# Are Large Language Models Robust in Understanding Code Against Semantics-Preserving Mutations?

arXiv: https://arxiv.org/abs/2505.10443 · HTML: https://mlsanity.com/paper/2505.10443/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Builds on 16 of its 30 references that are in the map.
- First posted on arXiv 15 May 2025.

## Builds on (most cited first, 16 of 16)

- [Code Llama: Open Foundation Models for Code](https://mlsanity.com/paper/2308.12950/) (2023)
- [LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code](https://mlsanity.com/paper/2403.07974/) (2024)
- [DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence](https://mlsanity.com/paper/2406.11931/) (2024)
- [CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution](https://mlsanity.com/paper/2401.03065/) (2024)
- [CodeGemma: Open Code Models Based on Gemma](https://mlsanity.com/paper/2406.11409/) (2024)
- [Granite Code Models: A Family of Open Foundation Models for Code Intelligence](https://mlsanity.com/paper/2405.04324/) (2024)
- [ReCode: Robustness Evaluation of Code Generation Models](https://mlsanity.com/paper/2212.10264/) (2022)
- [A Large-Scale Survey on the Usability of AI Programming Assistants: Successes and Challenges](https://mlsanity.com/paper/2303.17125/) (2023)
- [Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad](https://mlsanity.com/paper/2503.21934/) (2025)
- [Self-Supervised Bug Detection and Repair](https://mlsanity.com/paper/2105.12787/) (2021)
- [Semantic Robustness of Models of Source Code](https://mlsanity.com/paper/2002.03043/) (2020)
- [Logically Consistent Language Models via Neuro-Symbolic Integration](https://mlsanity.com/paper/2409.13724/) (2024)
- [The Larger They Are, the Harder They Fail: Language Models do not Recognize Identifier Swaps in Python](https://mlsanity.com/paper/2305.15507/) (2023)
- [The Counterfeit Conundrum: Can Code Language Models Grasp the Nuances of Their Incorrect Generations?](https://mlsanity.com/paper/2402.19475/) (2024)
- [Poisoned ChatGPT Finds Work for Idle Hands: Exploring Developers' Coding Practices with Insecure Suggestions from Poisoned AI Models](https://mlsanity.com/paper/2312.06227/) (2023)
- [Counterexample Guided Program Repair Using Zero-Shot Learning and MaxSAT-based Fault Localization](https://mlsanity.com/paper/2502.07786/) (2024)

## Built on by (most cited first, 7 of 7)

- [LLMs Lean on Priors, Not Programming Language Semantics](https://mlsanity.com/paper/2510.03415/) (2025)
- [RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations](https://mlsanity.com/paper/2605.26177/) (2026)
- [Names Don't Matter: Symbol-Invariant Transformer for Open-Vocabulary Learning](https://mlsanity.com/paper/2601.23169/) (2026)
- [Syntax- and Compilation-Preserving Evasion of LLM Vulnerability Detectors](https://mlsanity.com/paper/2602.00305/) (2026)
- [How Do Semantically Equivalent Code Transformations Impact Membership Inference on LLMs for Code?](https://mlsanity.com/paper/2512.15468/) (2025)
- [Evolutionary Task Discovery: Advancing Reasoning Frontiers via Skill Composition and Complexity Scaling](https://mlsanity.com/paper/2605.11666/) (2026)
- [From Brewing to Resolution: Tracing the Internal Lifecycle of Code Reasoning in LLMs](https://mlsanity.com/paper/2606.17648/) (2026)

## Abstract (arXiv)

With the widespread adoption of vibe coding, understanding the reasoning and robustness of Large Language Models (LLMs) is critical for their reliable use in programming tasks. While recent studies assess LLMs' ability to predict program outputs, most focus on accuracy alone, without evaluating the underlying reasoning. Moreover, it has been observed on mathematical reasoning tasks that LLMs can arrive at correct answers through flawed logic, raising concerns about similar issues in code understanding. In this paper we assess whether state-of-the-art LLMs can reason about Python programs or are simply guessing. We apply five semantics-preserving code mutations: renaming variables, mirroring comparison expressions, swapping if-else branches, converting for loops to while, and loop unrolling. These mutations maintain program semantics while altering its syntax. We evaluated nine LLMs, including both open-source and closed-access models, and performed a human expert analysis using LiveCodeBench to assess whether correct predictions are based on sound reasoning. We also evaluated prediction stability across different code mutations on LiveCodeBench and CruxEval. While proprietary models achieve the strongest predictive accuracy and reasoning quality in the expert evaluation, our robustness analysis reveals substantial fragility under semantics-preserving transformations. Our findings show that LLMs trained for code produce correct predictions based on flawed reasoning in between 10% and 50% of cases. Furthermore, LLMs often change predictions in response to our code mutations, with performance drops reaching up to 70%, indicating that they do not yet exhibit stable, semantically grounded reasoning, even when initial accuracy is high.