX-FEVER: A Dataset for Multi-hop Explainable Fact Verification

Huanhuan Ma, Weizhi Xu, Yifan Wei, Liuji Chen, Liang Wang, Qiang Liu, Shu Wu, Liang Wang

Findings of the Association for Computational Linguistics ACL 2024

We introduce a large scale Multi-hop fact checking dataset with textual explanations, which can be used to evaluate the explainability of fact verification models.

[PDF] [Code] [Poster] [Slide]

Interpretable Multimodal Out-of-Context Detection with Soft Logic Regularization

Huanhuan Ma^*, Jinghao Zhang^*, Qiang Liu, Shu Wu, Liang Wang

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2024

We introduce a novel multimodal out-of-context detection framework with soft logic regularization, which can effectively detect out-of-context information with interpretability.

[PDF] [Slide]

Does Knowledge Localization Hold True? Surprising Differences Between Entity and Relation Perspectives in Language Models

Yifan Wei, Xiaoyan Yu, Yixuan Weng, Huanhuan Ma, Yuanzhe Zhang, Jun Zhao, Kang Liu

Proceedings of the 33rd ACM International Conference on Information and Knowledge Management: CIKM 2024:

This study investigates the differences between entity and relational knowledge through knowledge editing. Our findings reveal that entity and relational knowledge cannot be directly transferred or mapped to each other.

[PDF] [Code]

MenatQA: A New Dataset for Testing the Temporal Comprehension and Reasoning Abilities of Large Language Models

Yifan Wei, Yisong Su, Huanhuan Ma, Xiaoyan Yu, Fangyu Lei, Yuanzhe Zhang, Jun Zhao, Kang Liu

Findings of the Association for Computational Linguistics: EMNLP 2023

We construct Multiple Sensitive Factors Time QA (MenatQA), which encompasses three temporal factors (scope factor, order factor, counterfactual factor) with total 2,853 samples for evaluating the time comprehension and reasoning abilities of LLMs.

[PDF] [Code]

Assessing knowledge editing in language models via relation perspective

Yifan Wei, Xiaoyan Yu, Huanhuan Ma, Fangyu Lei, Yixuan Weng, Ran Song, Kang Liu

[PDF] [ArXiv] [Code]

Multi-Cause Learning for Diagnosis Prediction

Liping Wang, Qiang Liu, Huanhuan Ma, Shu Wu, Liang Wang

Data Mining and Big Data (DMBD) 2022

[PDF]