LLM-based Vulnerability Discovery - arxiv.org/pdf/2509.19117 Our investigation leads to a disappointing outcome: despite the impressive capabilities of language models in other domains, their performance in vulnerability discovery is not significantly different from that of a simple baseline. The substantial resources required to train these models, along with the considerable effort in curating high-quality training datasets, do not yield a substantial advantage over simple techniques developed decades ago. Authors: @_thrsten, @mlsec. Felix Weißberg, Lukas Pirch, Erik Imgrund, Jonas Möller at @TUBerlin @bifoldberlin #AISecurity #LLMResearch #LineVul #PDBERT #UniXcoder #CodeGen25 #StarCoder2 #GPT35Turbo #GPT4o #Devign #ReVeal #RATS #Semgrep #SCM #CodeBERT #BigVul #CrossVul #CVEfixes #DiverseVul #SARD #CodeXGLUE