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.

Complex logic errors in LLM-generated code are challenging to diagnose and repair. While existing LLM-based self-repair approaches conduct intensive static semantic analysis or rely on superficial execution logs, they miss the in-depth runtime behaviors that often expose bug root causes—lacking the interactive dynamic analysis capabilities that make human debugging effective. We present InspectCoder, the first agentic program repair system that empowers LLMs to actively conduct dynamic analysis via interactive debugger control. Our dual-agent framework enables strategic breakpoint placement, targeted state inspection, and incremental runtime experimentation within stateful debugger sessions. Unlike existing methods that follow fixed log collection procedures, InspectCoder adaptively inspects and perturbs relevant intermediate states at runtime, and leverages immediate process rewards from debugger feedback to guide multi-step reasoning, transforming LLM debugging paradigm from blind trial-and-error into systematic root cause diagnosis. We conduct comprehensive experiments on two challenging self-repair benchmarks: BigCodeBench-R and LiveCodeBench-R. InspectCoder achieves 5.10%–60.37% relative improvements in repair accuracy over the strongest baseline, while delivering 1.67x-2.24x superior bug-fix efficiency respectively.We also contribute InspectWare, an open-source middleware that abstracts debugger complexities and maintains stateful debugging sessions across mainstream Python testing frameworks. Our work provides actionable insight into the interactive LLM-debugger systems, demonstrating the significant potential of LLM-driven dynamic analysis for automated software engineering.

Yunkun Wang, Yue Zhang, Zhen Qin, Chen Zhi, Binhua Li, Fei Huang, Yongbin Li, Shuiguang Deng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization.

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.

Jiaxi Yang, Binyuan Hui, Min Yang, Bailin Wang, Bowen Li, Binhua Li, Fei Huang, Yongbin Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

How to evaluate Large Language Models (LLMs) in code generation remains an open question. Many benchmarks have been proposed, but they have two limitations, i.e., data leakage and lack of domain-specific evaluation.The former hurts the fairness of benchmarks, and the latter hinders practitioners from selecting superior LLMs for specific programming domains.To address these two limitations, we propose a new benchmark - EvoCodeBench, which has the following advances: (1) Evolving data. EvoCodeBench will be dynamically updated every period (e.g., 6 months) to avoid data leakage. This paper releases the first version - EvoCodeBench-2403, containing 275 samples from 25 repositories.(2) A domain taxonomy and domain labels. Based on the statistics of open-source communities, we design a programming domain taxonomy consisting of 10 popular domains. Based on the taxonomy, we annotate each sample in EvoCodeBench with a domain label. EvoCodeBench provides a broad platform for domain-specific evaluations.(3) Domain-specific evaluations. Besides the Pass@k, we compute the Domain-Specific Improvement (DSI) and define LLMs' comfort and strange domains. These evaluations help practitioners select superior LLMs in specific domains and discover the shortcomings of existing LLMs.Besides, EvoCodeBench is collected by a rigorous pipeline and aligns with real-world repositories in multiple aspects (e.g., code distributions).We evaluate 8 popular LLMs (e.g., gpt-4, DeepSeek Coder, StarCoder 2) on EvoCodeBench and summarize some insights. EvoCodeBench reveals the actual abilities of these LLMs in real-world repositories. For example, the highest Pass@1 of gpt-4 on EvoCodeBench-2403 is only 20.74%. Besides, we evaluate LLMs in different domains and discover their comfort and strange domains. For example, gpt-4 performs best in most domains but falls behind others in the Internet domain. StarCoder 2-15B unexpectedly performs well in the Database domain and even outperforms 33B LLMs. We release EvoCodeBench, all prompts, and LLMs' completions for further community analysis.

Abstract Data-to-text (D2T) generation aims to transform structured data into natural language text. Data-to-text pre-training has proved to be powerful in enhancing D2T generation and yields impressive performance. However, previous pre-training methods either oversimplified structured data into a sequence without considering input structures or designed training objectives tailored for a specific data structure (e.g., table or knowledge graph). In this paper, we unify different types of structured data (i.e., table, key-value data, knowledge graph) into the graph format and cast different D2T generation tasks as graph-to-text generation. To effectively exploit the structural information of the input graph, we propose a structure-enhanced pre-training method for D2T generation by designing a structure-enhanced Transformer. Concretely, we devise a position matrix for the Transformer, encoding relative positional information of connected nodes in the input graph. In addition, we propose a new attention matrix to incorporate graph structures into the original Transformer by taking the available explicit connectivity structure into account. Extensive experiments on six benchmark datasets show the effectiveness of our model. Our source codes are available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/unid2t.

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.

Hao Sun, Yang Li, Liwei Deng, Bowen Li, Binyuan Hui, Binhua Li, Yunshi Lan, Yan Zhang, Yongbin Li. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Text-to-SQL parsing, which aims at converting natural language instructions into executable SQLs, has gained increasing attention in recent years. In particular, GPT-4 and Claude-2 have shown impressive results in this task. However, most of the prevalent benchmarks, i.e., Spider, and WikiSQL, focus on database schema with few rows of database contents leaving the gap between academic study and real-world applications. To mitigate this gap, we present BIRD, a BIg benchmark for laRge-scale Database grounded in text-to-SQL tasks, containing 12,751 pairs of text-to-SQL data and 95 databases with a total size of 33.4 GB, spanning 37 professional domains. Our emphasis on database values highlights the new challenges of dirty database contents, external knowledge between NL questions and database contents, and SQL efficiency, particularly in the context of massive databases. To solve these problems, text-to-SQL models must feature database value comprehension in addition to semantic parsing. The experimental results demonstrate the significance of database values in generating accurate text-to-SQLs for big databases. Furthermore, even the most popular and effective text-to-SQL models, i.e. GPT-4, only achieve 54.89% in execution accuracy, which is still far from the human result of 92.96%, proving that challenges still stand. We also provide an efficiency analysis to offer insights into generating text-to-efficient-SQLs that are beneficial to industries. We believe that BIRD will contribute to advancing real-world applications of text-to-SQL research.The leaderboard and source code are available: https://bird-bench.github.io/.

Table-based reasoning has shown remarkable progress in a wide range of table-based tasks. It is a challenging task, which requires reasoning over both free-form natural language (NL) questions and (semi-)structured tabular data. However, previous table-based reasoning solutions usually suffer from significant performance degradation on ''huge'' evidence (tables). In addition, most existing methods struggle to reason over complex questions since the essential information is scattered in different places. To alleviate the above challenges, we exploit large language models (LLMs) as decomposers for effective table-based reasoning, which (i) decompose huge evidence (a huge table) into sub-evidence (a small table) to mitigate the interference of useless information for table reasoning, and (ii) decompose a complex question into simpler sub-questions for text reasoning. First, we use a powerful LLM to decompose the evidence involved in the current question into the sub-evidence that retains the relevant information and excludes the remaining irrelevant information from the ''huge'' evidence. Second, we propose a novel ''parsing-execution-filling'' strategy to decompose a complex question into simper step-by-step sub-questions by generating intermediate SQL queries as a bridge to produce numerical and logical sub-questions with a powerful LLM. Finally, we leverage the decomposed sub-evidence and sub-questions to get the final answer with a few in-context prompting examples. Extensive experiments on three benchmark datasets (TabFact, WikiTableQuestion, and FetaQA) demonstrate that our method achieves significantly better results than competitive baselines for table-based reasoning. Notably, our method outperforms human performance for the first time on the TabFact dataset. In addition to impressive overall performance, our method also has the advantage of interpretability, where the returned results are to some extent tractable with the generated sub-evidence and sub-questions. For reproducibility, we release our source code and data at: https://github.com/AlibabaResearch/DAMO-ConvAI.

Most graph-to-text works are built on the encoder-decoder framework with cross-attention mechanism. Recent studies have shown that explicitly modeling the input graph structure can significantly improve the performance. However, the vanilla structural encoder cannot capture all specialized information in a single forward pass for all decoding steps, resulting in inaccurate semantic representations. Meanwhile, the input graph is flatted as an unordered sequence in the cross attention, ignoring the original graph structure. As a result, the obtained input graph context vector in the decoder may be flawed. To address these issues, we propose a Structure-Aware Cross-Attention (SACA) mechanism to re-encode the input graph representation conditioning on the newly generated context at each decoding step in a structure aware manner. We further adapt SACA and introduce its variant Dynamic Graph Pruning (DGP) mechanism to dynamically drop irrelevant nodes in the decoding process. We achieve new state-of-the-art results on two graph-to-text datasets, LDC2020T02 and ENT-DESC, with only minor increase on computational cost.

This paper aims to improve the performance of text-to-SQL parsing by exploring the intrinsic uncertainties in the neural network based approaches (called SUN). From the data uncertainty perspective, it is indisputable that a single SQL can be learned from multiple semantically-equivalent questions. Different from previous methods that are limited to one-to-one mapping, we propose a data uncertainty constraint to explore the underlying complementary semantic information among multiple semantically-equivalent questions (many-to-one) and learn the robust feature representations with reduced spurious associations. In this way, we can reduce the sensitivity of the learned representations and improve the robustness of the parser. From the model uncertainty perspective, there is often structural information (dependence) among the weights of neural networks. To improve the generalizability and stability of neural text-to-SQL parsers, we propose a model uncertainty constraint to refine the query representations by enforcing the output representations of different perturbed encoding networks to be consistent with each other. Extensive experiments on five benchmark datasets demonstrate that our method significantly outperforms strong competitors and achieves new state-of-the-art results.

The importance of building text-to-SQL parsers which can be applied to new databases has long been acknowledged, and a critical step to achieve this goal is schema linking, i.e., properly recognizing mentions of unseen columns or tables when generating SQLs. In this work, we propose a novel framework to elicit relational structures from large-scale pre-trained language models (PLMs) via a probing procedure based on Poincaré distance metric, and use the induced relations to augment current graph-based parsers for better schema linking. Compared with commonly-used rule-based methods for schema linking, we found that probing relations can robustly capture semantic correspondences, even when surface forms of mentions and entities differ. Moreover, our probing procedure is entirely unsupervised and requires no additional parameters. Extensive experiments show that our framework sets new state-of-the-art performance on three benchmarks. We empirically verify that our probing procedure can indeed find desired relational structures through qualitative analysis.

Semantic parsing has long been a fundamental problem in natural language processing. Recently, cross-domain context-dependent semantic parsing has become a new focus of research. Central to the problem is the challenge of leveraging contextual information of both natural language queries and database schemas in the interaction history. In this paper, we present a dynamic graph framework that is capable of effectively modelling contextual utterances, tokens, database schemas, and their complicated interaction as the conversation proceeds. The framework employs a dynamic memory decay mechanism that incorporates inductive bias to integrate enriched contextual relation representation, which is further enhanced with a powerful reranking model. At the time of writing, we demonstrate that the proposed framework outperforms all existing models by large margins, achieving new state-of-the-art performance on two large-scale benchmarks, the SParC and CoSQL datasets. Specifically, the model attains a 55.8% question-match and 30.8% interaction-match accuracy on SParC, and a 46.8% question-match and 17.0% interaction-match accuracy on CoSQL.

This paper proposes Dynamic Memory Induction Networks (DMIN) for few-shot text classification.The model utilizes dynamic routing to provide more flexibility to memory-based few-shot learning in order to better adapt the support sets, which is a critical capacity of fewshot classification models.Based on that, we further develop induction models with query information, aiming to enhance the generalization ability of meta-learning.The proposed model achieves new state-of-the-art results on the miniRCV1 and ODIC dataset, improving the best performance (accuracy) by 2∼4%.Detailed analysis is further performed to show the effectiveness of each component.

Ruiying Geng, Binhua Li, Yongbin Li, Xiaodan Zhu, Ping Jian, Jian Sun. 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.