Yixiao Yang

Wei Ma, Yixiao Yang, Jingquan Ge et al.

Code generation models are widely used in software development, yet their sensitivity to prompt phrasing remains under-examined. Identical requirements expressed with different emotions or communication styles can yield divergent outputs, while most benchmarks emphasize only peak performance. We present PromptSE (Prompt Sensitivity Evaluation), a framework that creates semantically equivalent prompt variants with emotion and personality templates, and that evaluates stability using probability aware continuous scoring or using binary pass rates when logits are unavailable. The results are aggregated into a proposed area under curve metric (AUC-E) for cross model comparison. Across 14 models from three families (Llama, Qwen, and DeepSeek), our study shows that performance and stability behave as largely decoupled optimization objectives, and it reveals architectural and scale related patterns that challenge common assumptions about model robustness. The framework supports rapid screening for closed-source models as well as detailed stability analysis in research settings. PromptSE enables practitioners to quantify performance stability trade offs for deployment and model selection, positioning prompt stability as a complementary evaluation dimension alongside performance and fairness, and contributing to more trustworthy AI-assisted software development tools.

Wei Ma, Yixiao Yang, Qiang Hu et al.

Applications of Large Language Models~(LLMs) have evolved from simple text generators into complex software systems that integrate retrieval augmentation, tool invocation, and multi-turn interactions. Their inherent non-determinism, dynamism, and context dependence pose fundamental challenges for quality assurance. This paper decomposes LLM applications into a three-layer architecture: \textbf{\textit{System Shell Layer}}, \textbf{\textit{Prompt Orchestration Layer}}, and \textbf{\textit{LLM Inference Core}}. We then assess the applicability of traditional software testing methods in each layer: directly applicable at the shell layer, requiring semantic reinterpretation at the orchestration layer, and necessitating paradigm shifts at the inference core. A comparative analysis of Testing AI methods from the software engineering community and safety analysis techniques from the AI community reveals structural disconnects in testing unit abstraction, evaluation metrics, and lifecycle management. We identify four fundamental differences that underlie 6 core challenges. To address these, we propose four types of collaborative strategies (\emph{Retain}, \emph{Translate}, \emph{Integrate}, and \emph{Runtime}) and explore a closed-loop, trustworthy quality assurance framework that combines pre-deployment validation with runtime monitoring. Based on these strategies, we offer practical guidance and a protocol proposal to support the standardization and tooling of LLM application testing. We propose a protocol \textbf{\textit{Agent Interaction Communication Language}} (AICL) that is used to communicate between AI agents. AICL has the test-oriented features and is easily integrated in the current agent framework.

Yixiao Yang, Ran Tao, Kaixuan Wei et al.

Compressive imaging aims to recover a latent image from under-sampled measurements, suffering from a serious ill-posed inverse problem. Recently, deep neural networks have been applied to this problem with superior results, owing to the learned advanced image priors. These approaches, however, require training separate models for different imaging modalities and sampling ratios, leading to overfitting to specific settings. In this paper, a dynamic proximal unrolling network (dubbed DPUNet) was proposed, which can handle a variety of measurement matrices via one single model without retraining. Specifically, DPUNet can exploit both the embedded observation model via gradient descent and imposed image priors by learned dynamic proximal operators, achieving joint reconstruction. A key component of DPUNet is a dynamic proximal mapping module, whose parameters can be dynamically adjusted at the inference stage and make it adapt to different imaging settings. Experimental results demonstrate that the proposed DPUNet can effectively handle multiple compressive imaging modalities under varying sampling ratios and noise levels via only one trained model, and outperform the state-of-the-art approaches.

Yixiao Yang

This paper is written because I receive several inquiry emails saying it is hard to achieve good results when applying token repetition learning techniques. If REP (proposed by me) or Pointer-Mixture (proposed by Jian Li) is directly applied to source code to decide all token repetitions, the model performance will decrease sharply. As we use pre-order traversal to traverse the Abstract Syntax Tree (AST) to generate token sequence, tokens corresponding to AST grammar are ignored when learning token repetition. For non-grammar tokens, there are many kinds: strings, chars, numbers and identifiers. For each kind of tokens, we try to learn its repetition pattern and find that only identifiers have the property of token repetition. For identifiers, there are also many kinds such as variables, package names, method names, simple types, qualified types or qualified names. Actually, some kinds of identifiers such as package names, method names, qualified names or qualified types are unlikely to be repeated. Thus, we ignore these kinds of identifiers that are unlikely to be repeated when learning token repetition. This step is crucial and this important implementation trick is not clearly presented in the paper because we think it is trivial and too many details may bother readers. We offer the GitHub address of our model in our conference paper and readers can check the description and implementation in that repository. Thus, in this paper, we supplement the important implementation optimization details for the already published papers.

Yixiao Yang

In the field of software engineering, applying language models to the token sequence of source code is the state-of-art approach to build a code recommendation system. The syntax tree of source code has hierarchical structures. Ignoring the characteristics of tree structures decreases the model performance. Current LSTM model handles sequential data. The performance of LSTM model will decrease sharply if the noise unseen data is distributed everywhere in the test suite. As code has free naming conventions, it is common for a model trained on one project to encounter many unknown words on another project. If we set many unseen words as UNK just like the solution in natural language processing, the number of UNK will be much greater than the sum of the most frequently appeared words. In an extreme case, just predicting UNK at everywhere may achieve very high prediction accuracy. Thus, such solution cannot reflect the true performance of a model when encountering noise unseen data. In this paper, we only mark a small number of rare words as UNK and show the prediction performance of models under in-project and cross-project evaluation. We propose a novel Hierarchical Language Model (HLM) to improve the robustness of LSTM model to gain the capacity about dealing with the inconsistency of data distribution between training and testing. The newly proposed HLM takes the hierarchical structure of code tree into consideration to predict code. HLM uses BiLSTM to generate embedding for sub-trees according to hierarchies and collects the embedding of sub-trees in context to predict next code. The experiments on inner-project and cross-project data sets indicate that the newly proposed Hierarchical Language Model (HLM) performs better than the state-of-art LSTM model in dealing with the data inconsistency between training and testing and achieves averagely 11.2\% improvement in prediction accuracy.

Yixiao Yang

Language models such as RNN, LSTM or other variants have been widely used as generative models in natural language processing. In last few years, taking source code as natural languages, parsing source code into a token sequence and using a language model such as LSTM to train that sequence are state-of-art methods to get a generative model for solving the problem of code completion. However, for source code with hundreds of statements, traditional LSTM model or attention-based LSTM model failed to capture the long term dependency of source code. In this paper, we propose a novel statement-level language model (SLM) which uses BiLSTM to generate the embedding for each statement. The standard LSTM is adopted in SLM to iterate and accumulate the embedding of each statement in context to help predict next code. The statement level attention mechanism is also adopted in the model. The proposed model SLM is aimed at token level code completion. The experiments on inner-project and cross-project data sets indicate that the newly proposed statement-level language model with attention mechanism (SLM) outperforms all other state-of-art models in token level code completion task.