While Large Language Models (LLMs) excel at code generation, their inherent tendency toward verbatim memorization of training data introduces critical risks like copyright infringement, insecurity emission, and deprecated API utilization, etc. A straightforward yet promising defense is unlearning, i.e., erasing or down-weighting the offending snippets through post-training. However, we find its application to source code often tends to spill over, damaging the basic knowledge of programming languages learned by the LLM and degrading the overall capability. To ease this challenge, we propose PROD for precise source code unlearning. PROD surgically zeroes out the prediction probability of the prohibited tokens, and renormalizes the remaining distribution so that the generated code stays correct. By excising only the targeted snippets, PROD achieves precise forgetting without much degradation of the LLM's overall capability. To facilitate in-depth evaluation against PROD, we establish an unlearning benchmark consisting of three downstream tasks (i.e., unlearning of copyrighted code, insecure code, and deprecated APIs), and introduce Pareto Dominance Ratio (PDR) metric, which indicates both the forget quality and the LLM utility. Our comprehensive evaluation demonstrates that PROD achieves superior overall performance between forget quality and model utility compared to existing unlearning approaches across three downstream tasks, while consistently exhibiting improvements when applied to LLMs of varying series. PROD also exhibits superior robustness against adversarial attacks without generating or exposing the data to be forgotten. These results underscore that our approach not only successfully extends the application boundary of unlearning techniques to source code, but also holds significant implications for advancing reliable code generation.

Yihong Dong, Xue Jiang, Yongding Tao, Huanyu Liu, Kechi Zhang, Lili Mou, Rongyu Cao, Yingwei MA, Jue Chen, Binhua Li, Zhi Jin, Fei Huang, Yongbin Li, Ge Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent advancements in software engineering agents have demonstrated promising capabilities in automating program improvements. However, their reliance on closed-source or resource-intensive models introduces significant deployment challenges in private environments, prompting a critical question: How can personally deployable open-source LLMs (e.g., 32B models running on a single GPU) achieve comparable code reasoning performanceƒ To this end, we propose a unified Test-Time Compute (TTC) scaling framework that leverages increased inference-time computation instead of larger models. Our framework incorporates two complementary strategies: internal TTC and external TTC. Internally, we introduce a development-contextualized trajectory synthesis method leveraging real-world software repositories to bootstrap multi-stage reasoning processes, such as fault localization and patch generation. We further enhance trajectory quality through rejection sampling, rigorously evaluating trajectories along accuracy and complexity. Externally, we propose a novel development-process-based search strategy guided by reward models and execution verification. This approach enables targeted computational allocation at critical development decision points, overcoming limitations of existing "end-point only" verification methods.Evaluations on SWE-bench Verified demonstrate our 32B model achieves a 46% issue resolution rate, surpassing significantly larger models such as DeepSeek R1 671B and OpenAI o1. Additionally, we provide the empirical validation of the test-time scaling phenomenon within SWE agents, revealing that models dynamically allocate more tokens to increasingly challenging problems, effectively enhancing reasoning capabilities. We publicly release all training data, models, and code to facilitate future research.1. In fact, our method has been deployed in Tongyi Lingma, an IDE-based coding assistant developed by Alibaba Cloud, where it helps developers solve real-world programming problems.

Liang Li, Ruiying Geng, Chengyang Fang, Bing Li, Can Ma, Rongyu Cao, Binhua Li, Fei Huang, Yongbin Li. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

In this paper, we revisit the problem of extracting the values of a given set of key fields from form-like documents. It is the vital step to support many downstream applications, such as knowledge base construction, question answering, document comprehension and so on. Previous studies ignore the semantics of the given keys by considering them only as the class labels, and thus might be incapable to handle zero-shot keys. Meanwhile, although these models often leverage the attention mechanism, the learned features might not reflect the true proxy of explanations on why humans would recognize the value for the key, and thus could not well generalize to new documents. To address these issues, we propose a Key-Aware and Trigger-Aware (KATA) extraction model. With the input key, it explicitly learns two mappings, namely from key representations to trigger representations and then from trigger representations to values. These two mappings might be intrinsic and invariant across different keys and documents. With a large training set automatically constructed based on the Wikipedia data, we pre-train these two mappings. Experiments with the fine-tuning step to two applications show that the proposed model achieves more than 70% accuracy for the extraction of zero-shot keys while previous methods all fail.

The initial public offering (IPO) market in Hong Kong is consistently one of the largest in the world. As part of its regulatory responsibilities, Hong Kong Exchanges and Clearing Limited (HKEX) reviews annual reports published by listed companies (issuers). The number of issuers has grown at a fast pace, reaching 2,538 as the end of 2020. This poses a challenge for manually reviewing these annual reports against the many diverse regulatory obligations (listing rules). We propose a system named Jura to improve the efficiency of annual report reviewing with the help of machine learning methods. This system checks the compliance of an issuer's published information against listing rules in four steps: panoptic document recognition, relevant passage location, fine-grained information extraction, and compliance assessment. This paper introduces in detail the passage location step, how it is critical for speeding up compliance assessment, and the various challenges faced. We argue that although a passage is a relatively independent unit, it needs to be combined with document structure and contextual information to accurately locate the relevant passages. With the help of Jura, HKEX reports saving 80% of the time on reviewing issuers' annual reports.

Table, widely used in documents from various vertical domains, is a compact representation of data. There is always some strong demand to automatically extract key information from tables for further analysis. In addition, the set of keys that need to be extracted information is usually time-varying, which arises the issue of zero-shot keys in this situation. To increase the efficiency of these knowledge workers, in this study we aim to extract the values of a given set of keys from tables. Previous table-related studies mainly focus on relational, entity, and matrix tables. However, their methods fail on mixed-style tables, in which table headers might exist in any non-merged or merged cell, and the spatial relationships between headers and corresponding values are diverse. Here, we address this problem while taking mixed-style tables into account. To this end, we propose an end-to-end neural-based model, called Information Extraction in Mixed-style Table (IEMT). IEMT first uses BERT to extract textual semantics of the given key and the words in each cell. Then, it uses multi-layer CNN to capture the spatial and textual interactions among adjacent cells. Furthermore, to improve the accuracy on zero-shot keys, we pre-train IEMT on a dataset constructed on 0.4 million tables from Wikipedia and 140 million triplets from Ownthink. Experiments with the fine-tuning step on 26,869 financial tables show that the proposed model achieves 0.9323 accuracy for zero-shot keys, obtaining more than 8% increase compared with the model without pre-training.

Job scheduling is crucial in high-performance computing (HPC), which is dedicated to deciding when and which jobs are allocated to the system and placing the jobs on which resources, by considering multiple scheduling goals. Along with the incremental of various resources and dazzling deep learning training (DLT) workloads, job failure becomes a quite common issue in HPC, which will affect user satisfaction and cluster utilization. To alleviate the influence of hardware and software errors as much as possible, in this paper, we aim to tackle the problem of failure-aware job scheduling in HPC clusters. Inspired by the success of previous studies of deep reinforcement learning-driven job scheduling, we propose a novel HPC scheduling agent named FARS (Failure-aware RL-based scheduler) by considering the effects of job failures. On the one hand, a neural network is applied to map the information of raw cluster and job states to job placement decisions. On the other hand, to consider the influence of job failure for user satisfaction and cluster utilization, FARS leverages make-span of the entire workload as the training objective. Additionally, effective exploration and experience replay techniques are applied to obtain effectively converged agent. To evaluate the capability of FARS, we design extensive trace-based simulation experiments with the popular DLT workloads. The experimental results show that, compared with the best baseline model, FARS obtains 5.69% improvement of average make-span under different device error rates. Together, our FARS is an ideal candidate for failure-aware job scheduler in HPC clusters.

Different from other sequential data, sentences in natural language are structured by linguistic grammars. Previous generative conversational models with chain-structured decoder ignore this structure in human language and might generate plausible responses with less satisfactory relevance and fluency. In this study, we aim to incorporate the results from linguistic analysis into the process of sentence generation for high-quality conversation generation. Specifically, we use a dependency parser to transform each response sentence into a dependency tree and construct a training corpus of sentence-tree pairs. A tree-structured decoder is developed to learn the mapping from a sentence to its tree, where different types of hidden states are used to depict the local dependencies from an internal tree node to its children. For training acceleration, we propose a tree canonicalization method, which transforms trees into equivalent ternary trees. Then, with a proposed tree-structured search method, the model is able to generate the most probable responses in the form of dependency trees, which are finally flattened into sequences as the system output. Experimental results demonstrate that the proposed X2Tree framework outperforms baseline methods over 11.15% increase of acceptance ratio.

To the same utterance, people's responses in everyday dialogue may be diverse largely in terms of content semantics, speaking styles, communication intentions and so on. Previous generative conversational models ignore these 1-to-n relationships between a post to its diverse responses, and tend to return high-frequency but meaningless responses. In this study we propose a mechanism-aware neural machine for dialogue response generation. It assumes that there exists some latent responding mechanisms, each of which can generate different responses for a single input post. With this assumption we model different responding mechanisms as latent embeddings, and develop a encoder-diverter-decoder framework to train its modules in an end-to-end fashion. With the learned latent mechanisms, for the first time these decomposed modules can be used to encode the input into mechanism-aware context, and decode the responses with the controlled generation styles and topics. Finally, the experiments with human judgements, intuitive examples, detailed discussions demonstrate the quality and diversity of the generated responses with 9.80% increase of acceptable ratio over the best of six baseline methods.