LLM Performance for Code Generation on Noisy Tasks

TL;DR

The paper investigates whether large language models genuinely reason about obfuscated programming and math problems or simply overfit to contaminated benchmarks. It introduces a reproducible obfuscation framework and decays-based evaluation using LeetCode OldLC/NewLC and MATH, revealing robust memorisation and eager pattern matching that inflate apparent capabilities on contaminated data. Key contributions include demonstrating distinct decay patterns between contaminated and unseen datasets, proposing practical contamination-detection metrics, and providing a reproducible pipeline and adversarial tasks to probe reasoning versus pattern matching. The findings underscore the need for dynamic, contamination-aware benchmarking and careful deployment of LLM-based code-generation systems, given safety and interpretability risks in automated software contexts.

Abstract

This paper investigates the ability of large language models (LLMs) to recognise and solve tasks which have been obfuscated beyond recognition. Focusing on competitive programming and benchmark tasks (LeetCode and MATH), we compare performance across multiple models and obfuscation methods, such as noise and redaction. We demonstrate that all evaluated LLMs can solve tasks obfuscated to a level where the text would be unintelligible to human readers, and does not contain key pieces of instruction or context. We introduce the concept of eager pattern matching to describe this behaviour, which is not observed in tasks published after the models' knowledge cutoff date, indicating strong memorisation or overfitting to training data, rather than legitimate reasoning about the presented problem. We report empirical evidence of distinct performance decay patterns between contaminated and unseen datasets. We discuss the implications for benchmarking and evaluations of model behaviour, arguing for caution when designing experiments using standard datasets. We also propose measuring the decay of performance under obfuscation as a possible strategy for detecting dataset contamination and highlighting potential safety risks and interpretability issues for automated software systems.

Figures (15)

• Figure 1: Example of an obfuscated task.
• Figure 2: Examples of obfuscated tasks.
• Figure 3: Simplified experiment pipeline.
• Figure 4: Adversarial task resembling the structure of the "median of two sorted arrays" task.
• Figure 5: The performance of LLMs on the two LeetCode datasets, averaged across the three obfuscation methods. Performance across the augmentation methods is presented in Appendix A.