Search arXiv⌕ Search

arXiv subjects

Zeyu Sun

Publications and source records attributed to Zeyu Sun.

At least 19 recordsLinked to original sources
arXiv2026

TRACE: Evaluating Execution Efficiency of LLM-Based Code Translation

While Large Language Models (LLMs) have substantially improved the functional correctness of code translation, the critical dimension of execution efficiency remains overlooked. We present \textbf{\textsc{Trace}}, the first benchmark to explicitly assess efficiency in LLM-translated code. \textsc{Trace} includes 1,000 efficiency-critical tasks across C++, Java, and Python, each augmented with stress tests that reveal efficiency disparities often overlooked by small-scale tests. Using \textsc{Trace}, we conduct an extensive evaluation of 28 representative LLMs and highlight several key insights: 1) Correctness and efficiency are often misaligned: the correctness leader Claude-Sonnet-4-Think achieves only moderate time efficiency, outperformed by smaller open-source LLMs such as Qwen2.5-Coder-14B-Instruct. 2) Inefficiency is both prevalent and patterned: 23.5\% of correct translations suffer from notable inefficiency, mainly arising from algorithm implementation discrepancy (11.9\%), language construct mismatch (66.4\%), and resource management inefficiency (21.7\%). 3) Inference-time prompt strategies bring only modest improvements, indicating that simple prompting alone is insufficient to improve translation efficiency. Together, our results establish execution efficiency as an essential dimension of code translation and position \textsc{Trace} as a principled foundation for efficiency-oriented evaluation.

cs.SE↗
arXiv2026

DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation

Current data science (DS) code generation benchmarks equate correctness with quality, overlooking execution time differences that span orders of magnitude between correct solutions. We introduce DSEffi-Bench, the first benchmark specifically targeting execution efficiency in LLM-generated DS code, comprising 1,000 instances across 10+ DS libraries with stress-testing harnesses and human-validated references. Evaluating 16 models across 3 tiers, we find that correctness alone fails to characterize efficiency: GPT-5.4 leads in correctness (Pass, 66.9\%) but its efficiency score (B$|$P, 71.7\%) nearly matches GPT-5.4-mini (71.6\%), which solves 47 fewer tasks; Kimi-K2.5 ranks lowest in correctness among frontier models (40.2\%) yet achieves the highest efficiency score (73.6\%) across all 16 models. A human-annotated five-category taxonomy reveals that 79.1\% of efficiency deficits extend beyond algorithmic complexity to domain-specific root causes, with distinct failure profiles across model tiers and libraries. Two exploratory experiments provide initial evidence that these diagnostics can guide improvement, yielding up to +14.7\% efficiency gains via taxonomy-guided optimization and approaching Claude-Opus-4.6 Best@3 in efficiency at 13.0$\times$ lower cost via library-conditioned routing.

cs.SE↗
arXiv2026

KBSpec: LLM-driven Formal Specification Generation with Evolving Domain Knowledge Base

Automated formal specification generation is a key step towards program understanding and formal verification. Recently, due to the success of large language models (LLMs) in code generation, researchers have made early attempts to adopt LLMs for generating formal specifications. However, the lack of formal specification language corpora in the wild often makes LLMs fail to generate syntactically correct and semantically verifiable specifications. To mitigate this gap, we propose KBSpec, which augments LLMs with dual-source knowledge of formal specification languages: external knowledge from official documentation, and internal knowledge distilled from verifier feedback on LLM-generated specifications. KBSpec maintains a self-evolving knowledge base that is continuously updated from successful generation and repair trajectories, without any LLM parameter tuning or labeled training data. We evaluate KBSpec on Java Modeling Language (JML) specification generation with three LLM backends, and the results show that KBSpec improves verification pass rates by 14-32% over state-of-the-art LLM-based approaches, while producing the largest number of high-completeness specifications.

cs.SE↗
arXiv2026

How Powerful are LLMs in Generating Formal Program Specifications?

Formal verification provides strong guarantees of software correctness, but its adoption is limited by the high cost of writing precise formal specifications. While recent large language models (LLMs) have shown strong capabilities in theorem proving and verified code generation, their true ability to generate program specifications remains unclear. Existing evaluations require either verifying implementation conformance or proving semantic equivalence between specifications, both of which are formidably difficult and may conflate proof difficulty with specification quality. To address this problem, we introduce Coins, a Rocq based evaluation framework that assesses specification quality by instantiating specifications under evaluation on trusted test cases and generating concrete proof obligations. This design aligns with the asymmetric nature of formal reasoning, where successful proofs provide reliable evidence while proof failures are inherently ambiguous. Using Coins, we conduct a large scale study on HumanEval with a curated set of human written Rocq specifications. Our results show that specification generation remains a formidable challenge, and that verification complexity can obscure genuine differences in specification quality. Overall, we find that accurate specification evaluation, rather than model scaling alone, is central to understanding the power of LLMs for specification synthesis, and that test case based formal reasoning offers a more faithful and discriminative measure of progress.

cs.SE↗
arXiv2026

Contamination Means Overestimation? A Fine-Grained Empirical Study in Code Intelligence

In recent years, code intelligence has gained increasing importance in the field of automated software engineering. Meanwhile, the widespread adoption of Pretrained Language Models (PLMs) and Large Language Models (LLMs) has raised concerns regarding data contamination and its potential impact on model performance evaluation. Previous studies mainly focused on sample-level contamination, ignoring partial contamination scenarios that are pervasive in code intelligence. This paper fills this gap and presents a systematic empirical study to investigate the fine-grained data contamination on mainstream code tasks. Our study involves diverse representative PLMs: RoBERTa and GPT-2, and LLMs: LLaMA and StarCoder, covering three major tasks: code translation, code generation, and code summarization, across two Programming Languages (PLs): Java and Python. We categorize contamination scenarios into four types according to the code intelligence practice, namely input-only, output-only, unpaired, and paired contamination settings, and construct corresponding experimental and control groups for exploration. Experimental results show that, under the pre-training, fine-tuning, and inference paradigm adopted by PLMs, even deliberately injecting paired contamination does not lead to significant performance overestimation. But direct inference or small-scale fine-tuning uncovers the contamination effects. In contrast, LLMs with pre-training and inference paradigm are significantly affected by the paired contamination. Apart from the above, other contamination scenarios have no impact on both PLMs and LLMs. Our findings challenge the conventional belief that contamination inevitably leads to performance overestimation, providing new insights into the evaluation and deployment of code intelligence models.

cs.SE↗
arXiv2026

WarmTuner: Program-Specific Warm Starts for Compiler Autotuning via Offline-to-Online Reinforcement Learning

Compilers are fundamental software tools that translate high-level programs into machine code. Modern compilers expose hundreds of optimizations, each turned on or off through an optimization flag, to improve the performance of the generated code. However, the number of possible flag combinations grows exponentially, making it difficult to find a flag configuration well suited to a given target program. Existing compiler auto-tuning techniques reduce tuning cost by pruning the search space, injecting search biases, or predicting configuration performance. Although some exploit program features, the knowledge they extract from historical data is frozen once search begins; runtime feedback then guides only the search itself, never the prior. As a result, when this prior mismatches the target program, these methods waste much of the limited online budget before the search reaches good configurations. We propose WarmTuner, an offline-to-online reinforcement learning framework that instead turns historical records into a program-conditioned policy that predicts each flag's setting over the full flag space and remains adaptable on the target program. Offline, WarmTuner learns this program-conditioned policy over the full flag space from historical good configurations. Online, it refines the same policy on the target program using real compile-run feedback, so that the policy is driven by measured speedups rather than limited to the historical data. We instantiate the online update with Group Relative Policy Optimization (GRPO), which compares candidates in the same round and avoids a separate value model. We evaluate WarmTuner on GCC 15.2.0 with cBench and PolyBench. The results show that WarmTuner achieves an average speedup of 1.732x over GCC -O3 and obtains the best result on 14/30 programs, significantly outperforming the compared techniques.

cs.SE↗
arXiv2026

SpecRL: Reinforcement Learning with Test-Based Completeness Rewards for Formal Specification Synthesis

Specification synthesis asks a model to generate specifications and auxiliary annotations for an existing program. In modern software verification projects, specification accuracy is critical. A specification is the abstraction of a method's behavior, and callers are verified against the callee's specification rather than its implementation. As a result, the specification for a callee can strongly affect what properties callers can prove. However, verifier feedback alone is a poor training signal for specification synthesis: verification can prove that a specification is sound for the implementation, yet it cannot tell whether the specification is too weak. We present SpecRL, a reinforcement learning framework that adds an empirical completeness signal to specification synthesis in Dafny. SpecRL constructs negative tests, or spectests, from implementation-impossible input-output pairs that weak specifications such as ensures true may still admit. During training, SpecRL rewards verified candidates according to the fraction of spectests their specifications reject, thereby ranking these candidates by how many implementation-impossible behaviors they rule out. On the out-of-distribution DafnyComp-Spec benchmark, the 7B SpecRL model improves verification success and completeness over supervised fine-tuning by 49.96% and 26.46%, respectively. These relative gains show that fine-grained spectest feedback improves both verifiability and specification accuracy.

cs.SE↗
arXiv2026

WarpagePINN: Thermal Warpage Prediction in Advanced Packaging via a Two-Stage Physics-Informed Neural Networks

Thermal warpage has become a critical issue in advanced packaging, primarily caused by the mismatch in coefficients of thermal expansion (CTE) among heterogeneously integrated materials. However, only a limited number of studies have focused on developing computational methods for coupled thermal-warpage prediction in the chiplet. This paper proposes a two-stage physics-informed neural network (WarpagePINN) framework to compute both temperature profile and warpage deformation of chiplets. The neural networks are trained without relying on labeled datasets generated by conventional simulators. In the first stage, the temperature field is modeled using a Fourier series representation that inherently satisfies boundary conditions, and the network is trained solely through a loss function derived from the governing equation. In the second stage, a multilayer perceptron (MLP) is employed for warpage prediction, utilizing a novel hybrid supervisory strategy to optimize the energy-based loss function instead of residual loss. A parametric WarpagePINN is also developed to quantify uncertainties associated with the CTE. Numerical results show that the proposed WarpagePINN framework achieves excellent agreement with conventional finite element methods, with a mean absolute error (MAE) of 0.2 μm, while achieving a speedup of approximately 1000 {\times} in CTE parameterization studies.

math.NA↗
arXiv2026

Can Large Language Models Reason About Complex Execution Paths? An Empirical Study on Python

Execution path reasoning is a key step towards program semantics understanding. It is crucial for generating test cases that cover certain branches/paths, or detecting bugs that are triggered by some paths without actually executing the program. Traditionally, execution path reasoning can be achieved by symbolic execution techniques, but existing SMT-based symbolic execution approaches struggle with complex data structures and external API calls. This challenge is even more pronounced in languages with highly flexible syntax, such as Python, resulting in a lack of widely adopted tools for reasoning on execution paths. Therefore, reasoning execution paths with AI-based approaches become a promising direction. In this paper, we investigate the feasibility of adopting large language models (LLMs) for execution path reasoning on Python, where traditional path-based symbolic execution tools are unavailable. We conduct an empirical study on two types of path reasoning tasks: generation tasks for test case generation and classification tasks for bug detection. We build new evaluation pipelines and benchmarks from both competition-level programs and real-world repositories. Our results show that state-of-the-art LLMs can perform correct reasoning on execution paths and improve test coverage on real-world software, though models with stronger reasoning abilities do not always outperform weaker ones. These findings highlight the potential of utilizing LLMs as a complementary heuristic for path-aware code reasoning, especially in program languages lacking mature symbolic execution tools. We have released our benchmark and evaluation scripts at https://github.com/jacobwwh/llm-path-study.

cs.SE↗
arXiv2026

Contextualized Code Pretraining for Code Generation

As code generation becomes increasingly central to improving software development efficiency, modern code models are largely trained and evaluated on code with natural-language descriptions. In real projects, developers often implement missing functions under limited project-specific artifacts, while the local call-site context is already available in the surrounding code. This usage context provides actionable cues about expected behavior, but existing models are not explicitly optimized to leverage it reliably, leading to implementations that may not integrate smoothly with surrounding usage in repository settings. In this work, we propose contextualized code pretraining, an invocation-aware framework that integrates calling context into both the training and evaluation of code models. Using static analysis, we automatically extract large-scale caller-callee pairs from real repositories to construct pretraining tasks and benchmarks that condition generation on the calling context. We train CallerGen, the first code models pretrained with invocation-aware objectives spanning multiple sizes, and evaluate them on CallerEval, a new benchmark featuring realistic scenarios. Experiments show that CallerGen outperforms comparable-scale models and remains competitive with larger ones across two benchmarks. Our 220M and 0.5B models achieve 16.58% and 22.81@% pass1, surpassing baselines on CallerEval. These results highlight the importance of calling context in realistic code generation.

cs.SE↗
arXiv2026

TestDecision: Sequential Test Suite Generation via Greedy Optimization and Reinforcement Learning

With the rapid evolution of LLMs, automated software testing is witnessing a paradigm shift. While proprietary models like GPT-4o demonstrate impressive capabilities, their high deployment costs and data privacy concerns make open-source LLMs the practical imperative for many academic and industrial scenarios. In the field of automated test generation, it has evolved to iterative workflows to construct test suites based on LLMs. When utilizing open-source LLMs, we empirically observe they lack a suite-level perspective, suffering from structural myopia-failing to generate new tests with large marginal gain based on the current covered status. In this paper, from the perspective of sequences, we formalize test suite generation as a MDP and demonstrate that its objective exhibits monotone submodularity, which enables an effective relaxation of this NP-hard global optimization into a tractable step-wise greedy procedure. Guided by this insight, we propose TestDecision, which transforms LLMs into neural greedy experts. TestDecision consists of two synergistic components: (1) an inference framework which implements test suite construction following a step-wise greedy strategy; and (2) a training pipeline of reinforcement learning which equips the base LLM with sequential test generation ability to maximize marginal gain. Comprehensive evaluations on the ULT benchmark demonstrate that TestDecision significantly outperforms existing advanced methods. It brings an improvement between 38.15-52.37% in branch coverage and 298.22-558.88% in execution pass rate over all base models, achieving a comparable performance on 7B backbone with a much larger proprietary LLM GPT-5.2. Furthermore, TestDecision can find 58.43-95.45% more bugs than vanilla base LLMs and exhibit superior generalization on LiveCodeBench, proving its capability to construct high-quality test suites.

cs.SE↗
arXiv2026

Bridging the Gap between User Intent and LLM: A Requirement Alignment Approach for Code Generation

Code generation refers to automatically producing executable programs from user requirements. Recently, researchers have explored approaches to enhance the correctness of generated code with advanced large language models. Although achieving improvements, existing approaches focus on designing reasoning strategies or post-refinement methods to enhance code generation performance. Despite their differences, all these methods share a common assumption: the LLM can correctly understand the given requirement. However, this assumption does not always hold. To fill this gap, we propose REA-Coder, a requirement alignment approach to enhance the code generation performance of LLMs. REA-Coder involves first identifying the requirement content that does not align with LLMs and aligning the requirements. Then, based on the aligned requirements, LLMs generate code and further verify whether the generated code aligns with the requirements, iterating this process of requirement alignment and code generation until generating correct code or achieving the maximum number of iterations. Experimental results show that REA-Coder outperforms all advanced baselines on four LLMs across five programming benchmarks. Concretely, REA-Coder achieves average improvements of 7.93%, 30.25%, 26.75%, 8.59%, and 8.64% on the five benchmark datasets, demonstrating the effectiveness of requirement alignment for improving the code generation performance of LLMs.

cs.SE↗
arXiv2026

TRACE: Evaluating Execution Efficiency of LLM-Based Code Translation

While Large Language Models (LLMs) have substantially improved the functional correctness of code translation, the critical dimension of \textit{execution efficiency} remains overlooked. We present \textbf{\textsc{trace}}, the first benchmark to explicitly assess efficiency in LLM-translated code. \textsc{trace} includes 1,000 efficiency-critical tasks across C++, Java, and Python, each augmented with stress tests that reveal efficiency degradations often overlooked by small-scale tests. Using \textsc{trace}, we conduct an extensive evaluation of 28 representative LLMs and highlight several key insights: 1) Correctness is not a reliable proxy for efficiency: the correctness leader \textit{Claude-4-think} achieves only mid-level time efficiency, outperformed by smaller open-source LLMs such as \textit{Qwen2.5-Coder-14B-Instruct}. 2) Inefficiency is both prevalent and patterned: 23.5\% of correct translations exhibit pronounced inefficiency, distributed across algorithmic faults (11.9\%), language construct mismatches (66.4\%), and resource mismanagement (21.7\%). 3) Inference-time prompt strategies bring only modest improvements, suggesting that current LLMs lack intrinsic efficiency awareness. Together, our results establish efficiency as an essential dimension of code translation and position \textsc{trace} as a principled foundation for efficiency-oriented evaluation.

cs.SE↗
arXiv2026

Search-Induced Issues in Web-Augmented LLM Code Generation: Detecting and Repairing Error-Inducing Pages

Web-augmented large language models (LLMs) offer promising capabilities for automatic code generation. However, integrating live web search exposes models to unreliable or malicious content, leading to Search-Induced Issues (SII), a novel failure mode in which external pages mislead LLMs into producing incorrect code. This paper presents a comprehensive empirical study of the prevalence and impact of SII across three commercial search APIs and six advanced LLMs. Our analysis reveals that all evaluated web-augmented LLMs are vulnerable to SII, with root causes arising from either misaligned specifications or flawed code implementations in the searched Error-Inducing Pages (EIPs). To address this challenge, we propose Sherlock, an automated framework that enables LLM service providers to proactively safeguard web-augmented generation systems at scale. Sherlock operates as a continuous pipeline that first detects potential SII instances, then debugs them to identify the responsible EIPs and pinpoint their root causes, and finally repairs them by either annotating misaligned content or replacing erroneous code snippets with evaluated solutions from trusted sources. Experiments show that Sherlock identifies EIPs with an F1 score of up to 95% and repairs 71% to 100% of affected generations across the evaluated models, with modest computational overhead. Our findings and framework provide practical guidance for improving the reliability of web-augmented LLM-based code generation systems in real-world software engineering scenarios.

cs.SE↗
arXiv2026

ClimaOoD: Improving Anomaly Segmentation via Physically Realistic Synthetic Data

Anomaly segmentation seeks to detect and localize unknown or out-of-distribution (OoD) objects that fall outside predefined semantic classes a capability essential for safe autonomous driving. However, the scarcity and limited diversity of anomaly data severely constrain model generalization in open-world environments. Existing approaches mitigate this issue through synthetic data generation, either by copy-pasting external objects into driving scenes or by leveraging text-to-image diffusion models to inpaint anomalous regions. While these methods improve anomaly diversity, they often lack contextual coherence and physical realism, resulting in domain gaps between synthetic and real data. In this paper, we present ClimaDrive, a semantics-guided image-to-image framework for synthesizing semantically coherent, weather-diverse, and physically plausible OoD driving data. ClimaDrive unifies structure-guided multi-weather generation with prompt-driven anomaly inpainting, enabling the creation of visually realistic training data. Based on this framework, we construct ClimaOoD, a large-scale benchmark spanning six representative driving scenarios under both clear and adverse weather conditions. Extensive experiments on four state-of-the-art methods show that training with ClimaOoD leads to robust improvements in anomaly segmentation. Across all methods, AUROC, AP, and FPR95 show notable gains, with FPR95 dropping from 3.97 to 3.52 for RbA on Fishyscapes LAF. These results demonstrate that ClimaOoD enhances model robustness, offering valuable training data for better generalization in open-world anomaly detection.

cs.CV↗
arXiv2026

LAURA: Enhancing Code Review Generation with Context-Enriched Retrieval-Augmented LLM

Code review is critical for ensuring software quality and maintainability. With the rapid growth in software scale and complexity, code review has become a bottleneck in the development process because of its time-consuming and knowledge-intensive nature and the shortage of experienced developers willing to review code. Several approaches have been proposed for automatically generating code reviews based on retrieval, neural machine translation, pre-trained models, or large language models (LLMs). These approaches mainly leverage historical code changes and review comments. However, a large amount of crucial information for code review, such as the context of code changes and prior review knowledge, has been overlooked. This paper proposes an LLM-based review knowledge-augmented, context-aware framework for code review generation, named LAURA. The framework integrates review exemplar retrieval, context augmentation, and systematic guidance to enhance the performance of ChatGPT-4o and DeepSeek v3 in generating code review comments. Besides, given the extensive low-quality reviews in existing datasets, we also constructed a high-quality dataset. Experimental results show that for both models, LAURA generates review comments that are either completely correct or at least helpful to developers in 42.2% and 40.4% of cases, respectively, significantly outperforming SOTA baselines. Furthermore, our ablation studies demonstrate that all components of LAURA contribute positively to improving comment quality.

cs.SE↗
arXiv2026

Learning to Guarantee Type Correctness in Code Generation through Type-Guided Program Synthesis

Language models have shown remarkable proficiency in code generation; nevertheless, ensuring type correctness remains a challenge. Although traditional methods, such as constrained decoding, alleviate this problem by externally rejecting untypable code, the model itself does not effectively learn type reasoning internally, which ultimately limits its overall performance. This paper introduces TyFlow, a novel system that internalizes type reasoning within code generation to guide the model to learn the type system. The core of our approach is a novel type-guided program synthesis system that maintains an isomorphism between type derivation trees and synthesis derivation trees, enabling a new code representation based on synthesis decision sequences rather than traditional text-based token sequences. By offloading the complexity of type system learning to the representation itself, models can redirect their computational resources toward higher-level program semantics. Our evaluation shows that TyFlow not only eliminates type errors but also significantly improves functional correctness, highlighting the importance of aligning LMs with type systems internally.

cs.PL↗
arXiv2025

CupCleaner: A Hybrid Data Cleaning Approach for Comment Updating

Comment updating is an emerging task in software evolution that aims to automatically revise source code comments in accordance with code changes. This task plays a vital role in maintaining code-comment consistency throughout software development. Recently, deep learning-based approaches have shown great potential in addressing comment updating by learning complex patterns between code edits and corresponding comment modifications. However, the effectiveness of these learning-based approaches heavily depends on the quality of training data. Existing datasets are typically constructed by mining version histories from open-source repositories such as GitHub, where there is often a lack of quality control over comment edits. As a result, these datasets may contain noisy or inconsistent samples that hinder model learning and generalization. In this paper, we focus on cleaning existing comment updating datasets, considering both the data's characteristics in the updating scenario and their implications on the model training process. We propose a hybrid statistical approach named CupCleaner (Comment UPdating's CLEANER) to achieve this purpose. Specifically, we combine static semantic information within data samples and dynamic loss information during the training process to clean the dataset. Experimental results demonstrate that, on the same test set, both the individual static strategy and the dynamic strategy can significantly filter out a portion of the data and enhance the performance of the model. Furthermore, employing a model ensemble approach can combine the characteristics of static and dynamic cleaning, further enhancing the performance of the model and the reliability of its output results.

cs.SE↗