Papers by Tianyu Zhao
Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and Question Answering Tasks (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are trained on vast corpora that contain substantial knowledge but their outputs often contain confidently stated inaccuracies. |
| Approach: | They propose to encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. |
| Outcome: | The proposed model can generalize to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. |
Topic-relevant Response Generation using Optimal Transport for an Open-domain Dialog System (2020.coling-main)
Copied to clipboard
| Challenge: | Conventional neural generative models generate safe and generic responses which have little connection with previous utterances semantically and would disengage users in a dialog system. |
| Approach: | They propose a method that employs topical constraint and semantic constraint to generate relevant responses by regularizing the decoding objective function with semantic distance. |
| Outcome: | The proposed method generates more topic-relevant and content-rich responses than conventional models. |
Beyond Polarity: Continuous Affect-Enhanced Multimodal Aspect-Based Sentiment Classification (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for multimodal aspect-based sentiment classification exploit discrete polarity patterns and generic visual embeddings. |
| Approach: | They propose a Valence–Arousal–Dominance(VAD)-Enhanced MABSC framework that integrates VAD signals into multimodal sentiment reasoning and learns emotion-sensitive image representations. |
| Outcome: | The proposed framework brings VAD signals into multimodal sentiment reasoning and learns emotion-sensitive image representations. |
M2PO: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation (2026.acl-long)
Copied to clipboard
Hao Wang, Linlong Xu, Heng Liu, Yangyang Liu, Xiaohu Zhao, Bo Zeng, Liangying Shao, Yichen Dong, Xinwei Wu, Jiang Zhou, Tianyu Dong, Xiangxiang Zeng, Longyue Wang, Weihua Luo
| Challenge: | prevailing methods for machine translation are often hindered by misleading reward signals. |
| Approach: | They propose a framework that aligns large language models to human preferences . they propose 'M2PO' to correct the bias towards partial errors . |
| Outcome: | The proposed framework outperforms open-source models and achieves parity with proprietary models. |
DialogUSR: Complex Dialogue Utterance Splitting and Reformulation for Multiple Intent Detection (2022.findings-emnlp)
Copied to clipboard
Haoran Meng, Zheng Xin, Tianyu Liu, Zizhen Wang, He Feng, Binghuai Lin, Xuemin Zhao, Yunbo Cao, Zhifang Sui
| Challenge: | DialogUSR is a plug-in and domain-agnostic module that empowers multi-intent detection for chatbots . a single user query triggers inquiries on highspeed train ticket price and weather of destination. |
| Approach: | They propose a dialog utterance splitting and reformulation task that splits multi-intent user query into multiple single-intention sub-queries and recovers all coreferred and omitted information in the sub-questions. |
| Outcome: | The proposed model can be used to split multi-intent user queries into multiple sub-queries . it can be trained in two stages and perform in-depth analyses on the proposed models . |
Focused Prefix Tuning for Controllable Text Generation (2023.acl-short)
Copied to clipboard
| Challenge: | Existing unannotated attributes could degrade models' performance . focus on the desired attribute can be achieved with focused prefix tuning . |
| Approach: | They propose focused prefix tuning to enable the control to focus on the desired attribute . they propose to reduce the number of unannotated attributes in a controllable text generation dataset . |
| Outcome: | The proposed approach achieves better control accuracy and text fluency than baseline models in single-attribute tasks. |
Designing Precise and Robust Dialogue Response Evaluators (2020.acl-main)
Copied to clipboard
| Challenge: | Existing automated dialogue response evaluators have only moderate correlation with human judgement and are not robust. |
| Approach: | They propose to build a reference-free dialogue response evaluator that exploits the power of semi-supervised training and pretrained (masked) language models. |
| Outcome: | The proposed model achieves strong correlation with human judgement and generalizes robustly to diverse responses and corpora. |
Named Entity Recognition with Small Strongly Labeled and Large Weakly Labeled Data (2021.acl-long)
Copied to clipboard
| Challenge: | Existing work focuses on learning deep NER models with weak supervision without any human annotation. |
| Approach: | They propose a framework that can suppress the noise of the weak labels and fine-tune over the strongly labeled data. |
| Outcome: | The proposed framework outperforms existing methods on Named Entity Recognition tasks with weak supervision and weakly labeled data. |
GiLT: Augmenting Transformer Language Models with Dependency Graphs (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work focuses on syntactic tree structures of languages, in particular constituency tree structures. |
| Approach: | They propose a Graph-Infused Layers Transformer Language Model which leverages dependency graphs to augment Transformer language models. |
| Outcome: | The proposed model achieves better syntactic generalization while maintaining competitive perplexity compared with baseline models. |
Learning Contextualized Knowledge Structures for Commonsense Reasoning (2021.findings-acl)
Copied to clipboard
Jun Yan, Mrigank Raman, Aaron Chan, Tianyu Zhang, Ryan Rossi, Handong Zhao, Sungchul Kim, Nedim Lipka, Xiang Ren
| Challenge: | Recent knowledge graph (KG) augmented models have achieved notable success on commonsense reasoning tasks. |
| Approach: | They propose a KG-augmented model that contextualizes extracted and generated knowledge by reasoning over both within a single graph structure. |
| Outcome: | The proposed model outperforms existing models on four commonsense reasoning benchmarks and a user study on edge validness and helpfulness. |
Physics: Benchmarking Foundation Models on University-Level Physics Problem Solving (2025.findings-acl)
Copied to clipboard
| Challenge: | a benchmark for university-level physics problem solving contains 1,297 expert-annotated problems . a proprietary model, o3-mini, achieves only 59.9% accuracy, highlighting fundamental weaknesses in scientific reasoning, conceptual understanding, and mathematical precision. |
| Approach: | They introduce Physics, a benchmark for university-level physics problem solving. |
| Outcome: | The proposed model achieves only 59.9% accuracy on the most advanced model, o3-mini . the proposed model is a powerful tool for evaluating models on advanced problems . |
ReEx-SQL: Reasoning with Execution-Aware Reinforcement Learning for Text-to-SQL (2026.acl-long)
Copied to clipboard
Yaxun Dai, Wenxuan Xie, Xialie Zhuang, Tianyu Yang, Ziyi Liu, Haiqin Yang, Yiying Yang, Yuhang Zhao, Pingfu Chao, Wenhao Jiang
| Challenge: | Current Text-to-SQL reasoning models lack integrated execution feedback during generation. |
| Approach: | They propose a text-to-SQL framework that interacts with the SQL execution engine during decoding and dynamically adjusts reasoning based on execution feedback. |
| Outcome: | The proposed framework achieves 89.1% accuracy on Spider and 65.3% on BIRD at the 7B scale. |
Integrating Pre-Trained Speech and Language Models for End-to-End Speech Recognition (2024.findings-acl)
Copied to clipboard
| Challenge: | Mainstream of automatic speech recognition (ASR) has shifted from pipeline methods to end-to-end (E2E) methods. |
| Approach: | They propose to integrate a pre-trained speech representation model and a large language model (LLM) for automatic speech recognition in an end-to-end manner. |
| Outcome: | The proposed model achieves comparable performance to modern E2E ASR models by utilizing powerful pre-training models with the proposed integrated approach. |
PCA-Bench: Evaluating Multimodal Large Language Models in Perception-Cognition-Action Chain (2024.findings-acl)
Copied to clipboard
Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Xiangdi Meng, Tianyu Liu, Baobao Chang
| Challenge: | a new multimodal decision-making benchmark evaluates the integrated capabilities of multimodal large language models. |
| Approach: | They propose a multimodal decision-making benchmark for evaluating MLLMs . they propose an automatic evaluation protocol to assess 10 prevalent ML models . |
| Outcome: | The proposed benchmark improves performance of multimodal large language models in three scenarios . the model is required to integrate multiple capabilities to make accurate decisions . |
A Survey of Multimodal Mathematical Reasoning: From Perception, Alignment to Reasoning (2026.acl-long)
Copied to clipboard
Tianyu Yang, Sihong Wu, Yilun Zhao, Zhenwen Liang, Lisen Dai, Chen Zhao, Minhao Cheng, Arman Cohan, Xiangliang Zhang
| Challenge: | Multimodal mathematical Reasoning (MMR) has attracted increasing attention for its ability to solve mathematical problems involving both textual and visual modalities. |
| Approach: | They review the theoretical frameworks of multimodal reasoning and examine the challenges they face in visual math tasks. |
| Outcome: | The proposed models can solve problems involving both textual and visual modalities. |
Release of Pre-Trained Models for the Japanese Language (2024.lrec-main)
Copied to clipboard
Kei Sawada, Tianyu Zhao, Makoto Shing, Kentaro Mitsui, Akio Kaga, Yukiya Hono, Toshiaki Wakatsuki, Koh Mitsuda
| Challenge: | democratization of AI aims to create a world where everyone can use AI . pre-trained models with high performance in Japanese are lagging in non-English-speaking communities . |
| Approach: | et al. released large-scale pre-trained models trained on large-data to improve access to AI . authors say the models are more accurate and more accurate than those trained in the English language . e-mail protected: email protected. |
| Outcome: | a new study shows that pre-trained models specialized for Japanese can achieve high performance in Japanese tasks. |
MirrorCAPTCHA: Wild CAPTCHA, Wild Distribution, Wild Web-based Platform Meet Multimodal LLM Agents (2026.acl-long)
Copied to clipboard
Xiangyu Wu, Yuwei Hu, Tianyu Cui, Yueying Tian, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Yang Yang, Jianfeng Lu
| Challenge: | Existing agent benchmarks fail to evaluate an agent's real-world capacity to handle CAPTCHA . Existing benchmarks ignore this practical challenge, failing to evaluate agents' ability to handle complex visual CAPTchas. |
| Approach: | They propose a benchmark annotated with Weighted Pass Rate and a new metric to measure agent's ability to handle CAPTCHA. |
| Outcome: | The proposed benchmark outperforms current state-of-the-art closed-source models on mirrorCAPTCHA and achieves 9.4% higher average weighted pass rate and 2.13% higher average Completion degree. |
Incentivizing Parametric Knowledge via Reinforcement Learning with Verifiable Rewards for Cross-Cultural Entity Translation (2026.acl-long)
Copied to clipboard
Jiang Zhou, Xiaohu Zhao, Xinwei Wu, Tianyu Dong, Hao Wang, Yangyang Liu, Heng Liu, Linlong Xu, Longyue Wang, Weihua Luo, Deyi Xiong
| Challenge: | Current systems often fall short of this goal in settings where translation hinges on culturally grounded entities such as books, films, places, songs and idioms. |
| Approach: | They propose a framework that anchors supervision on a verifiable, entity-level reward signal and incorporates lightweight structural gates to stabilize optimization. |
| Outcome: | The proposed framework improves on XC-Translate and shows that it can learn a robust reasoning process rather than imitating reference translations. |
Debiasing Large Language Models with Structured Knowledge (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to reduce biases in pre-training models are hampered by their performance. |
| Approach: | They propose a method that utilizes structured knowledge to mitigate bias in LLMs . their method obviates the need for training from scratch, thus offering enhanced scalability . |
| Outcome: | The proposed method outperforms state-of-the-art (SOTA) baselines in the debiasing ability. |
SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment (2026.findings-acl)
Copied to clipboard
Tianyu Dong, Yangyang Liu, Jiang Zhou, Xinwei Wu, Xiaohu Zhao, Hao Wang, Heng Liu, Linlong Xu, Longyue Wang, Weihua Luo, Shaolin Zhu, Deyi Xiong
| Challenge: | Low-resource language tokens are often routed to different experts than those activated by high-resourced inputs, which hinders their efficacy in multilingual contexts. |
| Approach: | They propose a framework to transfer specialized capabilities from high-resource languages as anchors to low-resourced languages by using a symmetric Jensen-Shannon constraint. |
| Outcome: | The proposed framework outperforms standard instruction tuning on 5 low-resource languages and 3 benchmarks. |
HSS-Synth: Humanities and Social Sciences Data Synthesis for LLMs (2026.findings-acl)
Copied to clipboard
Ru Peng, Tianyu Zhao, Xijun Gu, Zhiting Fan, Haokai Xu, Jinyang Zhang, Yawen Zeng, Yihong Zhuang, Kexin Yang, Junyang Lin, Dayiheng Liu, Junbo Zhao
| Challenge: | High-quality, diverse data are vital for large language models (LLMs) but remain scarce and costly. |
| Approach: | They define the first HSS domain system covering 14 mainstream fields and introduce HSS-Synth. |
| Outcome: | the proposed pipeline outperforms 14 leading baselines on 16 benchmarks. |
Bridging the Temporal Gap in Multimodal LLMs: Deeply Stacking Temporal Tokens for Audio-Visual Speech Recognition (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing audio-visual speech recognition systems suffer from a temporal gap . visual speech patterns captured from lip movements provide complementary information that remains inherently robust to acoustic noise. |
| Approach: | They propose a framework that deeply stacks temporal tokens across both encoding and decoding stages to bridge this temporal gap. |
| Outcome: | The proposed framework outperforms existing supervised, self-supervised, and LLM-based methods by 6.1% on LRS2 and 7.8% on LLS3. |
Geneverse: A Collection of Open-source Multimodal Large Language Models for Genomic and Proteomic Research (2024.findings-emnlp)
Copied to clipboard
| Challenge: | generative Large Language Models (LLMs) are a promising tool for biomedical and healthcare research. |
| Approach: | They propose to use finetuned LLMs and multimodal LLM for genomic and proteomics tasks. |
| Outcome: | The proposed models outperform closed-source models in genomic and proteomics tasks and are highly accurate. |
MMEvalPro: Calibrating Multimodal Benchmarks Towards Trustworthy and Efficient Evaluation (2025.naacl-long)
Copied to clipboard
Jinsheng Huang, Liang Chen, Taian Guo, Fu Zeng, Yusheng Zhao, Bohan Wu, Ye Yuan, Haozhe Zhao, Zhihui Guo, Yichi Zhang, Jingyang Yuan, Wei Ju, Luchen Liu, Tianyu Liu, Baobao Chang, Ming Zhang
| Challenge: | Large Multimodal Models (LMMs) exhibit impressive cross-modal understanding and reasoning abilities, but many benchmarks suffer from systematic biases. |
| Approach: | They propose a benchmark to avoid Type-I errors by creating one perception question and one knowledge anchor question through a meticulous annotation process. |
| Outcome: | The proposed benchmark avoids Type-I errors while maintaining reliability of MCQ evaluations. |