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. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Jia Li, Xuyuan Guo, Lei Li, Kechi Zhang, Ge Li, Jia Li, Zhengwei Tao, Fang Liu, Chongyang Tao, Yuqi Zhu, Zhi Jin. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Code generation is a latency-sensitive task that demands high timeliness. However, with the growing interest and inherent difficulty in repository-level code generation, most existing code generation studies focus on improving the correctness of generated code while overlooking the inference efficiency, which is substantially affected by the overhead during LLM generation. Although there has been work on accelerating LLM inference, these approaches are not tailored to the specific characteristics of code generation; instead, they treat code the same as natural language sequences and ignore its unique syntax and semantic characteristics, which are also crucial for improving efficiency. Consequently, these approaches exhibit limited effectiveness in code generation tasks, particularly for repository-level scenarios with considerable complexity and difficulty. To alleviate this issue, following draft-verification paradigm, we propose FastCoder, a simple yet highly efficient inference acceleration approach specifically designed for code generation, without compromising the quality of the output. FastCoder constructs a multi-source datastore, providing access to both general and project-specific knowledge, facilitating the retrieval of high-quality draft sequences. Moreover, FastCoder reduces the retrieval cost by controlling retrieval timing, and enhances efficiency through parallel retrieval and a context- and LLM preference-aware cache. Experimental results show that FastCoder can reach up to 2.53× and 2.54× speedup compared to autoregressive decoding in repository-level and standalone code generation tasks, respectively, outperforming state-of-the-art inference acceleration approaches by up to 88%. FastCoder can also be integrated with existing correctness-focused code generation approaches to accelerate the LLM generation process, and reach a speedup exceeding 2.6×.

The Transformer architecture has emerged as a landmark advancement within the broad field of artificial intelligence, effectively catalyzing the advent of large language models (LLMs). However, despite its remarkable capabilities and the substantial progress it has facilitated, the Transformer architecture still has some limitations. One such intrinsic limitation is its inability to effectively recognize regular expressions or deterministic context-free grammars. Standard Transformers lack an explicit mechanism for recursion and structured state transitions, which can hinder systematic generalization on nested and hierarchical patterns. Drawing inspiration from pushdown automata, which efficiently resolve deterministic context-free grammars using stacks, we equip layers with a differentiable stack and propose StackTrans with recursion to address the aforementioned issue within LLMs. Unlike previous approaches that modify the attention computation, StackTrans explicitly incorporates hidden state stacks between Transformer layers. This design maintains compatibility with existing frameworks like flash-attention. Specifically, our design features stack operations -- such as pushing and popping hidden states -- that are differentiable and can be learned in an end-to-end manner. Our comprehensive evaluation spans benchmarks for both Chomsky hierarchy and large-scale natural languages. Across these diverse tasks, StackTrans consistently outperforms standard Transformer models and other baselines. We have successfully scaled StackTrans up from 360M to 7B parameters. In particular, our from-scratch pretrained model StackTrans-360M outperforms several larger open-source LLMs with 2–3x more parameters, showcasing its superior efficiency and reasoning capability.

Large Language Models (LLMs) have shown impressive In-Context Learning (ICL) ability in code generation. LLMs take a prompt context consisting of a few demonstration examples and a new requirement as input, and output new programs without any parameter update. Existing studies have found that the performance of ICL-based code generation heavily depends on the quality of demonstration examples and thus arises research on selecting demonstration examples: given a new requirement, a few demonstration examples are selected from a candidate pool, where LLMs are expected to learn the pattern hidden in these selected demonstration examples. Existing approaches are mostly based on heuristics or randomly selecting examples. However, the distribution of randomly selected examples usually varies greatly, making the performance of LLMs less robust. The heuristics retrieve examples by only considering textual similarities of requirements, leading to sub-optimal performance. To fill this gap, we propose a L arge language model- A ware selection approach for I n-context- L earning-based code generation named LAIL. LAIL uses LLMs themselves to select examples. It requires LLMs themselves to label a candidate example as a positive example or a negative example for a requirement. Positive examples are helpful for LLMs to generate correct programs, while negative examples are trivial and should be ignored. Based on the labeled positive and negative data, LAIL trains a model-aware retriever to learn the preference of LLMs and select demonstration examples that LLMs need. During the inference, given a new requirement, LAIL uses the trained retriever to select a few examples and feed them into LLMs to generate desired programs. We apply LAIL to four widely used LLMs and evaluate it on five code generation datasets. Extensive experiments demonstrate that LAIL outperforms the state-of-the-art (SOTA) baselines by 11.58%, 3.33%, and 5.07% on CodeGen-Multi-16B, 1.32%, 2.29%, and 1.20% on CodeLlama-34B, and achieves 4.38%, 2.85%, and 2.74% improvements on Text-davinci-003 in terms of Pass@1 at MBJP, MBPP, and MBCPP, respectively. In addition to function-level code generation, LAIL improves the performance of LLMs on DevEval, a repository-level code generation dataset, which achieves 10.04%, 8.12%, and 4.63% improvements compared to the SOTA baselines at Pass@1, 3, and 5 on CodeLlama-7B. Human evaluation further verifies that the generated programs of LAIL are superior in correctness, code quality, and maintainability. Besides, LAIL has satisfactory transferability across different LLMs and datasets, where the retriever learned on one LLM (dataset) can be transferred to other LLMs (datasets).

Recently, Large Language Models (LLMs) have shown impressive abilities in code generation. However, existing LLMs' decoding strategies are designed for Natural Language (NL) generation, overlooking the differences between NL and programming languages (PL). Due to this oversight, a better decoding strategy for code generation remains an open question. In this paper, we conduct the first systematic study to explore a decoding strategy specialized in code generation. With an analysis of loss distributions of code tokens, we find that code tokens can be divided into two categories: challenging tokens that are difficult to predict and confident tokens that can be easily inferred. Among them, the challenging tokens mainly appear at the beginning of a code block. Inspired by the above findings, we propose a simple yet effective method: Adaptive Temperature (AdapT) sampling, which dynamically adjusts the temperature coefficient when decoding different tokens. We apply a larger temperature when sampling for challenging tokens, allowing LLMs to explore diverse choices. We employ a smaller temperature for confident tokens avoiding the influence of tail randomness noises. We apply AdapT sampling to LLMs with different sizes and conduct evaluations on two popular datasets. Results show that AdapT sampling significantly outperforms state-of-the-art decoding strategy.

Large Language Models (LLMs) have shown promise in automated code generation but typically excel only in simpler tasks such as generating standalone code units.However, real-world software development often involves complex code repositories with complex dependencies and extensive documentation.To enable LLMs to handle these realworld repo-level code generation, we present CODEAGENT, a novel LLM-based agent framework that employs external tools for effective repo-level code generation.CODEAGENT integrates five programming tools, enabling interaction with software artifacts for information retrieval, code implementation, and code testing.We implement four agent strategies to optimize these tools' usage.To the best of our knowledge, CODEAGENT is the first agent framework specifically for repolevel code generation.In order to measure the effectiveness of our method at the repository level, we design a repo-level benchmark CODEAGENTBENCH.The performance on this benchmark shows a significant improvement brought by our method, with improvements in pass rate ranging from 2.0 to 15.8.Further tests on the HumanEval benchmark confirm CODEAGENT's adaptability and efficacy across various code generation tasks.Notably, CODEAGENT outperforms commercial products like GitHub Copilot, showcasing superior accuracy and efficiency.These results demonstrate CODEAGENT's robust capabilities in code generation, highlighting its potential for real-world repo-level coding challenges.

Large Language Models (LLMs), such as GPT-4, StarCoder, and Code Llama, are transforming the way developers approach programming by automatically generating code based on given contexts, such as natural language descriptions or incomplete surrounding code. Despite advancements, generating syntactically and semantically correct code remains challenging, especially for complex programming tasks. Existing approaches typically generate multiple candidate solutions using LLMs to increase the likelihood of producing correct code. However, selecting the correct code from these candidates --- a process known as code ranking --- remains a major challenge. Current research on code ranking can be categorized into execution-based and non-execution-based methods. Execution-based methods, although effective, encounter notable limitations, such as scarcity of quality unit tests and security risks. Non-execution-based methods like CodeRanker, which rely solely on classification labels to train a code ranker, struggle to capture subtle errors and provide detailed error insights. Recognizing the strengths and limitations of both approaches, we propose a new method that integrates the advantages of execution-based and non-execution-based techniques. The key insight of our work is that an effective code ranker is expected to truly comprehend the underlying causes of erroneous code, as relying solely on classification labels is insufficient. Inspired by this, this paper puts forward RankEF, an innovative approach for code ranking that leverages execution feedback. RankEF employs multi-task learning to integrate code classification with execution feedback generation. This approach enables the model to understand the reasons behind incorrect code, distinguishing between correct and incorrect solutions without the need to execute the code during the ranking phase. Experiments on three code generation benchmarks---APPS, MBPP, and HumanEval---demonstrate that RankEF significantly outperforms the state-of-the-art CodeRanker, achieving relative improvements of +30.97%, +31.43%, and +19.51% in Pass@1, Pass@2, and Pass@5 on APPS test, respectively.

Abstract Bug localization, which is used to help programmers identify the location of bugs in source code, is an essential task in software development. Researchers have already made efforts to harness the powerful deep learning (DL) techniques to automate it. However, training bug localization model is usually challenging because it requires a large quantity of data labeled with the bug's exact location, which is difficult and time‐consuming to collect. By contrast, obtaining bug detection data with binary labels of whether there is a bug in the source code is much simpler. This paper proposes a WEakly supervised bug LocaLization (WELL) method, which only uses the bug detection data with binary labels to train a bug localization model. With CodeBERT finetuned on the buggy‐or‐not binary labeled data, WELL can address bug localization in a weakly supervised manner. The evaluations on three method‐level synthetic datasets and one file‐level real‐world dataset show that WELL is significantly better than the existing state‐of‐the‐art model in typical bug localization tasks such as variable misuse and other bugs.

There is growing interest in software migration as the development of software and society. Manually migrating projects between languages is error-prone and expensive. In recent years, researchers have begun to explore automatic program translation using supervised deep learning techniques by learning from large-scale parallel code corpus. However, parallel resources are scarce in the programming language domain, and it is costly to collect bilingual data manually. To address this issue, several unsupervised programming translation systems are proposed. However, these systems still rely on huge monolingual source code to train, which is very expensive. Besides, these models cannot perform well for translating the languages that are not seen during the pre-training procedure. In this paper, we propose SDA-Trans, a syntax and domain-aware model for program translation, which leverages the syntax structure and domain knowledge to enhance the cross-lingual transfer ability. SDA-Trans adopts unsupervised training on a smaller-scale corpus, including Python and Java monolingual programs. The experimental results on function translation tasks between Python, Java, and C++ show that SDA-Trans outperforms many large-scale pre-trained models, especially for unseen language translation.

Code search has been a critical software development activity in facilitating developers to retrieve a proper code snippet from open-source repositories given a user intent. In recent years, large-scale pre-trained models have shown impressive performance on code representation learning and have achieved state-of-the-art performance on code search task. However, it is challenging for these models to distinguish the functionally equivalent code snippets with dissimilar implementations or the non-equivalent code snippets that look similar. Due to the diversity of the code implementations, it is necessary for the code search engines to identify the functional similarities or dissimilarities of source code so as to return the functionally matched source code for a given query. Besides, existing pre-trained models mainly focus on learning the semantic representations of code snippets. The semantic correlation between the code snippet and natural language query is not sufficiently exploited. An effective code search tool not only needs to understand the relationship between queries and code snippets but also needs to identify the relationship between diversified code snippets. To address these limitations, we propose a novel multi-view contrastive learning model MCodeSearcher for code retrieval, aiming at sufficiently exploiting (1) the semantic correlation between queries and code snippets, and (2) the relationship between functionally equivalent code snippets. To achieve this, we design contrastive training objectives from three views and pre-train our model with these objectives. The experimental results on five representative code search datasets show that our approach significantly outperforms the state-of-the-art methods.

Developers introduce code clones to improve programming productivity. Many existing studies have achieved impressive performance in monolingual code clone detection. However, during software development, more and more developers write semantically equivalent programs with different languages to support different platforms and help developers translate projects from one language to another. Considering that collecting cross-language parallel data, especially for low-resource languages, is expensive and time-consuming, how designing an effective cross-language model that does not rely on any parallel data is a significant problem. In this paper, we propose a novel method named ZC3for Z_ero-shot Cross-language Code Clone detection. ZC3designs the contrastive snippet prediction to form an isomorphic representation space among different programming languages. Based on this, ZC3exploits domain-aware learning and cycle consistency learning to further constrain the model to generate representations that are aligned among different languages meanwhile are diacritical for different types of clones. To evaluate our approach, we conduct extensive experiments on four representative cross-language clone detection datasets. Experimental results show that ZC3outperforms the state-of-the-art baselines by 67.12%, 51.39%, 14.85%, and 53.01% on the MAP score, respectively. We further investigate the representational distribution of different languages and discuss the effectiveness of our method.

Aspect-based sentiment analysis (ABSA) plays an indispensable role in web mining and retrieval system as it involves a wide range of tasks, including aspect term extraction, opinion term extraction, aspect sentiment classification, etc. Early works are merely applicable to a part of these tasks, leading to computation-unfriendly models and a pipeline framework. Recently, a unified framework has been proposed to learn all these ABSA tasks in an end-to-end fashion. Despite its versatility, its performance is still sub-optimal since ABSA tasks depend heavily on both sentiment and syntax knowledge, but existing task-specific knowledge integration methods are hardly applicable to such a unified framework. Therefore, we propose a brand-new unified framework for ABSA in this work, which incorporates both implicit sentiment knowledge and explicit syntax knowledge to better complete all ABSA tasks. To effectively incorporate implicit sentiment knowledge, we first design a self-supervised pre-training procedure that is general enough to all ABSA tasks. It consists of conjunctive words prediction (CWP) task, sentiment-word polarity prediction (SPP) task, attribute nouns prediction (ANP) task, and sentiment-oriented masked language modeling (SMLM) task. Empowered by the pre-training procedure, our framework acquires strong abilities in sentiment representation and sentiment understanding. Meantime, considering a subtle syntax variation can significantly affect ABSA, we further explore a sparse relational graph attention network (SR-GAT) to introduce explicit aspect-oriented syntax knowledge. By combining both worlds of knowledge, our unified model can better represent and understand the input texts towards all ABSA tasks. Extensive experiments show that our proposed framework achieves consistent and significant improvements on all ABSA tasks.

Event detection is challenging in real-world application since new events continually occur and old events still exist which may result in repeated labeling for old events. Therefore, incremental event detection is essential where a model continuously learns new events and meanwhile prevents performance from degrading on old events. Although existing incremental event detection models achieve impressive performance, they face the data imbalance problem between old classes and new classes, and have the knowledge transfer problem which cannot adequately utilize the knowledge provided by the previous model and data. To this end, we propose a Balance-Normalization-Uncertainty (BNU) model to address above problems. Specifically, in order to mitigate the adverse effects of data imbalance, we incorporate a balanced fine-tuning stage and a cosine normalization module. Meanwhile, we consider aleatoric uncertainty to preserve previous knowledge while training for new events. Experimental results show that our proposed method resolves the above challenges effectively and achieves consistent and significant performance on ACE and TAC KBP datasets.

Recently, many neural context-response matching models have been developed for retrieval-based dialogue systems. Although existing models achieve impressive performance through learning on a large amount of in-domain parallel dialogue data, they usually perform worse in another new domain. How to transfer a response retrieval model trained in high-resource domains to other low-resource domains is a crucial problem for scalable dialogue systems. To this end, we investigate the unsupervised cross-domain adaptation for response selection when the target domain has no parallel dialogue data. Specifically, we propose a two-stage method to adapt a response selection model to a new domain using self-supervised and adversarial training based on pre-trained language models (PLMs). To efficiently incorporate domain awareness and target-domain knowledge to PLMs, we first design a self-supervised post-training procedure, including domain discrimination (DD) task, target-domain masked language model (MLM) task and target-domain next sentence prediction (NSP) task. Based on this, we further conduct the adversarial fine-tuning to empower the model to match the proper response with extracted domain-shared features as much as possible. Experimental results show that our proposed method achieves consistent and significant improvements on several cross-domain response selection datasets.

We propose TUTA, a unified pre-training architecture for understanding generally structured tables. Noticing that understanding a table requires spatial, hierarchical, and semantic information, we enhance transformers with three novel structure-aware mechanisms. First, we devise a unified tree-based structure, called a bi-dimensional coordinate tree, to describe both the spatial and hierarchical information of generally structured tables. Upon this, we propose tree-based attention and position embedding to better capture the spatial and hierarchical information. Moreover, we devise three progressive pre-training objectives to enable representations at the token, cell, and table levels. We pre-train TUTA on a wide range of unlabeled web and spreadsheet tables and fine-tune it on two critical tasks in the field of table structure understanding: cell type classification and table type classification. Experiments show that TUTA is highly effective, achieving state-of-the-art on five widely-studied datasets.

Jia Li, Chongyang Tao, Wei Wu, Yansong Feng, Dongyan Zhao, Rui Yan. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.