Papers by Haonan Wang
Removing Prompt-template Bias in Reinforcement Learning from Human Feedback (2025.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback (RLHF) has shown promise for enhancing pre-trained large language models to generate responses that align with human preferences and societal values. |
| Approach: | They propose a method to estimate prompt-template bias term during reward modeling and use it to calibrate reward scores. |
| Outcome: | The proposed method can be flexibly combined with existing algorithms of removing length bias, leading to a further improvement in the aspect of enhancing the quality of generated responses. |
LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Existing vector steering methods adjust the magnitude of answer vectors, but this creates a fundamental trade-off—reducing jailbreak increases over-refusal. |
| Approach: | They propose a method which aligns va with vb through closed-form weight updates, making the model’s willingness to respond causally dependent on its safety assessment. |
| Outcome: | Experiments on 12 LLMs show that the proposed method achieves 11.45% higher F1 than the best baseline while preserving 95.92% utility. |
Resource-Limited Joint Multimodal Sentiment Reasoning and Classification via Chain-of-Thought Enhancement and Distillation (2026.findings-acl)
Copied to clipboard
| Challenge: | Current approaches for Multimodal Sentiment Analysis (MSA) rely on parameter-heavy LLMs for classification, overlooking multimodal sentiment reasoning generation in resource-limited environments. |
| Approach: | They propose a multimodal sentiment reasoning distillation model that employs a teacher-assistant-student paradigm to address deployment constraints in resource-limited environments. |
| Outcome: | The proposed model performs well on a resource-limited JMSRC task with only 3B parameters and shows generalization and interpretability. |
A Chinese Dataset for Evaluating the Safeguards in Large Language Models (2024.findings-acl)
Copied to clipboard
Yuxia Wang, Zenan Zhai, Haonan Li, Xudong Han, Shom Lin, Zhenxuan Zhang, Angela Zhao, Preslav Nakov, Timothy Baldwin
| Challenge: | a recent study has shown that large language models can produce harmful responses, exposing users to unexpected risks. |
| Approach: | They propose a dataset for the safety evaluation of Chinese LLMs in Mandarin Chinese . they extend the dataset to better identify false negative and false positive examples . |
| Outcome: | The proposed dataset is for the safety evaluation of Chinese LLMs, and is based on a Chinese dataset. |
Prompt Space Optimizing Few-shot Reasoning Success with Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Prompt engineering is an essential technique for enhancing the abilities of large language models (LLMs) by providing explicit and specific instructions. |
| Approach: | They propose a new approach that uses text embeddings to obtain basis vectors by matrix decomposition and constructs a space for representing all prompts. |
| Outcome: | The proposed approach significantly outperforms state-of-the-art prompt paradigms on ten public reasoning benchmarks. |
Sticking to the Mean: Detecting Sticky Tokens in Text Embedding Models (2025.acl-long)
Copied to clipboard
| Challenge: | Sticky tokens, when repeatedly inserted into sentences, pull sentence similarity toward a certain value, disrupting the normal distribution of embedding distances and degrading downstream performance. |
| Approach: | They propose a method to detect “sticky tokens” by sentence and token filtering and apply it to 40 checkpoints across 14 model families. |
| Outcome: | The proposed method detects 868 sticky tokens across 14 models and shows that their presence does not correlate with model size or vocabulary size. |
MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct (2025.findings-acl)
Copied to clipboard
Run Luo, Haonan Zhang, Longze Chen, Ting-En Lin, Xiong Liu, Yuchuan Wu, Min Yang, Yongbin Li, Minzheng Wang, Pengpeng Zeng, Lianli Gao, Heng Tao Shen, Yunshui Li, Hamid Alinejad-Rokny, Xiaobo Xia, Jingkuan Song, Fei Huang
| Challenge: | a new framework for image-text instruction data evolution improves MLLM performance . lack of high-quality instruction data remains a major bottleneck in ML modeling . |
| Approach: | They propose a multimodal instruction data evolution framework that iteratively enhances data quality through fine-grained perception, cognitive reasoning, and interaction evolution. |
| Outcome: | The proposed approach improves MLLM performance in nine vision-language tasks while using significantly less data. |
LongSpec: Long-Context Lossless Speculative Decoding with Efficient Drafting and Verification (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can process extremely long contexts, requiring efficient inference over extended inputs. |
| Approach: | They propose a model that uses a constant-sized key-value cache to train long-context models. |
| Outcome: | Experimental results show that LongSpec achieves 3.26x speedup over strong Flash Attention baselines and 2.34x wall clock time on four math reasoning tasks. |
FinChain: A Symbolic Benchmark for Verifiable Chain-of-Thought Financial Reasoning (2026.acl-long)
Copied to clipboard
Zhuohan Xie, Daniil Orel, Rushil Thareja, Dhruv Sahnan, Hachem Madmoun, Fan Zhang, Debopriyo Banerjee, Georgi Nenkov Georgiev, Xueqing Peng, Lingfei Qian, Jimin Huang, Jinyan Su, Aaryamonvikram Singh, Rui Xing, Rania Elbadry, Chen Xu, Haonan Li, Fajri Koto, Ivan Koychev, Tanmoy Chakraborty, Yuxia Wang, Salem Lahlou, Veselin Stoyanov, Sophia Ananiadou, Preslav Nakov
| Challenge: | Existing benchmarks emphasize final numerical answers while neglecting intermediate reasoning steps. |
| Approach: | They propose a symbolic benchmark for verifiable Chain-of-Thought evaluation in finance . FINCHAIN spans 58 topics across 12 financial domains and three difficulty levels . |
| Outcome: | The proposed benchmark aims to bridge symbolic reasoning and factual verification. |
Loki: An Open-Source Tool for Fact Verification (2025.coling-demos)
Copied to clipboard
Haonan Li, Xudong Han, Hao Wang, Yuxia Wang, Minghan Wang, Rui Xing, Yilin Geng, Zenan Zhai, Preslav Nakov, Timothy Baldwin
| Challenge: | Loki is an open-source fact-checking tool designed to address the growing problem of misinformation. |
| Approach: | They propose a tool that breaks down the fact-checking task into five steps . they propose LOKI, which offers a semiautomated, human-in-the-loop approach . |
| Outcome: | a new open-source tool is designed to address the growing problem of misinformation . the tool breaks down the fact-checking task into five steps to assist human judgment . |
MRFD: Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Vision-Language Models often produce hallucinations due to the limited ability to verify information in different regions of the image. |
| Approach: | a new decoding method improves factual grounding by modeling inter-region consistency . the method identifies salient regions using cross-attention and generates initial responses for each . |
| Outcome: | a training-free decoding method reduces hallucinations and improves response consistency . the proposed method generates initial responses for each region and weights reliability weights among responses . |
SCALAR: Scientific Citation-based Live Assessment of Long-context Academic Reasoning (2026.eacl-long)
Copied to clipboard
| Challenge: | Long-context understanding is a critical capability for large language models . evaluating this capability requires extensive human annotation, which is time-consuming and costly. |
| Approach: | They propose a benchmark to assess citation-grounded long-context reasoning in academic writing. |
| Outcome: | The proposed benchmark compares state-of-the-art models with human experts on two tasks . human experts achieve 90% accuracy, but most models struggle with the cloze-style task . |
VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation (2024.emnlp-main)
Copied to clipboard
Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, Kai Wang, Quy Do, Yuansheng Ni, Bohan Lyu, Yaswanth Narsupalli, Rongqi Fan, Zhiheng Lyu, Bill Yuchen Lin, Wenhu Chen
| Challenge: | Existing video metrics are lagging behind in providing reliable scores over generated videos due to lack of large-scale human-annotated dataset. |
| Approach: | They propose to use VideoFeedback to train a human-annotated multi-aspect score over 37.6K synthesized videos from 11 existing video generative models. |
| Outcome: | The proposed model outperforms the prior best metrics by 50 points in the test. |
Do-Not-Answer: Evaluating Safeguards in LLMs (2024.findings-eacl)
Copied to clipboard
| Challenge: | a dataset evaluating harmful capabilities in large language models is available at https://github.com/Libr-AI/do-not-answer. |
| Approach: | They collect an open-source dataset to evaluate the safeguards in large language models . they find that simple BERT-style classifiers can achieve results comparable to GPT-4 . |
| Outcome: | The proposed dataset compares the safety of six popular LLMs to GPT-4 on automatic safety evaluation. |
NAT: Enhancing Agent Tuning with Negative Samples (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning and reinforcement learning use only positive examples, limiting their efficiency in low-resource scenarios. |
| Approach: | They propose a method that leverages both successful and failed trajectories for fine-tuning, maximizing the utility of limited resources. |
| Outcome: | The proposed method surpasses existing methods, including SFT, DPO, and PPO, across various tasks. |
Can Multimodal LLMs See Materials Clearly? A Multimodal Benchmark on Materials Characterization (2025.findings-emnlp)
Copied to clipboard
Zhengzhao Lai, Youbin Zheng, Zhenyang Cai, Haonan Lyu, Jingpu Yang, Hong-Qing Liang, Yan Hu, Benyou Wang
| Challenge: | characterization imaging data is fundamental to acquiring materials information. |
| Approach: | a team of researchers develop a benchmark for materials characterization imaging data . the goal is to bridge this gap by addressing 1,500 questions that require expert-level expertise. |
| Outcome: | a new benchmark for materials characterization imaging data is presented . the benchmark reveals that MLLMs perform poorly when addressing higher-level questions . |
From Query to Logic: Ontology-Driven Multi-Hop Reasoning in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) exhibit limitations in complex multi-hop question answering tasks that necessitate non-linear, structured reasoning. |
| Approach: | They propose an ontology-driven reasoning and chain framework that combines LLMs’ generative capabilities with the structural benefits of knowledge graphs. |
| Outcome: | Extensive experiments across a diverse set of models and standard MQA benchmarks demonstrate that the proposed framework achieves competitive performance while producing more interpretable reasoning chains. |
Little Giants: Synthesizing High-Quality Embedding Data at Scale (2025.naacl-long)
Copied to clipboard
| Challenge: | Synthetic data generation is an increasingly popular way of training models without the need for large, manually labeled datasets. |
| Approach: | They propose a framework that aligns open-source small models to efficiently generate large-scale embedding data. |
| Outcome: | The proposed framework outperforms state-of-the-art embedding models by using only 1/10 of the GPT API calls. |
InsCL: A Data-efficient Continual Learning Paradigm for Fine-tuning Large Language Models with Instructions (2024.naacl-long)
Copied to clipboard
| Challenge: | In order to perform downstream tasks, Large Language Models (LLMs) need continual adaptation without catastrophic forgetting. |
| Approach: | They propose a new paradigm that allows for continual adaptation without catastrophic forgetting . they propose to replay previous data based on task similarity with instructions . |
| Outcome: | The proposed method improves performance over 16 tasks with different training orders. |
Generalized Category Discovery with Large Language Models in the Loop (2024.findings-acl)
Copied to clipboard
Wenbin An, Wenkai Shi, Feng Tian, Haonan Lin, QianYing Wang, Yaqiang Wu, Mingxiang Cai, Luyan Wang, Yan Chen, Haiping Zhu, Ping Chen
| Challenge: | Generalized Category Discovery (GCD) is a crucial task that aims to recognize both known and novel categories from a set of unlabeled data. |
| Approach: | They propose a framework that introduces Large Language Models into the training loop to generate category names without human effort. |
| Outcome: | The proposed framework outperforms SOTA models on three benchmark datasets and generates accurate category names for the discovered clusters. |
RealMem: Benchmarking LLMs in Real-World Memory-Driven Interaction (2026.findings-acl)
Copied to clipboard
Haonan Bian, Zhiyuan Yao, Sen Hu, Zishan Xu, Shaolei Zhang, Yifu Guo, Ziliang Yang, Xueran Han, Huacan Wang, Ronghao Chen
| Challenge: | Existing benchmarks focus on casual conversation or task-oriented dialogue, failing to capture “long-term project-oriented” interactions where agents must track evolving goals. |
| Approach: | They propose a benchmark that simulates the dynamic evolution of memory in real-world projects. |
| Outcome: | The proposed benchmarks simulate the dynamic evolution of memory in real-world projects. |
mmE5: Improving Multimodal Multilingual Embeddings via High-quality Synthetic Data (2025.findings-acl)
Copied to clipboard
| Challenge: | Multimodal embedding models encode multimedia inputs into latent vector representations. |
| Approach: | They propose to synthesize multimodal multilingual data using a multimodal large language model . they identify three criteria for high-quality synthetic multimodal data . |
| Outcome: | The proposed model outperforms existing models on the MMEB Benchmark and the XTD benchmark. |
Getting More Juice Out of Your Data: Hard Pair Refinement Enhances Visual-Language Models Without Extra Data (2025.naacl-long)
Copied to clipboard
Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, Zhenguo Li, Hong Cheng, Kenji Kawaguchi
| Challenge: | Contrastive Language-Image Pre-training (CLIP) is a standard for cross-modal image-text representation learning. |
| Approach: | They propose a framework that enhances pre-trained CLIP models by exploiting challenging text-image pairs within existing datasets. |
| Outcome: | The proposed framework improves CLIP models by exploiting text-image pairs in training. |
Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation (2026.acl-long)
Copied to clipboard
| Challenge: | X (formerly Twitter) users can flag misleading posts, attach contextual notes, and rate the notes’ helpfulness, but there is a significant latency in Community Notes, which is unable to provide accurate notes. |
| Approach: | They propose a framework that augments Community Notes for faster and more reliable health misinformation governance. |
| Outcome: | The proposed framework outperforms human contributors in correctness, helpfulness, and evidence utility in health misinformation surges. |
Demystifying Instruction Mixing for Fine-tuning Large Language Models (2024.acl-srw)
Copied to clipboard
| Challenge: | Instruction tuning is effective for aligning large language models with human instructions, but the procedure to optimizing the mixing of instruction datasets is still unclear. |
| Approach: | They categorize instructions into three primary types: NLP downstream tasks, coding, and general chat. |
| Outcome: | The proposed method improves performance of large language models (LLMs) but it is difficult to combine different instruction datasets to optimize overall performance. |
MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to embed multimodal models face limitations such as suboptimal causal attention in VLMs and limited diversity in training objectives and data. |
| Approach: | They propose a framework for transforming pre-trained VLMs into bidirectional multimodal embedding models. |
| Outcome: | The proposed model improves performance across MMEB and ViDoRe-v2 benchmarks and exhibits strong scalability with both model size and training data on MMEF. |
AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use (2026.acl-industry)
Copied to clipboard
| Challenge: | Modern industrial applications increasingly demand language models capable of multi-step reasoning and tool use in real-world settings. |
| Approach: | They propose a model family that trains via multi-round reinforcement learning on synthetic data and open-source data. |
| Outcome: | The proposed model train on synthetic and open-source data achieves strong performance on multiple agentic benchmarks and in an industrial agent system. |
TinyJudge: Unverifiable Constraint Alignment via Lightweight Specialist Ensembles (2026.acl-long)
Copied to clipboard
Yirong Zeng, Yufei Liu, Xiao Ding, Yutai Hou, Yuxian Wang, Wu Ning, Haonan Song, Dandan Tu, Qixun Zhang, Yuxiang He, Bibo Cai, Ting Liu
| Challenge: | Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints. |
| Approach: | They propose a framework that uses tiny language models to evaluate instruction following . they propose to use a set of specialized tiny language model to provide rewards for soft constraints. |
| Outcome: | The proposed framework outperforms baseline models by 12% and speeds up training time by 3. |