Do Language Models Prefer Vulnerable Code? A Probabilistic Study of Insecure Code Preference

Rui Melo, Sofia Reis, Andre Catarino, Rui Abreu

ICST 2026 AAccepted 2026

Abstract—Large Language Models (LLMs) are increasingly integrated into software development and testing workflows, offering the promise of automated code generation, test synthesis, and program repair. However, ensuring the security of LLM-generated code remains a critical challenge for software verification and validation, as these models may inadvertently learn and propagate insecure patterns from their training data. In this paper, we present a probabilistic testing framework for evaluating the security alignment of code LLMs, analyzing their internal behavior across three dimensions: fluency (does the code appear natural?), preference (which version is the model more likely to generate?), and confidence (how certain is the model about its choice?). Using Delta-Secommits, a 2,422 real-world vulnerability-patch pairs spanning 25 CWE categories, we conduct the first empirical study of how code LLMs probabilistically favor secure versus insecure code. Our results reveal a significant security misalignment: LLMs exhibit a bias toward insecure code in approximately 92% of cases. Even when secure code is as fluent or confidently predicted, models still prefer the vulnerable version in the vast majority of comparisons. For researchers, our findings extend existing evaluation frameworks by introducing probabilistic security alignment, measuring not only generated outputs, but also the likelihoods that drive them. For tool builders, the implication is clear: AI coding assistants must be designed for and tested against secure defaults, or they risk amplifying vulnerabilities at scale.

[Preprint] [ ] [ ]

SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain

Pierre Colombo, Telmo Pires, Malik Boudiaf, Rui Melo, Dominic Culver, Sofia Morgado, Etienne Malaboeuf, Gabriel Hautreux, Johanne Charpentier, Michael Desa

NeurIPS A* 2024

In this paper, we introduce SaulLM-54B and SaulLM-141B, two large language models (LLMs) tailored for the legal sector. These models, which feature architectures of 54 billion and 141 billion parameters, respectively, are based on the Mixtral architecture. The development of SaulLM-54B and SaulLM-141B is guided by large-scale domain adaptation, divided into three strategies: (1) the exploitation of continued pretraining involving a base corpus that includes over 540 billion of legal tokens, (2) the implementation of a specialized legal instruction-following protocol, and (3) the alignment of model outputs with human preferences in legal interpretations. The integration of synthetically generated data in the second and third steps enhances the models' capabilities in interpreting and processing legal texts, effectively reaching state-of-the-art performance and outperforming previous open-source models on LegalBench-Instruct. This work explores the trade-offs involved in domain-specific adaptation at this scale, offering insights that may inform future studies on domain adaptation using strong decoder models. Building upon SaulLM-7B, this study refines the approach to produce an LLM better equipped for legal tasks. We are releasing base, instruct, and aligned versions on top of SaulLM-54B and SaulLM-141B under the MIT License to facilitate reuse and collaborative research.