Out of Distribution, Out of Luck: How Well Can LLMs Trained on Vulnerability Datasets Detect Top 25 CWE Weaknesses?

Automated vulnerability detection research has made substantial progress, yet its real-world impact remains limited. Current vulnerability datasets suffer from issues including label inaccuracy rates of 20-71%, extensive duplication, and poor coverage of critical CWE types. These issues create a significant "generalization gap" where models achieve misleading self-testing performance (measured on held-out data from the same dataset for training) by exploiting spurious correlations rather than learning true vulnerability patterns. Our analysis reveals that many models experience substantial performance drops of up to 33% when evaluated on independent data, with some performing close to random guessing. To address these limitations, we present a three-part solution. First, we introduce a manually curated test dataset, BenchVul, covering the MITRE Top 25 Most Dangerous CWEs. Second, we construct a high-quality training dataset, TitanVul, comprising 38,863 functions by aggregating seven public sources and applying deduplication and validation using a novel multi-agent LLM framework. Third, we propose a Realistic Vulnerability Generation (RVG) framework, which synthesizes context-aware vulnerability examples for underrepresented but critical CWE types through simulated development workflows. Our evaluation shows the strengths of each component in closing the generalization gap. First, BenchVul shows the limitations of self-testing: models trained on existing datasets, such as BigVul and CVEfixes, experience performance drops on BenchVul (from 0.776 to 0.519 and from 0.713 to 0.607). Second, training models on TitanVul demonstrates improved generalization, with model performance increasing from 0.584 when evaluated on the same dataset to 0.767 when tested on BenchVul. Third, supplementing TitanVul with RVG-generated data yields further gains, increasing model performance by 14.0% to 0.874.