Laboni Sarker

Laboni Sarker, Mara Downing, Achintya Desai et al.

Rapid advances in the field of Large Language Models (LLMs) have made LLM-based code generation an important area for investigation. An LLM-based code generator takes a prompt as input and produces code that implements the requirements specified in the prompt. Many software requirements include mathematical formulas that specify the expected behavior of the code to be generated. Given a code generation prompt that contains a mathematical formula, a reasonable expectation is that, if the formula is syntactically modified without changing its semantics, the generated code for the modified prompt should be semantically equivalent. We formalize this concept as syntactic robustness and investigate the syntactic robustness of LLMs as code generators. Our experimental assessment demonstrates that LLMs are not syntactically robust for code generation prompts with formulas, especially for the ones that require mathematical reasoning. We investigate attack strategies that can further deteriorate the syntactic robustness of LLMs. Finally, to mitigate syntactic robustness failures in LLMs, we propose a pre-processing step that uses reductions to transform formulas in prompts to a simplified form. Our experimental results demonstrate that the syntactic robustness of LLM-based code generation improves significantly using our approach, improving syntactic robustness of LLMs from 54.05% to 74.42%.

Laboni Sarker, Abdus Satter, Tevfik Bultan

Traditional equivalence checking classifies programs as equivalent or non-equivalent, providing insufficient information for tasks like patch impact analysis where it is expected the patched version of the program to be non-equivalent to the original program. When two program versions are non-equivalent, determining under what conditions they differ and what percentage of inputs are affected remains an open challenge. In this work, we introduce quantitative partial equivalence analysis, an approach for assessing software patches by quantifying behavioral differences between the original (vulnerable) code and the patched code. Using symbolic analysis, we identify input conditions under which patched and original programs exhibit identical or divergent behaviors. Our approach refines non-equivalence by measuring the extent of behavioral divergence across the input domain. For efficient quantitative analysis of numerical domains, we propose a range-based search heuristic that provides a sound lower bound on equivalence. We demonstrate our approach on 90 CVE patches from widely used open-source projects (Linux, Qemu, FFmpeg), as well as on a Juliet Test Suite-based dataset containing programs with CWEs. Our results show that quantitative partial equivalence analysis effectively characterizes and quantifies patch impact. Additionally, experiments on the EqBench benchmark reveal five C program pairs that are mislabeled as equivalent, and we identify the input conditions under which their behaviors diverge.