Papers by Su Dong
Self-Consistency Boosts Calibration for Math Reasoning (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing solutions for math reasoning tasks use semantic parsing or AST decoding, but performance can degrade dramatically even with slight changes to the questions. |
| Approach: | They propose three calibration methods based on self-consistency for math reasoning tasks. |
| Outcome: | The proposed methods bridge model confidence and accuracy better than existing methods based on p(True) or logit. |
Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Multi-agent LLMs are rapidly moving from prototype to real-world use . network topology is a first-order security parameter in multi-aggent systems . |
| Approach: | They propose a framework for comparing topology-conditioned memory leakage in multi-agent LLM systems. |
| Outcome: | The proposed framework evaluates topology-conditioned memory leakage in multi-agent LLM systems. |
SPAGBias: Uncovering and Tracing Structured Spatial Gender Bias in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) are being used in urban planning but there is concern that they reproduce or amplify such biases. |
| Approach: | They propose a framework to evaluate spatial gender bias in large language models . they use a taxonomy of 62 urban micro-spaces, a prompt library and three diagnostic layers . |
| Outcome: | The proposed framework identifies structured gender-space associations that go beyond the public-private divide, forming nuanced micro-level mappings. |
Multilingual Encoder Knows more than You Realize: Shared Weights Pretraining for Extremely Low-Resource Languages (2025.acl-long)
Copied to clipboard
| Challenge: | XLM-R and mBART have advanced multilingualism in NLP, but low-resource languages such as Tibetan, Uyghur, Kazakh, and Mongolian are underserved. |
| Approach: | They propose a framework for adapting multilingual encoders to text generation in extremely low-resource languages by reusing the weights between the encoder and the decoder. |
| Outcome: | The proposed framework performs better on various downstream tasks even when compared with much larger models. |
Crossing the Reward Bridge: Expanding Reinforcement Learning with Verifiable Rewards Across Diverse Domains (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has been effective on structured tasks, but its reliance on simple, rule-based verifiers creates a bottleneck. |
| Approach: | They propose a framework that uses a generative verifier to provide soft, probabilistic rewards. |
| Outcome: | The proposed framework outperforms existing models up to 10x their size and can be scalable and effective. |
Exploring the Generalizability of Factual Hallucination Mitigation via Enhancing Precise Knowledge Utilization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to improve LLMs' ability to align their responses with objective facts suffer from poor generalization and trade-offs in other different capabilities. |
| Approach: | They propose to introduce PKUE (Precise Knowledge Utilization Enhancement) which fine-tunes the model on self-generated responses to precise and simple factual questions through preference optimization. |
| Outcome: | The proposed enhancements improve LLM’s ability to precisely leverage its knowledge and improve FactualBench, a comprehensive and precise factual QA dataset containing 181k Chinese data spanning 21 domains. |
Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System (2025.acl-long)
Copied to clipboard
Haoyang Su, Renqi Chen, Shixiang Tang, Zhenfei Yin, Xinzhe Zheng, Jinzhe Li, Biqing Qi, Qi Wu, Hui Li, Wanli Ouyang, Philip Torr, Bowen Zhou, Nanqing Dong
| Challenge: | Recent AI methods have shown promise in tasks such as hypothesis generation and experimental design, but they fail to replicate the collaborative nature of real-world scientific practices. |
| Approach: | They propose a virtual scientific system that mimics the collaborative nature of scientific research by organizing a team of agents to generate, evaluate, and refine research ideas. |
| Outcome: | The proposed system outperforms the state-of-the-art method in producing new scientific ideas and offers valuable insights to guide future research. |
CMHG: A Dataset and Benchmark for Headline Generation of Minority Languages in China (2025.emnlp-main)
Copied to clipboard
| Challenge: | Minority languages in China face significant challenges due to their unique writing systems, which differ from international standards. |
| Approach: | They propose a dataset specifically curated for headline generation tasks for minority languages in China . they propose 50,000 entries each for Uyghur and Mongolian, and a test set annotated by native speakers . |
| Outcome: | The proposed dataset will help improve headline generation in minority languages . it includes 100,000 entries for Tibetan, 50,000 entries each for Uyghur and Mongolian . |
OpenFact: Factuality Enhanced Open Knowledge Extraction (2023.tacl-1)
Copied to clipboard
Linfeng Song, Ante Wang, Xiaoman Pan, Hongming Zhang, Dian Yu, Lifeng Jin, Haitao Mi, Jinsong Su, Yue Zhang, Dong Yu
| Challenge: | Existing OIE systems organize knowledge into subject-relation-object (SRO) triplets, and they use templates to extract such knowledge triplet. |
| Approach: | They propose a framework to handle expressiveness and groundedness in OpenFact . they propose to use templates, extra constraints, and adopt human efforts to ensure that most triplets contain enough details. |
| Outcome: | The proposed framework improves expressiveness and groundedness of OpenFact . it is more accurate and denser than OPIEC-Linked, which is grounded to Wikidata . |
Improving LLM Generations via Fine-Grained Self-Endorsement (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent large language models (LLMs) have demonstrated remarkable capabilities but can still fail frequently on knowledge-intensive tasks. |
| Approach: | They propose a self-endorsement framework that leverages fine-grained fact-level comparisons across multiple sampled responses. |
| Outcome: | The proposed framework can improve factuality of generations with simple prompts across scales of LLMs. |
AutoBreach: Universal and Adaptive Jailbreaking with Efficient Wordplay-Guided Optimization via Multi-LLMs (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing jailbreak research exhibits limitations in universality, validity, and efficiency . Existing methods for jailbreaking LLMs have limited validity and effectiveness . |
| Approach: | They propose a black-box approach that uses wordplay-guided mapping rule sampling to create universal adversarial prompts. |
| Outcome: | The proposed method efficiently identifies security vulnerabilities across various LLMs, achieving an average success rate of over 80% with fewer than 10 queries. |
Structural Information Preserving for Graph-to-Text Generation (2020.acl-main)
Copied to clipboard
| Challenge: | Existing models that mess up or drop the core structural information of input graphs are lacking in graph-to-text generation. |
| Approach: | They propose to leverage richer training signals to guide a graph-to-text generation model by focusing on autoencoding losses and back-propagating the losses to better calibrate the model. |
| Outcome: | Experiments on two benchmarks show the proposed model over a state-of-the-art model . two types of autoencoding losses are used to back-propagate the model based on multitask training . |
APPL: A Prompt Programming Language for Harmonious Integration of Programs and Large Language Model Prompts (2025.acl-long)
Copied to clipboard
Honghua Dong, Qidong Su, Yubo Gao, Zhaoyu Li, Yangjun Ruan, Gennady Pekhimenko, Chris J. Maddison, Xujie Si
| Challenge: | Large Language Models (LLMs) are capable of handling diverse tasks with well-crafted prompts and integration of external tools. |
| Approach: | They propose a prompt programming language that acts as a bridge between computer programs and LLMs by providing convenient conventions between them. |
| Outcome: | The proposed language is intuitive, concise, and efficient through representative scenarios including Chain-of-Thought with self-consistency (CoT-SC) and ReAct tool-use agent. |
Generative Pre-trained Speech Language Model with Efficient Hierarchical Transformer (2024.acl-long)
Copied to clipboard
| Challenge: | Experimental results indicate that GPST significantly outperforms the existing speech language models in terms of word error rate, speech quality, and speaker similarity. |
| Approach: | They propose a hierarchical transformer that quantizes audio waveforms into two distinct types of discrete speech representations and integrates them within a transformer architecture. |
| Outcome: | The proposed model outperforms existing speech language models in word error rate, speech quality, and speaker similarity. |
Don’t Get Lost in the Trees: Streamlining LLM Reasoning by Overcoming Tree Search Exploration Pitfalls (2025.acl-long)
Copied to clipboard
Ante Wang, Linfeng Song, Ye Tian, Dian Yu, Haitao Mi, Xiangyu Duan, Zhaopeng Tu, Jinsong Su, Dong Yu
| Challenge: | Recent advances in tree search algorithms guided by verifiers have significantly enhanced the reasoning capabilities of large language models (LLMs), but at the cost of increased computational resources. |
| Approach: | They propose an e ffici ent tree sear ch framework that is a plug-and-play system compatible with various tree search algorithms. |
| Outcome: | The proposed framework reduces computational costs and prioritizes resource allocation to harder tasks (Levels 3-4) over simpler ones (Level 1-2), addressing both over-exploration in basic problems and under-exploation in complex cases. |
FTibSuite: A Comprehensive Resource Suite for Tibetan Vision–Language Modeling (2026.findings-acl)
Copied to clipboard
| Challenge: | FTibSuite provides an end-to-end training-and-evaluation workflow for vision–language models . Tibetan is underserved due to the lack of infrastructure for reproducible training and evaluation. |
| Approach: | They propose a resource-centric workflow for Tibetan VLMs that provides an end-to-end training-and-evaluation workflow and human-verified multimodal annotations. |
| Outcome: | FTibSuite provides an end-to-end training-and-evaluation workflow and human-verified multimodal annotations. |
Multimodal Language Models See Better When They Look Shallower (2025.emnlp-main)
Copied to clipboard
Haoran Chen, Junyan Lin, Xinghao Chen, Yue Fan, Jianfeng Dong, Xin Jin, Hui Su, Jinlan Fu, Xiaoyu Shen
| Challenge: | Existing studies show that multimodal large language models extract visual features from the final layers of a pretrained Vision Transformer. |
| Approach: | They propose a feature fusion method that strategically incorporates shallower layers . they propose MLLMs that extract visual features from the final layers of a pretrained Vision Transformer . |
| Outcome: | The proposed method outperforms deep layers on fine-grained visual tasks . it is the first comprehensive study of visual layer selection for MLLMs . |
Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering (2026.acl-long)
Copied to clipboard
Shuliang Liu, Songbo Yang, Dong Fang, Sihang Jia, Yuqi Tang, Lingfeng Su, Ruoshui Peng, Yibo Yan, Xin Zou, Xuming Hu
| Challenge: | Existing approaches to overcome object hallucination are limited . Existing mitigations include costly retraining and a training-free inference framework . |
| Approach: | They propose a training-free inference framework that simulates a metacognitive self-correction process. |
| Outcome: | The proposed framework reduces object hallucination rates by 12.67% on MMHal-Bench and improves accuracy by 5.8% on POPE. |
XL-NBT: A Cross-lingual Neural Belief Tracking Framework (D18-1)
Copied to clipboard
| Challenge: | a multi-lingual approach to training dialog systems is expensive and tedious, but it can be useful for cross-lingual support. |
| Approach: | They propose to annotate data for multiple languages and train a multi-lingual dialog system for each language. |
| Outcome: | The proposed framework bypasses the expensive human annotation and achieves promising results. |
Optimal Partial Transport Based Sentence Selection for Long-form Document Matching (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods for document matching are limited by the partial nature of the sentence-level matching signals. |
| Approach: | They propose a matching approach that equips existing document matching models with an Optimal Partial Transport component, namely OPT-Match, which selects the key sentences that play a major role in matching. |
| Outcome: | The proposed approach outperforms existing models on four publicly available datasets and the key sentences selected by it are consistent with human-provided rationales. |
Middleware for LLMs: Tools Are Instrumental for Language Agents in Complex Environments (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are generalist agents capable of operating within complex environments. |
| Approach: | They propose a class of tools that can serve as a middleware layer shielding LLMs from environmental complexity. |
| Outcome: | The proposed tool can shield the LLM from environmental complexity in two representative complex environments. |
DecIF: Improving Instruction-Following through Decomposition (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to obtain high-quality instruction-following data rely heavily on existing documents and existing methods. |
| Approach: | They propose a data synthesis framework, DecIF, which automatically generates accurate and diverse instruction-following data from scratch for supervised fine-tuning and reinforcement learning. |
| Outcome: | Extensive experiments show that the proposed framework can synthesize accurate instruction-following data for both SFT and RL paradigms compared to baselines. |
FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models (2025.emnlp-main)
Copied to clipboard
Hengxing Cai, Jinhan Dong, Jingjun Tan, Jingcheng Deng, Sihang Li, Zhifeng Gao, Haidong Wang, Zicheng Su, Agachai Sumalee, Renxin Zhong
| Challenge: | Existing methods for vision-and-language navigation struggle with insufficient multimodal fusion, weak generalization, and poor interpretability. |
| Approach: | They propose a framework for UAV vision-and-language navigation that integrates natural language instructions with visual observations to improve multimodal fusion and interpretability. |
| Outcome: | The proposed framework achieves state-of-the-art performance across all scenarios, with a 9.22% higher success rate than the strongest baseline in unseen environments. |
Optimizing NLU Reranking Using Entity Resolution Signals in Multi-domain Dialog Systems (2021.naacl-industry)
Copied to clipboard
Tong Wang, Jiangning Chen, Mohsen Malmir, Shuyan Dong, Xin He, Han Wang, Chengwei Su, Yue Liu, Yang Liu
| Challenge: | In dialog systems, the Natural Language Understanding component makes the interpretation decision before the mentioned entities are resolved. |
| Approach: | They propose to leverage Entity Resolution (ER) features in NLU reranking to learn model weights . they propose a score distribution matching method to ensure the models are calibrated . |
| Outcome: | The proposed approach outperforms the baseline model on multiple domain evaluations. |
SeedBench: A Multi-task Benchmark for Evaluating Large Language Models in Seed Science (2025.acl-long)
Copied to clipboard
Jie Ying, Zihong Chen, Zhefan Wang, Wanli Jiang, Chenyang Wang, Zhonghang Yuan, Haoyang Su, Huanjun Kong, Fan Yang, Nanqing Dong
| Challenge: | Seed science is essential for modern agriculture, but its application in seed science remains limited due to a shortage of experts and limited availability of online resources. |
| Approach: | They evaluate 26 leading large language models and compare them against a set of benchmarks . they find that there is a gap between the power of LLMs and real-world seed science problems . |
| Outcome: | The new seed benchmark highlights the gap between the power of large language models and real-world seed science problems. |
Query-Aware Knowledge Retrieval via Hyperbolic Structuring (2026.acl-long)
Copied to clipboard
Chuang Zhou, Junnan Dong, Yilin Xiao, Shengyuan Chen, Su Dong, di Yin, Xing Sun, Zhaozhuo Xu, Xiao Huang
| Challenge: | Existing approaches focus primarily on retrieving isolated factual knowledge entities while neglecting the critical reasoning relationships. |
| Approach: | They propose a query-centric retrieval framework that explicitly integrates structured knowledge graphs to support complex reasoning tasks. |
| Outcome: | Extensive experiments on three benchmark datasets show that HyperRAG outperforms baselines. |
MM-LLMs: Recent Advances in MultiModal Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | MultiModal Large Language Models (MM-LLMs) have undergone significant advances in the past year . traditional MM models incur substantial computational costs, especially when trained from scratch . |
| Approach: | They propose a taxonomy encompassing 126 MM-LLMs and summarize key training recipes to enhance their potency. |
| Outcome: | The proposed models preserve the reasoning and decision-making capabilities of LLMs and empower diverse range of MM tasks. |
CodeDPO: Aligning Code Models with Self Generated and Verified Source Code (2025.acl-long)
Copied to clipboard
| Challenge: | Existing training methods for code generation do not improve code correctness and efficiency. |
| Approach: | They propose a framework that integrates preference learning into code generation to improve code correctness and efficiency. |
| Outcome: | The proposed framework improves code correctness and efficiency by integrating preference learning into code generation. |
Document Hashing with Mixture-Prior Generative Models (D19-1)
Copied to clipboard
| Challenge: | Existing generative hashing methods only consider the use of simple priors, which limits them to further improve their performance. |
| Approach: | They propose to use Gaussian and Bernoulli priors to generate hashing codes . they propose to cast a Gausssian latent representation into binary code . |
| Outcome: | The proposed models outperform existing methods on a benchmark dataset using Gaussian and Bernoulli priors. |
Investigating and Enhancing the Robustness of Large Multimodal Models Against Temporal Inconsistency (2025.acl-long)
Copied to clipboard
Jiafeng Liang, Shixin Jiang, Xuan Dong, Ning Wang, Zheng Chu, Hui Su, Jinlan Fu, Ming Liu, See-Kiong Ng, Bing Qin
| Challenge: | Large Multimodal Models (LMMs) have demonstrated impressive performance on general video comprehension benchmarks, but their robustness needs to be thoroughly investigated for broader applications. |
| Approach: | They propose a temporal robustness benchmark which introduces temporal inconsistency perturbations separately at the visual and textual modalities to assess the robustness of models. |
| Outcome: | The proposed method improves the model’s robustness and reliability in temporal analysis. |