arXiv subjects

Mohammad Abdollahi

Publications and source records attributed to Mohammad Abdollahi.

3 recordsLinked to original sources

arXiv2026

Can LLMs Reason About Runtime Behavior? A Repository-Level Dynamic Benchmark

Large language models (LLMs) are increasingly used in coding tasks, but their ability to reason about code execution remains unclear. Existing repository-level QA benchmarks mainly evaluate static code understanding and often rely on LLM-based evaluation, while execution-reasoning benchmarks are mostly limited to snippets or functions. We introduce SWE-Flux, a repository-level benchmark for dynamic execution reasoning containing 480 execution-grounded instances across 12 real Python repositories, with gold answers automatically harvested from instrumented test executions rather than written manually or judged by LLMs. The benchmark covers singletest and multi-test questions over control flow, loops, program state, dataflow, exceptions, and program invariants. Evaluating five LLMs shows that this task remains challenging. The best model achieves only 37% accuracy. Models perform better on localized behavior such as invariants, intra-procedural control flow, exceptions, and simple loops, but struggle with dataflow, inter-procedural execution, precise state reasoning, and suite-level aggregation. Finally, we show that the oracle-harvesting pipeline can generate fresh benchmark variants using input perturbation. It successfully harvests valid variants for almost 90% of the selected instances, and the resulting variants are substantially more challenging for the evaluated models.

cs.SE↗

arXiv2025

Demystifying Errors in LLM Reasoning Traces: An Empirical Study of Code Execution Simulation

Understanding a program's runtime reasoning behavior, meaning how intermediate states and control flows lead to final execution results, is essential for reliable code generation, debugging, and automated reasoning. Although large language models (LLMs) can accurately predict program outputs, most prior work has focused on output accuracy and performance, treating reasoning as a black box. As a result, little is known about the structure or failure modes of their reasoning traces. To address this gap, we conduct the first empirical study on runtime behavior inference with reasoning LLMs, aiming to uncover and characterize errors in their reasoning traces. We curate a benchmark from HumanEval Plus and LiveCodeBench, containing 427 code snippets. For each snippet, we test three input types: regular, edge, and invalid. Twelve input values are selected per snippet, each paired with its ground-truth execution result. We evaluate four state-of-the-art reasoning LLMs. Our results show that these models reach accuracies between 85 percent and 98 percent across input types. We also analyze the produced reasoning traces and develop a taxonomy with nine categories of inference errors. Finally, we explore tool-augmented reasoning. Using failures in the Computation Errors category as a case study, our experiments show that this approach corrects 58 percent of such errors, demonstrating the potential of tool support for improving LLM reasoning.

cs.SE↗

arXiv2025

Deep-Bench: Deep Learning Benchmark Dataset for Code Generation

Deep learning (DL) has revolutionized areas such as computer vision, natural language processing, and more. However, developing DL systems is challenging due to the complexity of DL workflows. Large Language Models (LLMs), such as GPT, Claude, Llama, Mistral, etc., have emerged as promising tools to assist in DL code generation, offering potential solutions to these challenges. Despite this, existing benchmarks such as DS-1000 are limited, as they primarily focus on small DL code snippets related to pre/post-processing tasks and lack a comprehensive coverage of the full DL pipeline, including different DL phases and input data types. To address this, we introduce DeepBench, a novel benchmark dataset designed for function-level DL code generation. DeepBench categorizes DL problems based on three key aspects: phases such as pre-processing, model construction, and training; tasks, including classification, regression, and recommendation; and input data types such as tabular, image, and text. GPT-4o -- the state-of-the-art LLM -- achieved 31% accuracy on DeepBench, significantly lower than its 60% on DS-1000. We observed similar difficulty for other LLMs (e.g., 28% vs. 54% for Claude, 21% vs. 41% for LLaMA, and 15% vs. 20% for Mistral). This result underscores DeepBench's greater complexity. We also construct a taxonomy of issues and bugs found in LLM-generated DL code, which highlights the distinct challenges that LLMs face when generating DL code compared to general code. Furthermore, our analysis also reveals substantial performance variations across categories, with differences of up to 7% among phases and 37% among tasks. These disparities suggest that DeepBench offers valuable insights into the LLMs' performance and areas for potential improvement in the DL domain.

cs.SE↗