Calibration Without Comprehension: Diagnosing the Limits of Fine-Tuning LLMs for Vulnerability Detection in Systems Software

This paper investigates whether Large Language Models (LLMs) actually learn to reason about software security or if they simply memorize patterns from contaminated training data. The authors seek to determine if fine-tuning on standard vulnerability datasets improves security reasoning or merely adjusts output distributions to match labels without understanding the underlying code semantics.

The authors introduce CWE-Trace, a framework for evaluating LLM vulnerability detection using 834 manually curated Linux kernel samples. To ensure rigor, the framework employs a strict temporal split (historical data pre-2025 vs. leakage-free data post-2025) and preserves multi-file context, which is critical for identifying complex vulnerabilities. The study evaluates eight vanilla LLMs and 15 LoRA fine-tuned variants across three tasks: non-targeted detection, targeted detection, and hierarchical CWE classification. They also introduce two diagnostic metrics, the Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD), to measure systematic biases and reasoning depth.

The analysis reveals that data contamination is not the primary driver of performance; instead, models exhibit stable, systematic failure modes (directional priors) that persist regardless of training. Fine-tuning effectively shifts the output threshold—making the model more or less likely to flag code—but it does not change the underlying decision policy or improve security reasoning. Even after fine-tuning, detection accuracy remains near chance (52.1%), and the ability to correctly classify the type of vulnerability (CWE) is negligible, with Top-1 accuracy below 1.3%. The authors conclude that detection and understanding are decoupled capabilities in current LLMs.

This work challenges the prevailing assumption that scaling and fine-tuning LLMs on existing vulnerability datasets will lead to reliable automated security tools. By demonstrating that current models lack genuine security reasoning, the paper highlights the need for new approaches that move beyond simple pattern matching and toward models capable of deep semantic analysis of multi-file dependencies.