Papers by Zhi Wang
CODERL+: Improving Code Generation via Reinforcement with Execution Semantics Alignment (2026.acl-long)
Copied to clipboard
Xue Jiang, Yihong Dong, Mengyang Liu, Deng Hongyi, Tian Wang, Yongding Tao, Zhi Jin, Wenpin Jiao, Ge Li
| Challenge: | Large Language Models excel at code generation by learning from vast code corpora, but a fundamental semantic gap remains between training on textual patterns and the goal of functional correctness . reinforcement learning with verifiable rewards (RLVR) approaches are inefficient for establishing a well-aligned connection between the textual representation of code and its execution semantics. |
| Approach: | They propose a novel approach that integrates execution semantics alignment into the RLVR training pipeline for code generation. |
| Outcome: | The proposed model outperforms baseline training and RLVR and shows strong applicability across RL and LLMs. |
Unsupervised Sentence Representation Learning with Syntactically Aligned Negative Samples (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing approaches to sentence representation learning often encounter semantic inconsistencies and feature suppression. |
| Approach: | They propose a method for generating syntactically aligned negative (SAN) samples using a semantic importance-aware Masked Language Model (MLM) approach. |
| Outcome: | The proposed method produces negative samples with substantial textual overlap with the original sentences while conveying different meanings. |
Learning from Contrasts: Synthesizing Reasoning Paths from Diverse Search Trajectories (2026.acl-long)
Copied to clipboard
| Challenge: | MCTS methods retain only the single highest-reward trajectory, discarding comparative signals present in the many explored paths. |
| Approach: | They propose a framework that transforms supervision extraction into a synthesis procedure. |
| Outcome: | The proposed framework matches or exceeds baselines on 60K CRPS-synthesized examples on out-of-domain benchmarks. |
I-AM-G: Interest Augmented Multimodal Generator for Item Personalization (2024.emnlp-main)
Copied to clipboard
| Challenge: | e-commerce and recommender systems lack a framework for personalized generation . a new framework extracts tags from multimodal information of items that the user has interacted with . |
| Approach: | They propose a framework that extracts tags from multimodal information and rewrites item description . they then use a decoupled text-to-text and image-to image retriever to search for similar item text . |
| Outcome: | The proposed framework can generate results aligned with user preferences . it can be used in e-commerce and recommender systems to win over diverse user base . |
DevEval: A Manually-Annotated Code Generation Benchmark Aligned with Real-World Code Repositories (2024.findings-acl)
Copied to clipboard
Jia Li, Ge Li, Yunfei Zhao, Yongmin Li, Huanyu Liu, Hao Zhu, Lecheng Wang, Kaibo Liu, Zheng Fang, Lanshen Wang, Jiazheng Ding, Xuanming Zhang, Yuqi Zhu, Yihong Dong, Zhi Jin, Binhua Li, Fei Huang, Yongbin Li, Bin Gu, Mengfei Yang
| Challenge: | Existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of Large Language Models (LLMs). |
| Approach: | They propose a repository-level benchmark named DevEval to evaluate LLMs' coding abilities in real-world code repositories. |
| Outcome: | The proposed benchmarks show that the LLMs perform better in real-world code repositories than existing benchmarks. |
SPARD: Self-Paced Curriculum for RL Alignment via Integrating Reward Dynamics and Data Utility (2026.acl-long)
Copied to clipboard
Xuyang Zhi, Peilun Zhou, Chengqiang Lu, Hang Lv, Yiwei Liang, Rongyang Zhang, Yan Gao, null Yiwu, Yao Hu, Hongchao Gu, Defu Lian, Hao Wang, Enhong Chen
| Challenge: | Large Language Models (LLMs) are shifting the focus from single verifiable tasks toward complex, open-ended real-world scenarios. |
| Approach: | They propose a framework that automatically adjusts reward weights and data importance to synchronize learning intent with data utility for optimal performance. |
| Outcome: | The proposed framework improves model capabilities across all domains and scales. |
Granular Entity Mapper: Advancing Fine-grained Multimodal Named Entity Recognition and Grounding (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for fine-grained content extraction are limited by long-tailed distribution of textual entity categories and performance of object detectors. |
| Approach: | They propose a multi-granularity entity recognition module and a reranking module to integrate hierarchical information of entity categories, visual cues, and external textual resources collectively. |
| Outcome: | The proposed framework achieves state-of-the-art on the fine-grained content extraction task. |
UnClE: Explicitly Leveraging Semantic Similarity to Reduce the Parameters of Word Embeddings (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to reduce word embedding parameters ignore semantic information . existing methods do not consider semantic information, allowing for performance degradation . |
| Approach: | They propose a method that leverages semantic similarity with weight sharing to reduce dimensionality of word embeddings. |
| Outcome: | The proposed method reduces word embedding parameters by more than 11x on a standard English-German dataset. |
ExploraCoder: Advancing Code Generation for Multiple Unseen APIs via Planning and Chained Exploration (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models face intrinsic limitations in coding with unseen APIs in training corpora. |
| Approach: | They propose a training-free framework that empowers LLMs to invoke multiple unseen APIs in code solution by planning a complex problem into several API invocation subtasks and experimenting with correct API usage at intermediate steps. |
| Outcome: | The proposed framework significantly improves performance for models lacking prior API knowledge, achieving 11.99% over retrieval-based approaches and 17.28% over pretraining-based methods in pass@10. |
Time-for-Accuracy: Formalizing Chain-of-Thought as an Expansion of Logical Depth (2026.findings-acl)
Copied to clipboard
| Challenge: | Chain-of-thought (CoT) prompting can improve multi-step reasoning, but it is unclear what kind of additional sequential computation longer traces actually enable. |
| Approach: | They propose a deletion-based measure of step necessity under a specified inference interface to operationalize realized depth beyond raw length. |
| Outcome: | The proposed method combines effective logical depth with Bennett's logical depth to show that it is more efficient than a linear model. |
From What Is Said to Why It Is Framed: Intent-Aware News Video Understanding (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing verification methods for short-form news videos neglect communicative intent . stylistic presentation and factual manipulation are often intertwined, resulting in shortcut learning . |
| Approach: | They propose a theory-grounded representation of communicative intent that captures creator stance, audience need activation, and communication strategy. |
| Outcome: | The proposed framework captures creator stance, audience need activation, and communication strategy. |
Focus-Constrained Attention Mechanism for CVAE-based Response Generation (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models generate high-frequency but trivial responses such as "I don't know" or "I'm ok" due to the discrepancy in discourse-level information, standard models generate one-to-many relationships. |
| Approach: | They propose to transform coarse-grained discourse-level information into fine-grounded word-level knowledge by introducing a fine-grain focus signal and a focus-constrained attention mechanism to take full advantage of focus. |
| Outcome: | The proposed model can generate more diverse and informative responses compared with state-of-the-art models. |
On Safety Risks in Experience-Driven Self-Evolving Agents (2026.findings-acl)
Copied to clipboard
Weixiang Zhao, Yichen Zhang, Yingshuo Wang, Yang Deng, Yanyan Zhao, Xuda Zhi, Yongbo Huang, Hao He, Wanxiang Che, Bing Qin, Ting Liu
| Challenge: | Experience-driven self-evolution has emerged as a promising paradigm for improving the autonomy of large language model agents, yet its reliance on self-curated experience introduces underexplored safety risks. |
| Approach: | They investigate how experience accumulation and utilization in self-evolving agents affect safety performance across web-based and embodied environments. |
| Outcome: | The findings expose inherent limitations of current self-evolving agents and call for more principled strategies to ensure safe and reliable adaptation. |
SelfAug: Mitigating Catastrophic Forgetting in Retrieval-Augmented Generation via Distribution Self-Alignment (2025.findings-emnlp)
Copied to clipboard
Yuqing Huang, Rongyang Zhang, Qimeng Wang, Chengqiang Lu, Yan Gao, null Yiwu, Yao Hu, Xuyang Zhi, Guiquan Liu, Xin Li, Hao Wang, Enhong Chen
| Challenge: | Existing solutions for supervised fine-tuning often lead to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities. |
| Approach: | They propose a self-distribution alignment method that aligns input sequence logits to preserve the model’s semantic distribution, thereby mitigating catastrophic forgetting and improving downstream performance. |
| Outcome: | The proposed method achieves a superior balance between downstream learning and general capability retention. |
Registering Source Tokens to Target Language Spaces in Multilingual Neural Machine Translation (2025.acl-long)
Copied to clipboard
| Challenge: | Multilingual neural machine translation (MNMT) aims for arbitrary translations across multiple languages. |
| Approach: | They propose a method that inserts a set of tokens specifying the target language into the input sequence between the source and target tokens. |
| Outcome: | The proposed method outperforms existing models on a large-scale benchmark. |
NeuroSym-Cal: Bridging the Reasoning-Execution Gap in Code Generation via Hierarchical Calibration (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing calibration methods rely on the assumption that consensus implies correctness . Existing methods fail under systematic errors, leading to miscalibrated high-confidence predictions. |
| Approach: | They propose a hierarchical calibration framework that measures confidence at two levels . they propose sensitivity analysis to measure local curvature of deductive process . |
| Outcome: | The proposed framework de-saturates overconfident errors and improves selective generation performance on OOD benchmarks. |
PATIENT-𝜓: Using Large Language Models to Simulate Patients for Training Mental Health Professionals (2024.emnlp-main)
Copied to clipboard
Ruiyi Wang, Stephanie Milani, Jamie Chiu, Jiayin Zhi, Shaun Eack, Travis Labrum, Samuel Murphy, Nev Jones, Kate Hardy, Hong Shen, Fei Fang, Zhiyu Chen
| Challenge: | Mental illness remains one of the most critical public health issues. |
| Approach: | They propose a patient simulation framework for cognitive behavior therapy training that uses large language models to act as a simulated therapy patient. |
| Outcome: | The proposed framework improves the skill acquisition and confidence of mental health trainees beyond textbooks, videos, and role-play with non-patients. |
AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models (2025.findings-acl)
Copied to clipboard
Qi Liu, Jingqing Ruan, Hao Li, Haodong Zhao, Desheng Wang, Jiansong Chen, Wan Guanglu, Xunliang Cai, Zhi Zheng, Tong Xu
| Challenge: | Existing multi-objective preference alignment methods for large language models face limitations such as auxiliary reward/reference models and computational complexity. |
| Approach: | They propose a framework that achieves dynamic balance across preference dimensions by using dimension-aware generation metrics as implicit rewards. |
| Outcome: | Empirical results show that AMoPO outperforms state-of-the-art methods by 28.5% . |
Double-Checker: Large Language Model as a Checker for Few-shot Named Entity Recognition (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have demonstrated remarkable performance on few-shot Named Entity Recognition tasks due to the high cost of obtaining high-quality labeled data. |
| Approach: | They propose to decompose the task into entity span detection and entity type classification using a type-independent entity span detector and then classify the detected spans based on their types. |
| Outcome: | The proposed method consistently yields improvements over two baseline approaches. |
MLeVLM: Improve Multi-level Progressive Capabilities based on Multimodal Large Language Model for Medical Visual Question Answering (2024.findings-acl)
Copied to clipboard
Dexuan Xu, Yanyuan Chen, Jieyi Wang, Yue Huang, Hanpin Wang, Zhi Jin, Hongxing Wang, Weihua Yue, Jing He, Hang Li, Yu Huang
| Challenge: | Existing MVQA models ignore multi-level progressive capabilities due to unspecific data and plain architecture. |
| Approach: | They propose a multi-level visual language model for medical visual question answering (MVQA) which covers multi- level questions and answers as well as reasoning processes from visual clues to semantic cognition. |
| Outcome: | The proposed model outperforms existing medical multimodal large language models on a multi-level instruction dataset and a feature alignment module. |
LLMs can Perform Multi-Dimensional Analytic Writing Assessments: A Case Study of L2 Graduate-Level Academic English Writing (2025.acl-long)
Copied to clipboard
| Challenge: | a growing number of studies have indicated the general usefulness of LLMs for automated writing assessments. |
| Approach: | They propose a framework that evaluates LLMs' ability to provide scores and comments based on multiple assessment criteria. |
| Outcome: | The proposed framework is interpretable, cost-efficient, scalable, and reproducible . it is compared to existing methods that rely on manual judgments . |
Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning (2026.findings-acl)
Copied to clipboard
Fanding Huang, Guanbo Huang, Xiao Fan, Yi He, Xiao Liang, Xiao Chen, Qinting Jiang, Faisal Nadeem Khan, Jingyan Jiang, Zhi Wang
| Challenge: | Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have substantially improved the reasoning abilities of Large Language Models (LLMs). |
| Approach: | They propose a method that balances exploration and exploitation in the hidden-state space of response trajectories. |
| Outcome: | The proposed model yields consistent improvements across models, algorithms and reasoning benchmarks. |
AlphaContext: An Evolutionary Tree-based Psychometric Context Generator for Creativity Assessment (2026.acl-long)
Copied to clipboard
Yixuan Wang, Yue Huang, Hong Qian, Yunzhao Wei, Yifei Ding, Wenkai Wang, Zhi Liu, Zhongjing Huang, Aimin Zhou, Jiajun Guo
| Challenge: | Existing LLM-based tools struggle with insufficient assessment cues, weak narrative coherence, limited stylistic diversity, and poor support for creative thinking. |
| Approach: | They propose an evolutionary tree-based psychometric context generator that integrates rule-guided outline planning, sentence-level MCTS generation, MAP-Elites quality-diversity optimization and assessment-guide refiner simulation. |
| Outcome: | The proposed tool outperforms strong LLMs and structured frameworks on 7 evaluation dimensions and shows higher alignment with expert-designed contexts. |
Think and Recall: Layer-Level Prompting for Lifelong Model Editing (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for lifelong model editing suffer from limitations in usability, such as requiring additional training corpora or lacking support for reversible and detachable edits. |
| Approach: | They propose a plug-and-play method for knowledge retrieval and storage, i.e., Layer-Level Prompting, which enables seamless and efficient lifelong model editing. |
| Outcome: | The proposed method outperforms existing methods on question answering and hallucination benchmarks across different LLMs. |
Infusing Sequential Information into Conditional Masked Translation Model with Self-Review Mechanism (2020.coling-main)
Copied to clipboard
| Challenge: | Existing non-autoregressive models generate target words in parallel, but with a large latency due to the left-to-right dependency. |
| Approach: | They propose to train a conditional masked translation model and refine results within several iterations to remedy a flawed translation by non-autoregressive models. |
| Outcome: | The proposed model outperforms state-of-the-art models by over 1 BLEU while using less training computations. |
TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text (2025.emnlp-main)
Copied to clipboard
| Challenge: | Current RAG systems concatenate and process numerous retrieved document chunks for prefill . this leads to significant latency in time-to-first-token (TTFT) Experimental results demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG system. |
| Approach: | They propose a hybrid offline-online paradigm that precomputes chunk-level key-value caches and stitches them together at inference time using independent–attention and reorderedRoPE techniques. |
| Outcome: | Experimental results show that TurboRAG reduces TTFT by 9.4x compared to the conventional RAG systems . long concatenated contexts consume disproportionate GPU memory, limiting throughput . |
Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing LLM-based recommenders lack explicit modeling of geographic signals . without explicit modeling geographic signals, recommenders struggle to capture core mobility patterns . |
| Approach: | They propose a framework that utilizes geography as a decision variable within the reasoning process. |
| Outcome: | The proposed framework achieves over 10% relative gains in hit rate over strongest LLM-based baselines and improves cross-city transfer. |
bert2BERT: Towards Reusable Pretrained Language Models (2022.acl-long)
Copied to clipboard
Cheng Chen, Yichun Yin, Lifeng Shang, Xin Jiang, Yujia Qin, Fengyu Wang, Zhi Wang, Xiao Chen, Zhiyuan Liu, Qun Liu
| Challenge: | Pre-training large language models can be expensive and wasteful. |
| Approach: | They propose a method which can transfer the knowledge of an existing smaller pre-trained model to a large model through parameter initialization and a two-stage learning method to further accelerate the pre-training. |
| Outcome: | The proposed method can transfer the knowledge of an existing smaller pre-trained model to a large model through parameter initialization and significantly improve the pre-training efficiency of the large model. |
Beyond A Single AI Cluster: A Survey of Decentralized LLM Training (2025.emnlp-main)
Copied to clipboard
| Challenge: | Decentralized LLM training leverages dispersed resources at varying scales. |
| Approach: | They propose a resource-driven paradigm that leverages dispersed resources across clusters, datacenters and even regions. |
| Outcome: | The proposed model scales are 175 billion to 660 billion parameters, and the exponential growth in computational requirements poses significant challenges. |
KoCo-Bench: Can Large Language Models Leverage Domain Knowledge in Software Development? (2026.acl-long)
Copied to clipboard
Xue Jiang, Ge Li, Jiaru Qian, Xianjie Shi, Chenjie Li, Hao Zhu, Ziyu Wang, Jielun Zhang, Zeyu Zhao, Kechi Zhang, Jia Li, Wenpin Jiao, Zhi Jin, Yihong Dong
| Challenge: | Existing domain-specific code benchmarks focus on assessing what knowledge LLMs possess rather than how they acquire and apply new knowledge. |
| Approach: | They propose a benchmark to evaluate domain specialization methods in real-world software development. |
| Outcome: | KOCO-bench is a new benchmark for evaluating domain specialization methods in real-world software development. |
MusTQ: A Temporal Knowledge Graph Question Answering Dataset for Multi-Step Temporal Reasoning (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing studies focus on fact-centered reasoning with limited attention to temporal reasoning. |
| Approach: | They propose a new TKGQA dataset, MusTQ, which contains 666K multi-step temporal reasoning questions and a TKG. |
| Outcome: | The proposed model achieves state-of-the-art multi-step temporal reasoning ability with entity-time attention mechanism and optimized temporal knowledge graph representation. |
GUI0: Self-Evolving Foundational GUI Agents in Super App Ecosystems (2026.acl-long)
Copied to clipboard
Xinyi Wang, Wei Dai, Kyle Qiao, Ke Wang, Peng Chen, Gang Cao, null Kangqin, Zhongpu Wang, Xiaode Zhang, Yanming Liu, Jihao Gu, Jingtao Xu, Gong Zhi
| Challenge: | Automated interaction with graphical user interfaces (GUIs) is central to general artificial intelligence, but remains challenging within Super App ecosystems. |
| Approach: | They propose a framework synergizing autonomous data synthesis with dual-agent co-evolution . GUI0 establishes a domain-aware foundation model via synthesized corpora and employs curriculum-driven reinforcement learning . |
| Outcome: | The proposed framework outperforms Gemini-2.5-Pro and Claude-4-Sonnet in the SuperAPP benchmark and has universal efficacy across base models. |
VulAgent: Hypothesis-Validation Driven Multi-Agent Architecture for Vulnerability Detection (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent reports indicate that software vulnerabilities caused by insecure coding practices remain a major security threat. |
| Approach: | They propose a multi-agent vulnerability detection framework based on hypothesis validation . they use multi-view analyzers to localize and localize security-sensitive operations . |
| Outcome: | The proposed framework reduces false positives and increases accuracy by 6.6 percentage points on PrimeVul and SVEN. |
SEAG: Structure-Aware Event Causality Generation (2023.findings-acl)
Copied to clipboard
Zhengwei Tao, Zhi Jin, Xiaoying Bai, Haiyan Zhao, Chengfeng Dou, Yongqiang Zhao, Fang Wang, Chongyang Tao
| Challenge: | Current methods for extracting event causality are limited by the lack of cross-task dependencies and may cause error propagation. |
| Approach: | They propose an approach for Structure-Aware Event Causality Generation (SEAG) they generate the ECG structure using a pre-trained language model and perform structural discriminative training alongside auto-regressive generation. |
| Outcome: | The proposed method is effective in extracting event causality from text. |
FinEval: A Chinese Financial Domain Knowledge Evaluation Benchmark for Large Language Models (2025.naacl-long)
Copied to clipboard
Xin Guo, Haotian Xia, Zhaowei Liu, Hanyang Cao, Zhi Yang, Zhiqiang Liu, Sizhe Wang, Jinyi Niu, Chuqi Wang, Yanhui Wang, Xiaolong Liang, Xiaoming Huang, Bing Zhu, Zhongyu Wei, Yun Chen, Weining Shen, Liwen Zhang
| Challenge: | Large language models have demonstrated outstanding performance in various natural language processing tasks, but their security capabilities in the financial domain have not been explored. |
| Approach: | They propose to use a benchmark to evaluate large language models' financial domain knowledge and practical abilities. |
| Outcome: | The proposed benchmark evaluates large language models' financial domain knowledge and practical abilities. |
Your Inference Request Will Become a Black Box: Confidential Inference for Cloud-based Large Language Models (2026.acl-long)
Copied to clipboard
Chung-ju Huang, Huiqiang Zhao, Yuanpeng He, Lijian Li, Wenpin Jiao, Zhi Jin, Peixuan Chen, Leye Wang
| Challenge: | Existing approaches fail to ensure privacy, maintain model performance, and preserve computational efficiency simultaneously. |
| Approach: | They propose a confidential inference framework that partitions the LLM pipeline between a client-verified Confidential Virtual Machine (CVM) and the public cloud to protect client data without compromising the cloud’s model intellectual property or inference quality. |
| Outcome: | The proposed framework can defend against state-of-the-art token inference attacks while preserving model privacy, performance, and efficiency. |
QueryAttack: Jailbreaking Aligned Large Language Models Using Structured Non-natural Query Language (2025.findings-acl)
Copied to clipboard
Qingsong Zou, Jingyu Xiao, Qing Li, Zhi Yan, Yuhang Wang, Li Xu, Wenxuan Wang, Kuofeng Gao, Ruoyu Li, Yong Jiang
| Challenge: | Existing methods to bypass security defenses of large language models (LLMs) are not effective, but QueryAttack can be jailbroken. |
| Approach: | They propose a framework to examine generalizability of safety alignment by translating malicious queries into structured non-natural query languages. |
| Outcome: | The proposed framework can achieve high attack success rates and jailbreak various defense methods on mainstream LLMs. |
Bridging Kernel Drivers and Virtual Device Models with LLM-Powered Automation (2026.acl-demo)
Copied to clipboard
| Challenge: | Linux kernel device drivers are tightly coupled with hardware, making them difficult to execute and test without physical devices. |
| Approach: | They present a tool that generates QEMU-based virtual devices directly from Linux driver source code. |
| Outcome: | The proposed tool generates QEMU-based virtual devices directly from Linux driver source code. |
CLEVA: Chinese Language Models EVAluation Platform (2023.emnlp-demo)
Copied to clipboard
Yanyang Li, Jianqiao Zhao, Duo Zheng, Zi-Yuan Hu, Zhi Chen, Xiaohui Su, Yongfeng Huang, Shijia Huang, Dahua Lin, Michael Lyu, Liwei Wang
| Challenge: | Large language models (LLMs) have revolutionized natural language processing. |
| Approach: | They propose a Chinese-based platform that assesses Chinese LLMs using a standardized workflow and a unique sampling strategy. |
| Outcome: | CLEVA evaluates Chinese LLMs on a standardized workflow and a competitive leaderboard with minimal coding. |