Papers by Lei Song
CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning (2023.acl-long)
Copied to clipboard
| Challenge: | HKUST-KnowComp proposes a framework for commonsense reasoning that can be used to conceptualize commonsence knowledge bases at scale. |
| Approach: | They propose a framework that integrates event conceptualization and instantiation to conceptualize commonsense knowledge bases at scale. |
| Outcome: | The proposed framework achieves state-of-the-art on two conceptualization tasks and the acquired abstract commonsense knowledge significantly improves commonsence inference modeling. |
We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning? (2025.acl-long)
Copied to clipboard
Runqi Qiao, Qiuna Tan, Guanting Dong, MinhuiWu MinhuiWu, Chong Sun, Xiaoshuai Song, Jiapeng Wang, Zhuoma GongQue, Shanglin Lei, YiFan Zhang, Zhe Wei, Miaoxuan Zhang, Runfeng Qiao, Xiao Zong, Yida Xu, Peiqing Yang, Zhimin Bao, Muxi Diao, Chen Li, Honggang Zhang
| Challenge: | Existing benchmarks focus more on end-to-end performance, but neglect the underlying principles of knowledge acquisition and generalization. |
| Approach: | They propose a benchmark specifically designed to explore the problem-solving principles by decomposing 6.5K visual math problems into 10.9K step-level questions for evaluation. |
| Outcome: | The proposed benchmark covers 6.5K visual math problems and 10.9K step-level questions spanning 5 layers of knowledge granularity and 67 hierarchical knowledge concepts. |
Multilingual Blending: Large Language Model Safety Alignment Evaluation with Language Mixture (2025.findings-naacl)
Copied to clipboard
| Challenge: | a range of representative Large Language Models have exhibited remarkable generalization capabilities across numerous downstream tasks. |
| Approach: | They propose a query-response scheme to evaluate the safety alignment of LLMs . they found that multilingual query-responding significantly amplifies the detriment of malicious queries . |
| Outcome: | The proposed scheme improves the safety alignment of state-of-the-art LLMs under multilingual conditions. |
From Hypothesis to Publication: A Comprehensive Survey of AI-Driven Research Support Systems (2025.findings-emnlp)
Copied to clipboard
Zekun Zhou, Xiaocheng Feng, Lei Huang, Xiachong Feng, Ziyun Song, Ruihan Chen, Liang Zhao, Weitao Ma, Yuxuan Gu, Baoxin Wang, Dayong Wu, Guoping Hu, Ting Liu, Bing Qin
| Challenge: | rapid development of artificial intelligence (AI) technologies has inspired researchers to explore how AI can accelerate and enhance research. |
| Approach: | They organize the relevant studies into three main categories: hypothesis formulation, hypothesis validation, and manuscript publication. |
| Outcome: | The authors summarize the current state of research in three main areas: hypothesis formulation, hypothesis validation, and manuscript publication. |
Hire a Linguist!: Learning Endangered Languages in LLMs with In-Context Linguistic Descriptions (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing LLMs rarely perform well in unseen, endangered languages . Existing models such as Llama and GPT-4 lack a rich corpus of training data . |
| Approach: | They propose a training-free approach to enable an LLM to process unseen languages that hardly occur in its pre-training. |
| Outcome: | The proposed approach elevates translation capability from GPT-4’s 0 to 10.5 BLEU for 10 language directions. |
Structural Bias for Aspect Sentiment Triplet Extraction (2022.coling-1)
Copied to clipboard
| Challenge: | Existing structural bias adapters for aspect sentiment triplet extraction are under-confident . a large-scale dataset for ASTE shows the adapter is effective and efficient to a larger scale. |
| Approach: | They propose to use a structural adapter to integrate structural bias into pretrained language models . they propose to add a relative position structure in place of the syntactic dependency structure . |
| Outcome: | The proposed adapter achieves state-of-the-art performance over strong baselines, but with a light parameter demand and low latency. |
RPC-Bench: A Fine-grained Benchmark for Research Paper Comprehension (2026.acl-long)
Copied to clipboard
Yelin Chen, Fanjin Zhang, Suping Sun, Yunhe Pang, Yuanchun Wang, Jian Song, XiaoYan Li, Lei Hou, Shu Zhao, Jie Tang, Juanzi Li
| Challenge: | Existing benchmarks for understanding research papers offer limited fine-grained evaluation at scale. |
| Approach: | They propose a large-scale question-answering benchmark built from review–rebuttal exchanges of high-quality computer science papers. |
| Outcome: | The proposed model is based on human-verified QA pairs and contains 15K questions. |
MixTEA: Semi-supervised Entity Alignment with Mixture Teaching (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to learn informative entity embeddings are insufficient for semi-supervised entity alignment. |
| Approach: | They propose a semi-supervised method which guides the model learning with an end-to-end mixture teaching of manually labeled mappings and probabilistic pseudo mappings. |
| Outcome: | The proposed method is superior to existing methods on benchmark datasets and further analyses. |
Time-RA: Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback (2026.findings-acl)
Copied to clipboard
Yiyuan Yang, Zichuan Liu, Lei Song, Kai Ying, Stephen Wang, Joshua Thomas Bamford, Svitlana Vyetrenko, Jiang Bian, Qingsong Wen
| Challenge: | Time series anomaly detection (TSAD) has traditionally focused on binary classification and lacks the fine-grained categorization and explanatory reasoning required for transparent decision-making. |
| Approach: | They propose a time-series reasoning task that reformulates TSAD from discriminative to reasoning-intensive paradigm. |
| Outcome: | The proposed task reformulates TSAD from discriminative to reasoning-intensive paradigm. |
MTG: A Benchmark Suite for Multilingual Text Generation (2022.findings-naacl)
Copied to clipboard
| Challenge: | Using MTG, we train and evaluate multilingual text generation models using human-annotated data. |
| Approach: | They propose a multilingual multiway text generation dataset with 400k human-annotated data that includes four generation tasks across five languages. |
| Outcome: | The proposed dataset includes four generation tasks across five languages (English, German, French, Spanish and Chinese) it provides comprehensive evaluations with diverse generation scenarios. |
From Observation to Understanding: Front-Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs (2025.findings-acl)
Copied to clipboard
Shenshen Li, Wenxin Meng, Lei Wang, Hao Yang, Chong Peng, Peng Yan, Fumin Shen, Jingkuan Song, Heng Tao Shen, Xing Xu
| Challenge: | Existing methods that adapt LVLMs to egocentric tasks overlook critical agent-environment interactions, limiting their ability to perform egoic reasoning. |
| Approach: | They propose a zero-shot paradigm to enhance egocentric reasoning by simulating human causal reasoning by formalizing ego-centric reasoning using a structural causal model. |
| Outcome: | The proposed method improves egocentric reasoning abilities on six tasks. |
Empathy in Diversity: Personalized Depression and Anxiety Therapy via Dialogue State Tracking and Patient-Aware Planning (2026.acl-long)
Copied to clipboard
Xinwei Yang, Junyi Fan, Yuqing Liu, Jiaxuan Wang, Jiashuai Zhang, Hongru Liang, Wenqiang Lei, Yao Song
| Challenge: | Recent efforts have turned to large language models (LLMs) as therapeutic agents for psychological therapy tasks, yet robustness across diverse patients remains underexplored. |
| Approach: | They propose a realistic role-play protocol for evaluating therapeutic dialogue agents and a de-identified, expert-annotated corpus of therapist–patient dialogues. |
| Outcome: | The proposed framework outperforms baselines on therapeutic outcomes and dialogue quality while improving conversational efficiency. |
Do you have the right scissors? Tailoring Pre-trained Language Models via Monte-Carlo Methods (2020.acl-main)
Copied to clipboard
| Challenge: | Pre-trained language models can be fine-tuned on task-specific datasets, but fine-timing can lead to over- and/or under-estimation problems. |
| Approach: | They propose a method to transfer probability mass from over-estimated regions to under-estimates by truncating and transferring probability mass between over- and under-estimating regions. |
| Outcome: | The proposed method outperforms the fine-tuning approach on a variety of datasets. |
TESTEVAL: Benchmarking Large Language Models for Test Case Generation (2025.findings-naacl)
Copied to clipboard
Wenhan Wang, Chenyuan Yang, Zhijie Wang, Yuheng Huang, Zhaoyang Chu, Da Song, Lingming Zhang, An Ran Chen, Lei Ma
| Challenge: | Existing methods to generate test cases using large language models are limited in their ability to generate unit test cases. |
| Approach: | They propose a test case generation benchmark that uses large language models to generate unit test cases. |
| Outcome: | The proposed test case generation benchmarks compare LLMs with commercial and open-source LLM platforms and find that they lack the ability to comprehend program logic and execution paths. |
When "Correct" Is Not Safe: Can We Trust Functionally Correct Patches Generated by Code Agents? (2026.acl-long)
Copied to clipboard
Yibo Peng, James Song, Lei Li, Xinyu Yang, Mihai Christodorescu, Ravi Mangal, Corina S. Pasareanu, Haizhong Zheng, Beidi Chen
| Challenge: | Code agents are increasingly trusted to autonomously fix bugs on platforms such as GitHub, yet their security evaluation focuses on functional correctness. |
| Approach: | They propose to attack functionally correct yet vulnerable (FCV) patches by combining multi-turn reasoning with tool invocation and environment interaction. |
| Outcome: | The proposed FCV-Attack achieves an attack success rate of 40.7% on GPT-5 Mini + OpenHands. |
INSTRUCTSCORE: Towards Explainable Text Generation Evaluation with Automatic Feedback (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to evaluate the quality of language generation do not provide explicit explanation of their verdicts. |
| Approach: | They propose a fine-grained explainable evaluation metric for text generation that harnesses human instruction and implicit knowledge of GPT-4 to fine-tune it. |
| Outcome: | The proposed model outperforms all other unsupervised metrics on translation, captioning, data-to-text, and commonsense generation tasks. |
Learning from Diverse Reasoning Paths with Routing and Collaboration (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies suggest that the reasoning abilities of large language models (LLMs) grows with model size and pre-training data. |
| Approach: | They propose to combine quality filtering, conditional routing, and cooperative peer teaching to transfer knowledge from powerful teacher models to compact and transparent students. |
| Outcome: | Experiments show that QR-Distill is superior to traditional methods. |
MARS: Unleashing the Power of Speculative Decoding via Margin-Aware Verification (2026.findings-acl)
Copied to clipboard
Jingwei Song, Xinyu Wang, Hanbin Wang, Xiaoxuan Lei, Tianyu Shi, Shixin Han, Eric Yang, Xiao-Wen Chang, Lynn Ai
| Challenge: | Autoregressive large language models suffer from high inference latency due to memorybandwidth constraints. |
| Approach: | They propose a method that decouples generation and verification by decoupling tokens and a lightweight draft model. |
| Outcome: | The proposed method delivers consistent and significant speedups over state-of-the-art baselines while preserving generation quality across diverse benchmarks. |
Smart-Searcher: Incentivizing the Dynamic Knowledge Acquisition of LLMs via Reinforcement Learning (2025.findings-emnlp)
Copied to clipboard
Huatong Song, Jinhao Jiang, Wenqing Tian, Zhipeng Chen, Yuhuan Wu, Jiahao Zhao, Yingqian Min, Xin Zhao, Lei Fang, Ji-Rong Wen
| Challenge: | Large Language Models (LLMs) are powerful but prone to hallucinations due to static knowledge. Retrieval-augmented generation (RAG) helps by injecting external information, but current methods are costly, generalize poorly, or ignore the model’s internal knowledge. |
| Approach: | They propose a framework to train large language models to leverage both internal and external knowledge sources. |
| Outcome: | The proposed framework outperforms existing methods and achieves efficient retrieval-augmented reasoning. |
XPrompt: Exploring the Extreme of Prompt Tuning (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prompt tuning learns soft prompts to condition pre-trained Language Models for performing downstream tasks in a parameter-efficient manner. |
| Approach: | They propose a Prompt tuning model with an eXtremely small scale that learns soft prompts to condition the frozen Pre-trained Language Models for performing downstream tasks in a parameter-efficient manner. |
| Outcome: | The proposed model outperforms the vanilla Prompt-Tuning and can significantly improve across tasks and model scales. |
Graph Neural Network Enhanced Retrieval for Question Answering of Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing retrieval methods divide reference documents into passages, treating them in isolation. Existing methods only use contiguous passages or keywords. |
| Approach: | They propose a retrieval method that leverages graph neural networks to exploit relatedness between passages to enhance retrieval. |
| Outcome: | The proposed method improves retrieval by exploiting the relatedness between passages. |
Act as you think: Reinforcing Consistent Reasoning in Medical Visual Question Answering (2026.acl-long)
Copied to clipboard
Songtao Jiang, Yuan Wang, Ruizhe Chen, Yan Zhang, Ruilin Luo, Bohan Lei, Yeying Jin, Sibo Song, ZhiBo Yang, Jimeng Sun, Jian Wu, Zuozhu Liu
| Challenge: | Recent advances have improved the accuracy of medical visual question answering (Med-VQA) however, the high stakes nature of the medical domain has precipitated a shift towards interpretability and transparency of reasoning processes. |
| Approach: | They propose a reinforcement learning from verifiable rewards framework that rewards internal consistency and logical coherence. |
| Outcome: | The proposed framework rewards internal consistency and logical coherence, and is highly versatile, the authors show. |
Calibrating Factual Knowledge in Pretrained Language Models (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies show that Pretrained Language Models can store factual knowledge, but facts stored in PLMs are not always correct. |
| Approach: | They propose a lightweight method to calibrate factual knowledge in PLMs without re-training from scratch. |
| Outcome: | The proposed method can be used to calibrate factual knowledge in PLMs without re-training from scratch. |
Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets (2026.acl-long)
Copied to clipboard
| Challenge: | Existing mitigation strategies focus on reactively addressing jailbreak incidents after safety guardrails have been compromised. |
| Approach: | They investigate the degradation of safety guardrails through the lens of representation similarity between upstream alignment datasets and downstream fine-tuning tasks. |
| Outcome: | The proposed model reduces harmfulness score by 10.33% when compared to baseline models. |
From Cross-Task Examples to In-Task Prompts: A Graph-Based Pseudo-Labeling Framework for In-context Learning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | In-context learning (ICL) enables large language models to perform novel tasks without parameter updates by conditioning on a few input-output examples. |
| Approach: | They propose a cost-efficient two-stage pipeline that reduces reliance on LLMs for data labeling. |
| Outcome: | The proposed pipeline reduces reliance on LLMs for data labeling . it leverages readily available cross-task examples to prompt an LLM and pseudo-label a small set of target task instances. |
An Efficient Conversational Smart Compose System (2023.acl-demo)
Copied to clipboard
Yun Zhu, Xiayu Chen, Lei Shu, Bowen Tan, Xinying Song, Lijuan Liu, Maria Wang, Jindong Chen, Ning Ruan
| Challenge: | a cloud-based smart compose system is designed to improve human-to-human conversation efficiency. |
| Approach: | They propose a cloud-based smart compose system to improve conversation efficiency . they propose heuristics to achieve the best trade-off between quality and latency . |
| Outcome: | The proposed system reduces latency without losing composing quality further. |
Hard Prompts Made Interpretable: Sparse Entropy Regularization for Prompt Tuning with RL (2024.acl-long)
Copied to clipboard
Yunseon Choi, Sangmin Bae, Seonghyun Ban, Minchan Jeong, Chuheng Zhang, Lei Song, Li Zhao, Jiang Bian, Kee-Eung Kim
| Challenge: | Prompt tuning is an important technique for directing model behaviors and eliciting desired responses. |
| Approach: | They propose to find optimal prompt tokens using soft Q-learning to optimize models for prompt tuning. |
| Outcome: | The proposed method improves on baseline prompt tuning, and the results are more natural and interpretable. |
Learning to Select In-Context Demonstration Preferred by Large Language Model (2025.findings-acl)
Copied to clipboard
| Challenge: | In-context learning (ICL) enables large language models to perform tasks with only a few examples as demonstrations. |
| Approach: | They propose a generative preference learning framework that leverages LLM feedback to directly optimize demonstration selection for ICL. |
| Outcome: | Experiments on 19 datasets across 11 task categories show that GenICL achieves superior performance than existing methods in selecting the most effective demonstrations. |
UNO Arena for Evaluating Sequential Decision-Making Capability of Large Language Models (2024.emnlp-main)
Copied to clipboard
Zhanyue Qin, Haochuan Wang, Deyuan Liu, Ziyang Song, Cunhang Fan, Zhao Lv, Jinlin Wu, Zhen Lei, Zhiying Tu, Dianhui Chu, Xiaoyan Yu, Dianbo Sui
| Challenge: | Existing LLMs demonstrate powerful capabilities between tasks, but can they make sequential decisions? |
| Approach: | They propose to evaluate sequential decision-making capability of large language models (LLMs) using novel metrics based Monte Carlo methods. |
| Outcome: | The proposed benchmark improves sequential decision-making performance compared to the vanilla LLM player. |
SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis (2025.findings-emnlp)
Copied to clipboard
Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Junjie Zhang, Fei Bai, Jia Deng, Xin Zhao, Zheng Liu, Lei Fang, Zhongyuan Wang, Ji-Rong Wen
| Challenge: | Existing approaches to deep search training lack high-quality training trajectories, prohibitive computational costs and lack of high-fidelity training data. |
| Approach: | They propose a framework that synthesizes high-quality training data by simulating real user interactions in live web search environments. |
| Outcome: | The proposed framework synthesizes high-quality training data by simulating user interactions in live web search environments. |
Making Pretrained Language Models Good Long-tailed Learners (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prompt-tuning has shown appealing performance in few-shot classification . however, it is less promising in long-tailed classification due to long tail . |
| Approach: | They propose to use prompt-tuning to make pretrained language models at least good long-tailed learners by bridging the gap between prompt- and commonly used finetun. |
| Outcome: | The proposed method makes pretrained language models at least good long-tailed learners, bridging the gap between prompt-tuning and finetunation. |
GazeVQA: A Video Question Answering Dataset for Multiview Eye-Gaze Task-Oriented Collaborations (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on the use of exocentric and egocentric videos in video question answering are focusing on eye-gaze information. |
| Approach: | They propose a task-oriented VQA dataset that captures eye-gaze information . they propose assisting models that ground the perceptual input into semantic information based on three different answer types . |
| Outcome: | The proposed model can ground the perceptual input into semantic information while reducing ambiguities. |
Polarity Calibration for Opinion Summarization (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing opinions summarization models emphasize the majority opinions while ignoring the minority opinions. |
| Approach: | They propose a method to align output summary and input text to achieve polarity calibration. |
| Outcome: | The proposed model can mitigate the polarity mismatch between output summary and input text, and maintain the content semantic and language quality. |
OMGM: Orchestrate Multiple Granularities and Modalities for Efficient Multimodal Retrieval (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for Knowledge-Based Visual Question Answering lack multimodal retrieval . large language models (LLMs) have demonstrated remarkable generalization and reasoning capabilities in text-based systems. |
| Approach: | They propose a multimodal vision-language retrieval-augmented generation system that harmonizes multiple modalities and modality to enhance retrieval. |
| Outcome: | The proposed system achieves state-of-the-art retrieval performance and competitive answers on InfoSeek and Encyclopedic-VQA benchmarks. |