Papers by Junjie Li

From Allies to Adversaries: Manipulating LLM Tool-Calling through Adversarial Injection (2025.naacl-long)

Challenge:	Current multimodal large language models (MLLMs) show limited understanding of dental images.
Approach:	They propose a dental-specialized multimodal large language model trained via staged multimodal alignment and reinforcement learning.
Outcome:	The proposed model outperforms state-of-the-art models on disease classification and dental VQA tasks.

Copied to clipboard

Rupeng Zhang, Haowei Wang, Junjie Wang, Mingyang Li, Yuekai Huang, Dandan Wang, Qing Wang

Repairing Catastrophic-Neglect in Text-to-Image Diffusion Models via Attention-Guided Feature Enhancement (2024.findings-emnlp)

Challenge:	Toolcalling has changed Large Language Model (LLM) applications by integrating external tools, but it also introduces new security vulnerabilities, particularly in the tool scheduling mechanisms of LLM, which have not been extensively studied.
Approach:	They propose a framework that exploits vulnerabilities in Large Language Models through adversarial tool injection to execute privacy theft, launch denial-of-service attacks, and manipulate business competition.
Outcome:	The proposed framework exploits vulnerabilities in LLM tool-calling systems through adversarial tool injection.

Copied to clipboard

Zhiyuan Chang, Mingyang Li, Junjie Wang, Yi Liu, Qing Wang, Yang Liu

AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments (2026.acl-long)

Challenge:	Text-to-Image Diffusion models generate high-quality images from textual descriptions, but they often produce images that do not fully align with the input prompts, resulting in semantic inconsistencies.
Approach:	They propose an automated repair approach to address catastrophic-neglect in T2I DMs.
Outcome:	The proposed model achieves 10.1%-16.3% higher Correct Rate in image generation compared to baselines.

Copied to clipboard

Zhiheng Xi, Dingwen Yang, Jiaqi Liu, Jixuan Huang, Honglin Guo, Baodai Huang, Tinggang Chen, Qi Zhang, Zhonghang Lu, Chenyu Liu, Jiajun Sun, Jiazheng Zhang, Dingwei Zhu, Xin Guo, Junzhe Wang, Zhihao Zhang, Yuming Yang, Junjie Ye, Minghe Gao, Dongrui Liu, Jiaming Ji, Guohao Li, Tao Gui, Qi Zhang, Xuanjing Huang

Generative Text-to-Image Retrieval via Hierarchical Identifiers and Semantic Internalization (2026.findings-acl)

Challenge:	Existing benchmarks evaluate agents in simplified, idealized settings, relying on pre-packaged tool interfaces, overlooking critical steps, and assume inputs are clean and fully specified.
Approach:	They propose a framework that evaluates language agents in simplified, idealized settings . they show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .
Outcome:	Experiments on 15 proprietary and open-source models show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .

Copied to clipboard

Jie Huang, Junjie Wang, Xin Liao, Ziyou Jiang, Wenshuo Wang, Shoubin Li, Qing Wang

RATE-Nav: Region-Aware Termination Enhancement for Zero-shot Object Navigation with Vision-Language Models (2025.findings-acl)

Challenge:	Existing text-to-image retrieval methods suffer from limited semantic discriminability, alignment bias, and closed-set restrictions.
Approach:	They propose a framework for semantic internalization for Generative Multimodal Alignment . they construct multi-granularity hierarchical identifiers to ensure unique, semantically consistent image representations .
Outcome:	The proposed framework outperforms state-of-the-art frameworks on Flickr30K and MS-COCO datasets . it achieves average Recall@1, Recall @5, and Recall_10 improvements of 10.65%, 8.50%, and 7.00% .

Copied to clipboard

Junjie Li, Nan Zhang, Xiaoyang Qu, Kai Lu, Guokuan Li, Jiguang Wan, Jianzong Wang

S2G-RAG: Structured Sufficiency and Gap Judging for Iterative Retrieval-Augmented QA (2026.acl-long)

Challenge:	Object navigation is a fundamental task in embodied artificial intelligence.
Approach:	They propose a region-aware Termination-Enhanced method that incorporates visual language models and exploration rates to enable efficient termination.
Outcome:	The proposed method achieves a success rate of 67.8% and an SPL of 31.3% on the HM3D dataset.

Copied to clipboard

Minghan Li, Junjie Zou, Xinxuan Lv, Chao Zhang, Guodong Zhou

Question Answering

Challenge:	Retrieval-augmented generation grounds language models in external evidence, but multi-hop question answering remains difficult . iterative pipelines must control what to retrieve next and when evidence is adequate.
Approach:	They propose an iterative framework with an explicit controller, S2G-Judge . they map structured gap items into the next retrieval query to produce stable retrieval trajectories .
Outcome:	Experiments on TriviaQA, HotpotQA, and 2WikiMultiHopQA show that S2G-RAG improves multi-hop QA performance and robustness under multi-turn retrieval.

Document-level Multi-aspect Sentiment Classification by Jointly Modeling Users, Aspects, and Overall Ratings (C18-1)

Copied to clipboard

Junjie Li, Haitong Yang, Chengqing Zong

Information Extraction & Text Mining

Challenge:	Existing approaches focus on text information, but authors and overall ratings are ignored, both of which are proved to be significant on interpreting the sentiments of different aspects.
Approach:	They propose a hierarchical user-aspect rating network model to consider user preference and overall ratings jointly.
Outcome:	The proposed model can predict aspects of a product in two real-world datasets.

Agentic Episodic Control (2026.findings-acl)

Copied to clipboard

Xidong Yang, Wenhao Li, Junjie Sheng, Yun Hua, Haosheng Chen, Chuyun Shen, Xiangfeng Wang

Green & Sustainable NLP

Challenge:	Existing methods for reinforcement learning (RL) are limited by poor data efficiency and weak generalization.
Approach:	They propose a novel architecture that integrates large language models into episodic RL.
Outcome:	The proposed architecture achieves 2–6 higher data efficiency than baselines and is the only method to solve complex tasks like UnlockLocal with over 90% success.

Play Guessing Game with LLM: Indirect Jailbreak Attack with Implicit Clues (2024.findings-acl)

Copied to clipboard

Zhiyuan Chang, Mingyang Li, Yi Liu, Junjie Wang, Qing Wang, Yang Liu

Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric (2025.acl-long)

Challenge:	Existing jailbreak attacks primarily utilize scenario camouflage techniques, however their explicit mention of malicious intent will be easily recognized and defended by LLMs.
Approach:	They propose an indirect jailbreak attack approach, Puzzler, which can bypass LLM’s defensive strategies and obtain malicious response by implicitly providing LLMs with some clues about the original malicious query.
Outcome:	The proposed approach can bypass the LLM’s defensive strategies and obtain malicious response by implicitly providing LLMs with some clues about the original malicious query.

Copied to clipboard

Yuming Yang, Yang Nan, Junjie Ye, Shihan Dou, Xiao Wang, Shuo Li, Huijie Lv, Tao Gui, Qi Zhang, Xuanjing Huang

Generative Cross-Domain Data Augmentation for Aspect and Opinion Co-Extraction (2022.naacl-main)

Challenge:	Existing studies have explored various diversity-aware data selection methods to construct high-quality datasets and enhance model performance.
Approach:	They propose to use data diversity to measure instruction tuning of large language models.
Outcome:	The proposed diversity metric outperforms existing methods on simulated and real-world data and shows that it captures diversity variations and achieves a 0.97 correlation with instruction tuning.

Copied to clipboard

Junjie Li, Jianfei Yu, Rui Xia

Low-Resource NLP

Challenge:	Existing approaches to perform aspect and opinion co-extraction are difficult due to the lack of fine-grained annotations.
Approach:	They propose a framework to transfer knowledge from a labeled source domain to an unlabeled target domain.
Outcome:	The proposed framework is more effective than previous domain adaptation methods on three datasets.

PHMOSpell: Phonological and Morphological Knowledge Guided Chinese Spelling Check (2021.acl-long)

Copied to clipboard

Li Huang, Junjie Li, Weiwei Jiang, Zhiyu Zhang, Minchuan Chen, Shaojun Wang, Jing Xiao

Syntactic Text Processing

Challenge:	False gram and phonological errors make Chinese spelling check difficult . a novel end-to-end trainable model outperforms existing methods .
Approach:	They propose a trainable Chinese spelling check model that integrates phonological and visual information into a pre-trained language model.
Outcome:	The proposed model outperforms existing state-of-the-art models on three benchmarks.

Mimicking the Familiar: Dynamic Command Generation for Information Theft Attacks in LLM Tool-Learning System (2025.acl-long)

Copied to clipboard

Ziyou Jiang, Mingyang Li, Guowei Yang, Junjie Wang, Yuekai Huang, Zhiyuan Chang, Qing Wang

Data Augmentation using LLMs: Data Perspectives, Learning Paradigms and Challenges (2024.findings-acl)

Challenge:	Existing approaches to attack Large Language Model (LLM) tool-learning systems are black-box oriented and rely on static commands that cannot adapt flexibly to the changes in user queries and the invocation chain.
Approach:	They propose a dynamic attack comment generation approach for information theft attacks in LLM tool-learning systems that mimics the familiar by inferring the information utilized by upstream tools.
Outcome:	The proposed approach outperforms baselines with +13.2% ASRTheft and can be generalized to new tool-learning systems to expose their information leakage risks.

Copied to clipboard

Bosheng Ding, Chengwei Qin, Ruochen Zhao, Tianze Luo, Xinze Li, Guizhen Chen, Wenhan Xia, Junjie Hu, Anh Tuan Luu, Shafiq Joty

FocusLLM: Precise Understanding of Long Context by Dynamic Condensing (2025.acl-long)

Challenge:	Data augmentation (DA) is a key technique for enhancing model performance by diversifying training examples without the need for additional data collection.
Approach:	They examine various strategies that utilize LLMs for data augmentation, including a novel exploration of learning paradigms where LLM-generated data is used for diverse forms of further training.
Outcome:	The proposed approach addresses the primary open challenges faced by LLMs in the field of large language models and aims to serve as a comprehensive guide for researchers and practitioners.

Copied to clipboard

Zhenyu Li, Yike Zhang, Tengyu Pan, Yutao Sun, Zhichao Duan, Junjie Fang, Rong Han, Zixuan Wang, Jianyong Wang

Challenge:	Existing context condensing methods cannot accurately understand the full context, as there is a considerable amount of information loss in the condensed process.
Approach:	They propose a framework to extend the fixed context length of any decoder-only LLM by distilling crucial information from long sequences.
Outcome:	The proposed framework extends the fixed context length of any decoder-only LLM, allowing it to focus on relevant information from very long sequences.

All Information is Valuable: Question Matching over Full Information Transmission Network (2022.findings-naacl)

Copied to clipboard

Le Qi, Yu Zhang, Qingyu Yin, Guidong Zheng, Wen Junjie, Jinlong Li, Ting Liu

Reasoning

Challenge:	Existing methods for question matching only transmit one kind of information while failing to utilize both kinds of information simultaneously.
Approach:	They propose a question matching network that can transmit both representation and interactive information together in a simultaneous fashion.
Outcome:	The proposed approach outperforms strong baseline models on two standard benchmarks.

CalibraEval: Calibrating Prediction Distribution to Mitigate Selection Bias in LLMs-as-Judges (2025.acl-long)

Copied to clipboard

Haitao Li, Junjie Chen, Qingyao Ai, Zhumin Chu, Yujia Zhou, Qian Dong, Yiqun Liu

ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages (2024.acl-long)

Challenge:	Empirical evaluations of large language models demonstrate that they improve performance in a wide range of tasks.
Approach:	They propose a label-free method for mitigating selection bias during inference by reformulating debiasing as an optimization task.
Outcome:	The proposed method mitigates selection bias and improves performance compared to existing methods.

Copied to clipboard

Junjie Ye, Sixian Li, Guanyu Li, Caishuang Huang, Songyang Gao, Yilong Wu, Qi Zhang, Tao Gui, Xuanjing Huang

Semantic Text Processing

Challenge:	Existing research focuses on enhancing LLMs capabilities through tool utilization.
Approach:	They propose a framework to investigate safety issues in large language models in tool learning . they propose malicious queries and jailbreak attacks in the input stage .
Outcome:	The proposed framework investigates six safety scenarios for LLMs in tool learning . the data will be released upon acceptance of the proposed framework .

The Stochastic Parrot on LLM’s Shoulder: A Summative Assessment of Physical Concept Understanding (2025.naacl-long)

Copied to clipboard

Mo Yu, Lemao Liu, Junjie Wu, Tsz Ting Chung, Shunchi Zhang, Jiangnan Li, Dit-Yan Yeung, Jie Zhou

Challenge:	Recent years have witnessed remarkable advancements in large language models (LLMs) many researchers argue that LLMs may not * Equal contribution.
Approach:	They propose a task that summarises the memorization issue by using grid inputs that abstractly describe physical phenomena.
Outcome:	The proposed task alleviates the memorization issue by using grid-format inputs that abstractly describe physical phenomena.

LLM can Achieve Self-Regulation via Hyperparameter Aware Generation (2024.findings-acl)

Copied to clipboard

Siyin Wang, Shimin Li, Tianxiang Sun, Jinlan Fu, Qinyuan Cheng, Jiasheng Ye, Junjie Ye, Xipeng Qiu, Xuanjing Huang

Text Generation

Challenge:	Existing decoding strategies and hyperparameters may not be optimal for each sample.
Approach:	They propose a model that auto-regulates decoding strategies and hyperparameters . this approach eliminates the need for extensive manual tuning, they argue .
Outcome:	The proposed model eliminates the need for extensive manual tuning, offering a more autonomous, self-regulate model behavior.

Lost in Stories: Consistency Bugs in Long Story Generation by LLMs (2026.findings-acl)

Copied to clipboard

Junjie Li, Xinrui Guo, Yuhao Wu, Roy Ka-Wei Lee, Hongzhi Li, Yutao Xie

Text Generation

Challenge:	Existing story generation benchmarks focus mainly on plot quality and fluency, leaving consistency errors unexplored.
Approach:	They propose a benchmark to evaluate narrative consistency in long-form story generation.
Outcome:	Evaluating LLMs, we find consistency errors are common in factual and temporal dimensions . authors say the findings can inform future efforts to improve consistency in long-form narrative generation.

LegalAgentBench: Evaluating LLM Agents in Legal Domain (2025.acl-long)

Copied to clipboard

Haitao Li, Junjie Chen, Jingli Yang, Qingyao Ai, Wei Jia, Youfeng Liu, Kai Lin, Yueyue Wu, Guozhi Yuan, Yiran Hu, Wuyue Wang, Yiqun Liu, Minlie Huang

Challenge:	Existing general-domain benchmarks do not capture complexity of real-world judicial cognition and decision-making.
Approach:	They propose a benchmark specifically designed to evaluate LLM Agents in the legal domain.
Outcome:	The proposed benchmark includes 17 corpora from real-world legal scenarios and provides 37 tools for interacting with external knowledge.

TL-Training: A Task-Feature-Based Framework for Training Large Language Models in Tool Use (2025.findings-emnlp)

Copied to clipboard

Junjie Ye, Yilong Wu, Sixian Li, Yuming Yang, Zhiheng Xi, Tao Gui, Qi Zhang, Xuanjing Huang, Peng Wang, Zhongchao Shi, Jianping Fan, Zhengyin Du

ToolEyes: Fine-Grained Evaluation for Tool Learning Capabilities of Large Language Models in Real-world Scenarios (2025.coling-main)

Challenge:	a new approach to training large language models (LLMs) overlooks task-specific characteristics in tool use, leading to performance bottlenecks.
Approach:	They propose a task-feature-based framework that mitigates the effects of suboptimal training data . they use a dataset to train large-scale LLMs and a reward mechanism tailored to error categories .
Outcome:	The proposed framework matches or surpasses open- and closed-source LLMs in tool-use performance using only 1,217 training data points.

Copied to clipboard

Junjie Ye, Guanyu Li, SongYang Gao, Caishuang Huang, Yilong Wu, Sixian Li, Xiaoran Fan, Shihan Dou, Tao Ji, Qi Zhang, Tao Gui, Xuanjing Huang

One Shot Dominance: Knowledge Poisoning Attack on Retrieval-Augmented Generation Systems (2025.findings-emnlp)

Challenge:	Existing evaluations of tool learning focus on validation of tools for large language models with expected outcomes, but this focus ignores the complex capabilities required for LLMs to effectively use tools.
Approach:	They propose a fine-grained system for evaluation of large language models’ tool learning capabilities in authentic scenarios.
Outcome:	The proposed system examines seven real-world scenarios, analyzing five dimensions crucial to LLMs in tool learning: format alignment, intent comprehension, behavior planning, tool selection, and answer organization.

Copied to clipboard

Zhiyuan Chang, Mingyang Li, Xiaojun Jia, Junjie Wang, Yuekai Huang, Ziyou Jiang, Yang Liu, Qing Wang

Supervised Optimism Correction: Be Confident When LLMs Are Sure (2025.findings-acl)

Challenge:	Existing knowledge poisoning attacks against RAG systems require multiple poisoned documents or can only function effectively on simplistic queries.
Approach:	They propose a more realistic knowledge poisoning attack that poisons only a single document while remaining effective for complex multi-hop questions involving complex relationships between multiple elements.
Outcome:	The proposed attack achieves success by poisoning only a single document while remaining effective for complex multi-hop questions involving complex relationships between multiple elements.

Copied to clipboard

Junjie Zhang, Rushuai Yang, Shunyu Liu, Ting-En Lin, Fei Huang, Yi Chen, Yongbin Li, Dacheng Tao

Reasoning

Challenge:	Recent advances in Large Language Models (LLMs) have demonstrated remarkable success across diverse tasks such as instruction following, code generation, and medical diagnosis.
Approach:	They propose a supervised fine-tuning-based auxiliary loss for Q-value estimations during supervised refinement.
Outcome:	The proposed method outperforms beam search on GSM8K, MATH, and GAOKAO on reasoning benchmarks.

RoTBench: A Multi-Level Benchmark for Evaluating the Robustness of Large Language Models in Tool Learning (2024.emnlp-main)

Copied to clipboard

Junjie Ye, Yilong Wu, Songyang Gao, Caishuang Huang, Sixian Li, Guanyu Li, Xiaoran Fan, Qi Zhang, Tao Gui, Xuanjing Huang

All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction (2026.acl-long)

Challenge:	Current research emphasizes LLMs’ capacity to utilize tools in well-structured environments while overlooking their stability when confronted with the inevitable noise of the real world.
Approach:	They propose a multi-level benchmark to evaluate the robustness of large language models in tool learning by establishing five external environments with varying levels of noise.
Outcome:	The proposed model outperforms the GPT-4 model in tool learning in three critical phases: tool selection, parameter identification, and content filling.

Copied to clipboard

Ziyou Jiang, Mingyang Li, Junjie Wang, Yuekai Huang, Jie Huang, Zhiyuan Chang, Zhaoyang Li, Qing Wang

Know Thy Enemy: Securing LLMs Against Prompt Injection via Diverse Data Synthesis and Instruction-Level Chain-of-Thought Learning (2026.findings-acl)

Challenge:	Existing methods for harmful meme detection only learn the combination of harmful elements and lack understanding of these implicit expressions.
Approach:	They propose a method that detects harmful memes by replicating the design concept of malicious users.
Outcome:	The proposed method achieves the highest accuracy with 81.1% and has slight accuracy decreases when generalized to type-shifting and temporal-evolving memes.

Copied to clipboard

Zhiyuan Chang, Mingyang Li, Yuekai Huang, Ziyou Jiang, Xiaojun Jia, Qian Xiong, Junjie Wang, Zhaoyang Li, Qing Wang

Entity Relation Extraction as Dependency Parsing in Visually Rich Documents (2021.emnlp-main)

Challenge:	Large language model (LLM)-integrated applications face security vulnerabilities from prompt injection (PI) attacks.
Approach:	They propose a model enhancement method that synthesizes diverse training data and employs instruction-level chain-of-thought fine-tuning to enable LLMs to effectively identify and reject malicious instructions regardless of their source or position in the context.
Outcome:	The proposed method outperforms baselines in three critical dimensions while maintaining utility performance without degradation.

Copied to clipboard

Yue Zhang, Zhang Bo, Rui Wang, Junjie Cao, Chen Li, Zuyi Bao

Syntactic Text Processing

Challenge:	Existing studies on key information extraction from visually rich documents focus on labeling the text within bounding boxes, while relations between words are unexplored.
Approach:	They propose to use a dependency parsing model to extract semantic entities from visually rich documents by combining entity labeling and relation extraction tasks.
Outcome:	The proposed model achieves 65.96% F1 score on the FUNSD dataset.

Where Did It Go Wrong? Capability-Oriented Failure Attribution for Vision-and-Language Navigation Agents (2026.findings-acl)

Copied to clipboard

Jianming Chen, Yawen Wang, Junjie Wang, Xiaofei Xie, Shoubin Li, Qing Wang, Fanjiang Xu

Probing Social Identity Bias in Chinese LLMs with Gendered Pronouns and Social Groups (2026.findings-acl)

Challenge:	Existing testing methods are system-level and provide limited insight into which capability deficiencies cause task failures.
Approach:	They propose a capability-oriented testing approach that enables failure detection and attribution by seed selection and mutation.
Outcome:	The proposed method detects more failure cases and pinpoints capability-level deficiencies than state-of-the-art baselines, providing more interpretable and actionable guidance for improving embodied agents.

Copied to clipboard

Geng Liu, Li Feng, Junjie Mu, Mengxiao Zhu, Francesco Pierri

SAGE: Synergistic Adaptive Gating of Experts for Hateful Video Detection (2026.acl-long)

Challenge:	Large language models (LLMs) are increasingly deployed in user-facing applications, raising concerns that they reflect and amplify social biases.
Approach:	They propose a Mandarin-specific evaluation framework to examine social identity biases in Chinese LLMs using Mandarin-based prompts.
Outcome:	The proposed framework compares ingroup (“We”) and outgroup (“They”) framings across 240 social groups salient in the Chinese context, using a two-tiered measurement framework that assesses both sentiment and toxicity.

Copied to clipboard

Jie Huang, Xin Liao, Junjie Wang, Mingyang Li, Wenshuo Wang, Ziyou Jiang, Shoubin Li, Qing Wang

Visual Data in NLP

Challenge:	Existing methods for hateful video detection rely on multimodal feature fusion . existing methods rely only on blind feature mixing, which leads to feature dilution .
Approach:	They propose a framework that shifts from blind feature mixing to decision-level arbitration . it instantiates disentangled experts to rigorously preserve modality-specific semantics .
Outcome:	The proposed framework outperforms state-of-the-art methods on HateMM and MultiHateClip benchmarks.

Attribute-aware Sequence Network for Review Summarization (D19-1)

Copied to clipboard

Junjie Li, Xuepeng Wang, Dawei Yin, Chengqing Zong

Summarization

Challenge:	Existing review summarization systems generate summary only based on review content and neglect the authors’ attributes (e.g., gender, age, and occupation).
Approach:	They propose an Attribute-aware Sequence Network (ASN) to take the aforementioned users’ characteristics into account by encoding their attributes over the words.
Outcome:	The proposed model outperforms existing systems on tripAtt and human evaluation by taking the authors' attributes into account and incorporating attribute embedding and word-using habits into word prediction.

Is Fine-tuning Needed? Pre-trained Language Models Are Near Perfect for Out-of-Domain Detection (2023.acl-long)

Copied to clipboard

Rheeya Uppaal, Junjie Hu, Yixuan Li

Semantic Text Processing

Challenge:	Out-of-distribution (OOD) detection is critical for reliable predictions over text . fine-tuning with pre-trained language models has been a de facto procedure .
Approach:	They propose to leverage pre-trained language models for OOD detection without fine-tuning on ID data.
Outcome:	The proposed approach outperforms the fine-tuned model under distributional shifts.

Towards Scalable Lightweight GUI Agents via Multi-role Orchestration (2026.findings-acl)

Copied to clipboard

Ziwei Wang, Junjie Zheng, Leyang Yang, Sheng Zhou, Xiaoxuan Tang, Fang Zhouhua, Zhiwei Liu, Dajun Chen, Yong Li, Jiajun Bu

Challenge:	Advanced GUI agents suffer from prohibitive deployment costs on resource-constrained devices.
Approach:	They propose a lightweight GUI agent with GUI-specific knowledge and task scalability . LAMO-3B supports monolithic execution and MAS-style orchestration .
Outcome:	The proposed GUI agent LAMO-3B supports monolithic execution and MAS-style orchestration.

Analyzing the Effects of Supervised Fine-Tuning on Model Knowledge from Token and Parameter Levels (2025.emnlp-main)

Copied to clipboard

Junjie Ye, Yuming Yang, Yang Nan, Shuo Li, Qi Zhang, Tao Gui, Xuanjing Huang, Peng Wang, Zhongchao Shi, Jianping Fan

Situated Embedding Models for Context-Aware Dense Retrieval (2026.acl-short)

Challenge:	Large language models (LLMs) acquire substantial world knowledge during pretraining, which is further shaped by post-training techniques such as supervised fine-tuning (SFT).
Approach:	They evaluate closed-book question answering (CBQA) performance across five LLMs from the LLaMA-2 and LLama-3 families and examine the impact of supervised fine-tuning on model knowledge.
Outcome:	The proposed model performance is 14% worse than models fine-tuned on 1,920 samples and 12% worse on 240 samples.

Copied to clipboard

Junjie Wu, Jiangnan Li, Yuqing Li, Lemao Liu, Liyan Xu, Jiwei Li, Dit-Yan Yeung, Jie Zhou, Mo Yu

Representation Learning

Challenge:	Existing embedding models are not well-equipped to encode situated context effectively, i.e., situating a chunk’s meaning within its context.
Approach:	They propose to represent short chunks in a way that is conditioned on a broader context window to enhance retrieval performance.
Outcome:	The proposed model outperforms state-of-the-art embedding models on a book-plot retrieval dataset.

A.S.E: A Repository-Level Benchmark for Evaluating Security in AI-Generated Code (2026.findings-acl)

Copied to clipboard

Keke Lian, Wang Bin, Lei Zhang, Libo Chen, Junjie Wang, Ziming Zhao, Yujiu Yang, Miaoqian Lin, Haotong Duan, Haoran Zhao, Shuang Liao, Mingda Guo, Quan Jiazheng, Yilu Zhong, Chenhao He, Chen Zichuan, Jie Wu, Haoling Li, Zhaoxuan Li, Jiongchi Yu, Hui LI, Dong Zhang

Programming Languages in NLP

Challenge:	Existing security evaluation benchmarks lack relevance to real-world AI programming tasks . current LLMs struggle with secure coding, research shows .
Approach:	They propose a repository-level evaluation benchmark to assess security of AI-generated code.
Outcome:	The proposed framework mirrors real-world AI programming tasks and offers valuable insights into the state of AI code generation.

Invisible Prompts, Visible Threats: Malicious Font Injection in External Resources for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Junjie Xiong, Changjia Zhu, Shuhang Lin, Chong Zhang, Yongfeng Zhang, Yao Liu, Lingyao Li