Biological systems continually incorporate new experiences while preserving prior knowledge through coordinated interactions across distributed memory systems. Artificial neural networks, by contrast, typically address continual learning by reducing interference, often through replay buffers or parameter isolation. Although effective for limiting forgetting, these strategies do not explicitly support the integration of new information into existing knowledge. Here we introduce the Cascaded Memory Network (CaMNet), a neuro-inspired continual learning framework that organizes memory into a progressive cascade of modules and uses compact cues to trigger targeted reinstatement of prior representations. Instead of retraining on stored samples, CaMNet stabilizes task-relevant knowledge by coordinating internal memory states and enforcing consistency across the cascade. Across continual vision–language and reinforcement learning benchmarks, CaMNet improves both retention and forward transfer relative to strong continual learning baselines. CaMNet also enables cue-driven recovery of latent memory traces after substantial performance degradation, indicating that previously acquired knowledge can be functionally re-engaged without revisiting prior inputs. In dynamic physical environments, the framework allows a humanoid robot to rapidly adapt to perceptual shifts and bridge the sim-to-real gap while preserving pre-trained capabilities. These results identify cascaded memory coordination as an effective computational principle for continual generalization and replay-free continual learning across domains.

Code generation has emerged as a valuable technique for accelerating robotic software development. However, unlike general-purpose programming, generating code for Cyber-Physical Systems (CPS) requires navigating complex, domain-specific frameworks like the Robot Operating System (ROS). Although Large Language Models (LLMs) excel in general-purpose code generation, their capability to synthesize executable, middleware-compliant ROS code remains poorly understood. This task demands that LLMs not only solve robotic tasks but also generate code that seamlessly integrates within the distributed ROS framework. To address this gap, we construct ROSDevEval, a specialized benchmark comprising 240 real-world ROS programming tasks, and conduct a comprehensive empirical study to evaluate four state-of-the-art LLMs alongside a specialized code assistant (GitHub Copilot). Our results reveal a severe domain capability gap: while state-of-the-art LLMs achieve up to 88.4 https://github.com/zyx065/rosdeveval .

Code differencing is a fundamental technique in software engineering practice and research. While researchers have proposed text-based differencing techniques capable of identifying line changes over the past decade, existing methods exhibit a notable limitation in identifying edit actions (EAs) that operate on text blocks spanning multiple lines. Such EAs are common in developers' practice, such as moving a code block for conditional branching or duplicating a method definition block for overloading. Existing tools represent such block-level operations as discrete sequences of line-level EAs, compelling developers to manually correlate them and thereby substantially impeding the efficiency of change comprehension. To address this issue, we propose BDiff, a text-based differencing algorithm capable of identifying two types of block-level EAs and five types of line-level EAs. Building on traditional differencing algorithms, we first construct a candidate set containing all possible line mappings and block mappings. Leveraging the Kuhn-Munkres algorithm, we then compute the optimal mapping set that can minimize the size of the edit script (ES) while closely aligning with the original developer's intent. To validate the effectiveness of BDiff, we selected five state-of-the-art tools, including large language models (LLMs), as baselines and adopted a combined qualitative and quantitative approach to evaluate their performance in terms of ES size, result quality, and running time. Experimental results show that BDiff produces higher-quality differencing results than baseline tools while maintaining competitive runtime performance. Our experiments also show the unreliability of LLMs in code differencing tasks regarding result quality and their infeasibility in terms of runtime efficiency. We have implemented a web-based visual differencing tool.

Exemplar-free Class-Incremental Learning (EFCIL) poses a significant challenge in mitigating catastrophic forgetting, due to the absence of exemplars. Recently, analytic learning-based methods propose a recursive alignment procedure to execute EFCIL in a phase-invariant manner and show state-of-the-art performance. However, they heavily rely on a frozen feature extractor trained with the initial dataset to avoid the misalignment between feature and label spaces, ignoring the importance of acquiring generalizable features across incremental tasks for performance improvement. To tackle this, we rethink the obscured sub-optimality of analytic learning-based methods, particularly through empirical reevaluation, and then introduce the Multi-head analytic learning (Muheal) approach. Muheal forms the multi-head model with a delicate feature extractor, thereby introducing a feature optimization procedure and a forgetting compensation module to balance the learning and forgetting. Specifically, within the feature optimization procedure, the feature extractor seeks to learn more generalizable features in a self-supervised manner using the fully-connected classification head. An analytic learning-based classification head follows to align the feature-label space. Additionally, we employ the compensation module to generate and align pseudo-features with a replicated analytic head, thus preventing overfitting and testing. Comprehensive experiments on several benchmark datasets have demonstrated that Muheal significantly outperforms existing state-of-the-art EFCIL methods and is comparable, if not superior, to methods that use replay techniques.

Code adaptation is a fundamental but challenging task in software development, requiring developers to modify existing code for new contexts. A key challenge is to resolve Context Adaptation Bugs (CtxBugs), which occurs when code correct in its original context violates constraints in the target environment. Unlike isolated bugs, CtxBugs cannot be resolved through local fixes and require cross-context reasoning to identify semantic mismatches. Overlooking them may lead to critical failures in adaptation. Although Large Language Models (LLMs) show great potential in automating code-related tasks, their ability to resolve CtxBugs remains a significant and unexplored obstacle to their practical use in code adaptation. To bridge this gap, we propose CtxBugGen, a novel framework for generating CtxBugs to evaluate LLMs. Its core idea is to leverage LLMs' tendency to generate plausible but context-free code when contextual constraints are absent. The framework generates CtxBugs through a four-step process to ensure their relevance and validity: (1) Adaptation Task Selection, (2) Task-specific Perturbation,(3) LLM-based Variant Generation and (4) CtxBugs Identification. Based on the benchmark constructed by CtxBugGen, we conduct an empirical study with four state-of-the-art LLMs. Our results reveal their unsatisfactory performance in CtxBug resolution. The best performing LLM, Kimi-K2, achieves 55.93

All-reduce is a widely used communication technique for distributed and parallel applications typically implemented using either a tree-based or ring-based scheme. Each of these approaches has its own limitations: tree-based schemes struggle with efficiently exchanging large messages, while ring-based solutions assume constant communication throughput, an unrealistic expectation in modern network communication infrastructures. We present FMCC-RT, an all-reduce approach that combines the advantages of tree- and ring-based implementations while mitigating their drawbacks. FMCC-RT dynamically switches between tree and ring-based implementations depending on the size of the message being processed. It utilizes an analytical model to assess the impact of message sizes on the achieved throughput, enabling the derivation of optimal work partitioning parameters. Furthermore, FMCC-RT is designed with an OpenMPI-compatible API, requiring no modification to user code. We evaluated FMCC-RT through micro-benchmarks and real-world application tests. Experimental results show that FMCC-RT outperforms state-of-the-art tree- and ring-based methods, achieving speedups of up to 5.6×.

The recognition of Application Programming Interface (API) mentions in software-related texts is vital for extracting API-related knowledge, providing deep insights into API usage and enhancing productivity efficiency. Previous research identifies two primary technical challenges in this task: (1) differentiating APIs from common words and (2) identifying morphological variants of standard APIs. While deep learning-based methods have demonstrated advancements in addressing these challenges, they rely heavily on high-quality labeled data, leading to another significant data-related challenge: (3) the lack of such high-quality data due to the substantial effort required for labeling. To overcome these challenges, this paper proposes a context-aware API recognition method named CARLDA. This approach utilizes two key components, namely, Bidirectional Encoder Representations from Transformers (BERT) and Bidirectional Long Short-Term Memory (BiLSTM), to extract context at both the word and sequence levels, capturing syntactic and semantic information to address the first challenge. For the second challenge, it incorporates a character-level BiLSTM with an attention mechanism to grasp global character-level context, enhancing the recognition of morphological features of APIs. To address the third challenge, we developed specialized data augmentation techniques using large language models (LLMs) to tackle both in-library and cross-library data shortages. These techniques generate a variety of labeled samples through targeted transformations (e.g., replacing tokens and restructuring sentences) and hybrid augmentation strategies (e.g., combining real-world and generated data while applying style rules to replicate authentic programming contexts). Given the uncertainty about the quality of LLM-generated samples, we also developed sample selection algorithms to filter out low-quality samples (i.e., incomplete or incorrectly labeled samples). Moreover, specific datasets have been constructed to evaluate CARLDA's ability to address the aforementioned challenges. Experimental results demonstrate that (1) CARLDA significantly enhances F1 by 11.0% and the Matthews correlation coefficient (MCC) by 10.0% compared to state-of-the-art methods, showing superior overall performance and effectively tackling the first two challenges, and (2) LLM-based data augmentation techniques successfully yield high-quality labeled data and effectively alleviate the third challenge.

Code snippet adaptation is a fundamental activity in the software development process. Unlike code generation, code snippet adaptation is not a “free creation”, which requires developers to tailor a given code snippet in order to fit specific requirements and the code context. Recently, large language models (LLMs) have confirmed their effectiveness in the code generation task with promising results. However, their performance on code snippet adaptation, a reuse-oriented and context-dependent code change prediction task, is still unclear. To bridge this gap, we conduct an empirical study to investigate the performance and issues of LLMs on the adaptation task. We first evaluate the adaptation performances of three popular LLMs and compare them to the code generation task. Our result indicates that their adaptation ability is weaker than generation, with a nearly 15% decrease on pass@1 and more context-related errors. By manually inspecting 200 cases, we further investigate the causes of LLMs' sub-optimal performance, which can be classified into three categories, i.e., Unclear Requirement, Requirement Misalignment and Context Misapplication. Based on the above empirical research, we propose an interactive prompting approach to eliciting LLMs' ability on the adaptation task. Specifically, we enhance the prompt by enriching the context and decomposing the task, which alleviates context misapplication and improves requirement understanding. Besides, we enable LLMs' reflection by requiring them to interact with a human or a LLM counselor, compensating for unclear requirement. Our experimental result reveals that our approach greatly improve LLMs' adaptation performance. The best-performing Human-LLM interaction successfully solves 159 out of the 202 identified defects and improves the pass@1 and pass@5 by over 40% compared to the initial instruction-based prompt. Considering human efforts, we suggest multi-agent interaction as a trade-off, which can achieve comparable performance with excellent generalization ability. We deem that our approach could provide methodological assistance for autonomous code snippet reuse and adaptation with LLMs.

Recent advancements in large language models (LLMs) have automated various software engineering tasks, with benchmarks emerging to evaluate their capabilities. However, for adaptation, a critical activity during code reuse, there is no benchmark to assess LLMs' performance, leaving their practical utility in this area unclear. To fill this gap, we propose AdaptEval, a benchmark designed to evaluate LLMs on code snippet adaptation. Unlike existing benchmarks, AdaptEval incorporates the following three distinctive features: First, Practical Context. Tasks in AdaptEval are derived from developers' practices, preserving rich contextual information from Stack Overflow and GitHub communities. Second, Multi-granularity Annotation. Each task is annotated with requirements at both task and adaptation levels, supporting the evaluation of LLMs across diverse adaptation scenarios. Third, Fine-grained Evaluation. AdaptEval includes a two-tier testing framework combining adaptation-level and function-level tests, which enables evaluating LLMs' performance across various individual adaptations. Based on AdaptEval, we conduct the first empirical study to evaluate six instruction-tuned LLMs and especially three reasoning LLMs on code snippet adaptation. Experimental results demonstrate that AdaptEval enables the assessment of LLMs' adaptation capabilities from various perspectives. It also provides critical insights into their current limitations, particularly their struggle to follow explicit instructions. We hope AdaptEval can facilitate further investigation and enhancement of LLMs' capabilities in code snippet adaptation, supporting their real-world applications.

Extracting API knowledge from Stack Overflow has become a crucial way to assist developers in using APIs. Existing research has primarily focused on extracting relevant API-related knowledge at the sentence level to enhance API documentation. However, this level of extraction can lead to a loss of crucial context, especially when sentences contain context-dependent entities (i.e., whose understanding requires reference to the surrounding context) that may hinder developers' understanding. To investigate this issue, we conducted an empirical study of 384 Stack Overflow posts and found that (1) approximately one-third of API functionality sentences contain context-dependent entities, and (2) these entities fall into two categories: Referential ContextDependent Entities and Local Variable Context-Dependent Entities. In response, we developed a novel method, CEDCC, which combines an entity filtering strategy informed by insights from our empirical study, with a large language model (LLM) to construct coreference chains for detecting context-dependent entities. Additionally, it employs a step-by-step approach with the LLM to complete the necessary context for understanding these entities. To evaluate CEDCC, we constructed a dataset of 1,023 API knowledge sentences, including 567 context-dependent entities and their required contexts. The results demonstrate the effectiveness of CEDCC in accurately detecting contextdependent entities and completing context tasks, achieving an F1score of 0.865 and a BERTScore of 0.373, significantly surpassing the baseline methods. Human evaluations further confirmed that CEDCC effectively improves the comprehensibility of API knowledge sentences.

Serverless computing is a novel cloud computing paradigm that enables developers to develop, deploy, and run applications in the cloud without complex and error-prone cloud resource management. However, its characteristics also introduce new types of faults (e.g., faults due to insufficient computing resource allocation) and challenges to serverless computing-based applications (abbreviated as serverless applications). While prior studies have highlighted that serverless developers encounter various challenges, no attempts have been made to understand the faults in serverless applications. These faults may cause catastrophic consequences such as application crash, thereby hindering the further spread of serverless computing. We aim in this paper to understand the symptoms, root causes, and fix patterns of faults in serverless applications. To this end, we conduct an empirical study investigating developers' issues on GitHub and posts on Stack Overflow (SO). We first identify 546 real-world serverless-related faults from GitHub and SO. Then, we manually analyze and construct taxonomies of the symptoms, root causes, and fix patterns for these faults, respectively. Our study leads to the first taxonomy for symptoms of serverlessrelated faults, covering 5 categories and 21 subcategories. The findings of our study inform that the Permission Denied error is the most common type ($\mathbf{1 0. 8 1 \%}$) of faults. Furthermore, the Incorrect Code Logic is the main cause ($\mathbf{1 7. 9 5 \%}$) behind the faults. Furthermore, we summarize 15 fix patterns that can resolve $\mathbf{7 3. 6 3 \%}$ of faults in this study. Based on the results, we provide actionable implications that can potentially facilitate research and assist developers in improving the development of serverless applications. Finally, we implement a knowledge-based Q&A tool named SafHelper to help developers understand and fix faults.

Logit-based knowledge distillation (KD) is commonly used to mitigate catastrophic forgetting in class-incremental learning (CIL) caused by data distribution shifts. However, the strict match of logit values between student and teacher models conflicts with the cross-entropy (CE) loss objective of learning new classes, leading to significant recency bias (i.e. unfairness). To address this issue, we rethink the overlooked limitations of KD-based methods through empirical analysis. Inspired by our findings, we introduce a plug-and-play pre-process method that normalizes the logits of both the student and teacher across all classes, rather than just the old classes, before distillation. This approach allows the student to focus on both old and new classes, capturing intrinsic inter-class relations from the teacher. By doing so, our method avoids the inherent conflict between KD and CE, maintaining fairness between old and new classes. Additionally, recognizing that overconfident teacher predictions can hinder the transfer of inter-class relations (i.e., dark knowledge), we extend our method to capture intra-class relations among different instances, ensuring fairness within old classes. Our method integrates seamlessly with existing logit-based KD approaches, consistently enhancing their performance across multiple CIL benchmarks without incurring additional training costs.

Class-Incremental Learning (CIL) enables models to continuously learn new classes while mitigating catastrophic forgetting. Recently, Pre-Trained Models (PTMs) have greatly enhanced CIL performance, even when fine-tuning is limited to the first task. This advantage is particularly beneficial for CIL methods that freeze the feature extractor after first-task fine-tuning, such as analytic learning-based approaches using a least squares solution-based classification head to acquire knowledge recursively. In this work, we revisit the analytical learning approach combined with PTMs and identify its limitations in adapting to new classes, leading to sub-optimal performance. To address this, we propose the Momentum-based Analytical Learning (MoAL) approach. MoAL achieves robust knowledge memorization via an analytical classification head and improves adaptivity to new classes through momentum-based adapter weight interpolation, leading to forgetting outdated knowledge. Importantly, we introduce a knowledge rumination mechanism that leverages refined adaptivity, allowing the model to revisit and reinforce old knowledge, thereby improving performance on old classes. MoAL facilitates the acquisition of new knowledge and consolidates old knowledge, achieving a win-win outcome between plasticity and stability. Extensive experiments on various incremental settings show MoAL’s state-of-the-art performance1.

Pre-trained code models are essential for various code intelligence tasks. Yet, their effectiveness is heavily influenced by the quality of the pre-training dataset, particularly human-written reference comments, which usually serve as a bridge between the programming language and natural language. One significant challenge is that such comments could become inconsistent with the corresponding code as the software evolves, leading to suboptimal model performance. Large language models (LLMs) have demonstrated superior capabilities in generating high-quality code comments. This work investigates whether substituting original human-written comments with LLM-generated ones can improve pre-training datasets for more effective pretrained code models. As existing reference-based metrics cannot evaluate the quality of human-written reference comments themselves, to enable direct comparison between LLM-generated and human reference comments, we introduce two auxiliary tasks as novel reference-free metrics, including code-comment inconsistency detection and semantic code search. Experimental results show that LLM-generated comments exhibit superior semantic consistency with the code compared to human-written reference comments. Our manual evaluation also corroborates this conclusion, which indicates the potential of utilizing LLMs to enhance the quality of the pre-training dataset. Based on this finding, we rebuilt the CodeSearchNet dataset with LLM-generated comments and re-pre-trained the CodeT5 model. Evaluations on multiple code intelligence tasks demonstrate that models pretrained by LLM-enhanced data outperform their counterparts (pre-trained by original human reference comments data) on code summarization, code generation, and code translation tasks. This research validates the feasibility of rebuilding the pre-training dataset by LLMs to advance code intelligence tasks. It advocates rethinking the reliance on human reference comments for code-related tasks.

Class-incremental learning (CIL) aims to learn a family of classes incrementally with data available in order rather than training all data at once. One main drawback of CIL is that standard deep neural networks suffer from catastrophic forgetting (CF), especially when the model only has access to data from the current incremental step. Knowledge Distillation (KD) is a widely used technique that utilizes old models as the teacher model to alleviate CF. However, based on a case study, our investigation reveals that the vanilla KD is insufficient with a strict point-to-point restriction. Instead, a relaxed match between the teacher and student improves distillation performance and model stability. In this article, we propose a simple yet effective method to mitigate CF without any additional training costs or requiring any exemplars. Specifically, we apply the linear correlation between the features of the teacher and student to measure the distillation loss rather than vanilla point-to-point loss, which significantly improves the model stability. Then, we utilize label augmentation to improve feature generalization and save prototypes to alleviate classification bias further. The proposed method significantly outperforms state-of-the-art methods in the various settings of benchmarks, including CIFAR-100 and Tiny-ImageNet, demonstrating its effectiveness and robustness.

Reusing code snippets from online programming Q&A communities has become a common development practice, in which developers often need to adapt code snippets to their code contexts to satisfy their own programming needs. However, how developers make these code adaptations based on contexts is still unclear. To bridge this gap, we first conduct a semi-structured interview of 21 developers to investigate their adaptation practices and perceived challenges during this process. The result suggests that code snippet adaptation is a challenging and exhausting task for developers, as they should tailor the snippets to guarantee their correctness and quality with laborious work. We also note that developers all resort to their intra-file context to complete adaptations, which motivates us to further study how developers performed context-based adaptations (CAs) in real scenarios. To this end, we conduct a quantitative study on an adaptation dataset comprising 300 code snippet reuse cases with 1,384 adaptations from Stack Overflow to GitHub. For each adaptation, we manually annotate its intention and relationship with the context. Based on our annotated data, we employ frequent itemset mining to obtain four CA patterns from our dataset, including Fortification , Code Wiring , Attribute-ization and Parameterization . Our main findings reveal that: (1) more than half of the code snippet reuse cases include CAs and 23.3% of the adaptations are CAs; (2) more than half of the CAs are corrective adaptations and variable is the primary adapted language construct; (3) attribute is the most frequently utilized context and 88% of the local contexts are within the nearest 10 LOCs; and (4) CAs towards different intentions are repetitive, which are useful for automatic adaptation. Overall, our study provides valuable insights into code snippet adaptation and has important implications for research, practice, and tool design.

The sparsity of extrinsic rewards presents a significant challenge for deep reinforcement learning (DRL). As an alternative, researchers have focused on intrinsic rewards to improve exploration efficiency. One of the most representative approaches is utilizing curiosity. However, the challenge of designing effective intrinsic rewards remains, as artificial curiosity differs significantly from human curiosity. In this article, we introduce a novel curiosity approach for DRL, named DyMeCu, which stands for Dynamic Memory-based Curiosity. Inspired by human curiosity and information theory, DyMeCu constructs a dynamic memory using the online learner following the bootstrap paradigm. Additionally, we design a two-learner architecture inspired by ensemble techniques to access curiosity better. The information gap between the two learners serves as the intrinsic reward for agents, and the state information is consistently consolidated into the dynamic memory. Compared with previous curiosity methods, DyMeCu can better mimic human curiosity by using a dynamic memory that can be dynamically grown based on a bootstrap paradigm with two learners. Large-scale empirical experiments on multiple benchmarks, including DeepMind Control Suite and Atari Suite, demonstrate that DyMeCu outperforms competitive curiosity-based methods with or without extrinsic rewards. In the Atari Suite, DyMeCu achieves a mean human-normalized score of 5.076 on a subset of 26 Atari games, achieving a 77.4% relative improvement over the best other baselines. In the DeepMind Control Suite, DyMeCu presents new state-of-the-art results across 11 tasks of all 12 when compared to curiosity-based methods and other pre-training strategies.

Under sparse extrinsic reward settings, reinforcement learning has remained challenging, despite surging interests in this field. Previous attempts suggest that intrinsic reward can alleviate the issue caused by sparsity. In this article, we present a novel intrinsic reward that is inspired by human learning, as humans evaluate curiosity by comparing current observations with historical knowledge. Our method involves training a self-supervised prediction model, saving snapshots of the model parameters, and using nuclear norm to evaluate the temporal inconsistency between the predictions of different snapshots as intrinsic rewards. We also propose a variational weighting mechanism to assign weight to different snapshots in an adaptive manner. Our experimental results on various benchmark environments demonstrate the efficacy of our method, which outperforms other intrinsic reward-based methods without additional training costs and with higher noise tolerance. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.