Papers by Haozhe Wang
How Do Humans Write Code? Large Models Do It the Same Way Too (2024.emnlp-main)
Copied to clipboard
| Challenge: | Program-of-Thought (PoT) replaces natural language-based Chain-ofThough (CoT) but introduces more reasoning errors, such as incorrect formulas or flawed logic, compared to CoT. |
| Approach: | They propose a method that integrates CoT and Program-of-Thought to achieve more accurate reasoning and reinforcement learning. |
| Outcome: | The proposed method achieves an average improvement of 6.5% on the Llama-Base model and 4.3% on the Mistral-Bass model across 8 mathematical calculation datasets. |
Knowledge-enhanced Multimodal ECG Representation Learning with Arbitrary-Lead Inputs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Current methods for multimodal representation learning for electrocardiograms often result in suboptimal alignment of ECG signals with their corresponding text reports. |
| Approach: | They propose a framework to learn ECG representations by aligning ECG signals with paired free-text reports. |
| Outcome: | The proposed framework outperforms existing methods in zero-shot classification and linear probing tasks using 12 leads. |
FaithLens: Detecting and Explaining Faithfulness Hallucination (2026.findings-acl)
Copied to clipboard
Shuzheng Si, Qingyi Wang, Haozhe Zhao, Yuzhuo Bai, Guanqiao Chen, Kangyang Luo, Gang Chen, Fanchao Qi, Minjia Zhang, Baobao Chang, Maosong Sun
| Challenge: | Recent progress in large language models (LLMs) has revolutionized text generation. |
| Approach: | They propose a faithfulness hallucination detection model that can provide binary predictions and corresponding explanations to improve trustworthiness. |
| Outcome: | The proposed model outperforms advanced models on 12 diverse tasks. |
Infinity-Parser: Layout-Aware Reinforcement Learning with High-quality Document Parsing Dataset (2026.findings-acl)
Copied to clipboard
Baode Wang, Biao Wu, Weizhen Li, Meng Fang, Zuming Huang, Jun Huang, Yanjie Liang, Haozhe Wang, Ling Chen, Wei Chu, Yuan Qi
| Challenge: | Existing supervised fine-tuning methods struggle to generalize across document types, leading to poor performance. |
| Approach: | They propose layoutRL, a reinforcement learning framework that optimizes layout understanding through composite rewards integrating normalized edit distance, paragraph count accuracy, and reading order preservation. |
| Outcome: | The proposed model outperforms specialized document parsing systems and general-purpose vision-language models on a broad range of document types, languages, and structural complexities. |
Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning (2024.acl-long)
Copied to clipboard
| Challenge: | Experimental results show that fine-tuning of large language models for specific tasks can be challenging . distribution shift during fine-timing can lead to performance degradation in general task capabilities . |
| Approach: | They propose a new approach that bridges the distribution gap between task datasets and LLMs by guiding fine-tuning with a distilled dataset generated by the model itself. |
| Outcome: | The proposed approach achieves comparable or superior performance on downstream tasks compared to the vanilla approach. |
ACECODER: Acing Coder RL via Automated Test-Case Synthesis (2025.acl-long)
Copied to clipboard
| Challenge: | Recent coder models have been driven by supervised fine-tuning, but the potential of reinforcement learning remains unexplored due to the lack of reliable reward data/model in the code domain. |
| Approach: | They propose a pipeline that generates extensive test-case pairs from existing code data and constructs preference pairs based on pass rates over sampled programs. |
| Outcome: | The proposed pipeline generates extensive (question, test-cases) pairs from existing code data and trains them with Bradley-Terry loss. |
SWE-QA-Pro: A Representative Benchmark and Scalable Training Recipe for Repository-Level Code Understanding (2026.findings-acl)
Copied to clipboard
Songcheng Cai, Zhiheng Lyu, Yuansheng Ni, Xiangchao Chen, Baichuan Zhou, Shenzhe Zhu, Yi Lu, Haozhe Wang, Chi Ruan, Benjamin Schneider, Weixu Zhang, Xiang Li, Andy Zheng, Yuyu Zhang, Ping Nie, Wenhu Chen
| Challenge: | Existing benchmarks for agentic repository-level code understanding overlook long tail topics and rely on memorized knowledge. |
| Approach: | They propose a repository-level agentic code understanding benchmark that uses long-tail repositories with executable environments to enforce topical balance. |
| Outcome: | Empirically, a Qwen3-8B model trained with the proposed benchmark outperforms GPT-4o by 2.3 points. |
Rethinking Semantic Parsing for Large Language Models: Enhancing LLM Performance with Semantic Hints (2025.acl-short)
Copied to clipboard
| Challenge: | Semantic Parsing improves performance of smaller models, but it is unclear whether it extends similarly to large language models. |
| Approach: | They propose a prompting approach that embeds semantic hints within the prompt to improve LLM performance. |
| Outcome: | The proposed approach improves LLMs’ performance across various tasks, highlighting the potential of integrating semantic information to improve LLM capabilities. |
Argus: Benchmarking and Enhancing Vision-Language Models for 3D Radiology Report Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing work on 3D radiograph report generation focuses on 2D images, but 3D medical images provide more comprehensive diagnostic information. |
| Approach: | They propose a comprehensive training recipe for building high-performing VLMs for 3DRRG using a publicly available 3D CT-report dataset. |
| Outcome: | The proposed model achieves superior performance across different model sizes and input 3D medical image resolutions. |
JointGT: Graph-Text Joint Representation Learning for Text Generation from Knowledge Graphs (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing pre-trained models for knowledgegraph-to-text generation ignore graph structure during encoding and lack elaborate pre-training tasks to explicitly model graph-text alignments. |
| Approach: | They propose a graph-text joint representation learning model called JointGT which incorporates a structure-aware semantic aggregation module into each Transformer layer to preserve the graph structure. |
| Outcome: | The proposed model achieves state-of-the-art performance on various KG-to-text datasets. |
Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering (2025.acl-long)
Copied to clipboard
Shuzheng Si, Haozhe Zhao, Gang Chen, Cheng Gao, Yuzhuo Bai, Zhitong Wang, Kaikai An, Kangyang Luo, Chen Qian, Fanchao Qi, Baobao Chang, Maosong Sun
| Challenge: | Existing studies show that training LLMs on data containing unfamiliar knowledge during instruction tuning can encourage hallucinations. |
| Approach: | They propose a framework that measures how familiar the LLM is with instruction data and introduce an expert-aligned reward model to ensure the quality of selected samples. |
| Outcome: | The proposed framework reduces hallucinations while maintaining a competitive ability to follow instructions. |
To Code or not to Code? Adaptive Tool Integration for Math Language Models via Expectation-Maximization (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing tools that integrate chain-of-thought reasoning and code execution lack metacognitive awareness to integrate tools. |
| Approach: | They propose a framework that synergizes structured exploration with off-policy RL optimization to create a cycle between metacognitive tool-use decisions and evolving capabilities. |
| Outcome: | The proposed framework improves over 11% on MATH500 and 9.4% on AIME without o1-like CoT. |
Do Large Language Models Discriminate in Hiring Decisions on the Basis of Race, Ethnicity, and Gender? (2024.acl-short)
Copied to clipboard
| Challenge: | We study whether large language models exhibit race- and gender-based name discrimination in hiring decisions . |
| Approach: | They propose templatic prompts to LLMs to write an email to a named job applicant informing them of a hiring decision. |
| Outcome: | The proposed model generates an acceptance or rejection email based on the applicant's first name . |
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 . |
Beyond Single Labels: Improving Conversational Recommendation through LLM-Powered Data Augmentation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for enhancing recommendation quality face false negatives . only one "silly cop movie" is labeled as positive, leading to suboptimal recommendations . |
| Approach: | They propose a data augmentation framework that leverages an LLM-based semantic retriever to identify diverse and semantically relevant items and filter them by a relevance scorer to remove noisy candidates. |
| Outcome: | The proposed approach improves performance on two benchmark datasets and user simulators. |
Can Medical Vision-Language Pre-training Succeed with Purely Synthetic Data? (2025.findings-acl)
Copied to clipboard
Che Liu, Zhongwei Wan, Haozhe Wang, Yinda Chen, Talha Qaiser, Chen Jin, Nikolay Burlutskiy, Fariba Yousefi, Rossella Arcucci
| Challenge: | Medical Vision-Language Pretraining (MedVLP) models typically require large-scale datasets with paired, high-quality image-text data. |
| Approach: | They propose to generate large-scale synthetic image-text pairs using off-the-shelf generative models . they propose to isolate model and training settings, focusing entirely from the data perspective. |
| Outcome: | The proposed pipeline outperforms models trained on real data by 3.8% on averaged AUC on zero-shot classification tasks. |
Massive End-to-end Speech Recognition Models with Time Reduction (2024.naacl-long)
Copied to clipboard
Weiran Wang, Rohit Prabhavalkar, Haozhe Shan, Zhong Meng, Dongseong Hwang, Qiujia Li, Khe Chai Sim, Bo Li, James Qin, Xingyu Cai, Adam Stooke, Chengjian Zheng, Yanzhang He, Tara Sainath, Pedro Moreno Mengibar
| Challenge: | Using the neural architecture of Google’s universal speech model, we reduce the frame rate and speed up training and inference. |
| Approach: | They propose to use the neural architecture of Google’s universal speech model with additional funnel pooling layers to significantly reduce the frame rate and speed up training and inference. |
| Outcome: | The proposed methods work with both connectionist temporal classification (CTC) and RNN-Transducer (RNN-T) and over two domains. |
Missing Modality meets Meta Sampling (M3S): An Efficient Universal Approach for Multimodal Sentiment Analysis with Missing Modality (2022.aacl-main)
Copied to clipboard
| Challenge: | Existing methods to address missing modalities often assume a particular modality is completely missing due to recording or transmission error. |
| Approach: | They propose a missing modality-based meta-sampling approach for multimodal sentiment analysis with missing modalities . they conduct experiments on IEMOCAP, SIMS and CMU-MOSI datasets . |
| Outcome: | The proposed method significantly improves on existing models with a mixture of missing modalities. |