Newest papers

Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability

Zhengyue Zhao, Liwei Jiang, Yejin Choi, Chaowei Xiao1 · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26809 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Sin embargo, la seguridad puede degradarse cuando el entrenamiento posterior ya no incluye datos de seguridad, y los beneficios en seguridad a menudo vienen al coste de la capacidad general. Esta tensión sugiere que la seguridad y el razonamiento general no deben ser tratados como objetivos de entrenamiento separados, mientras que el trabajo reciente también ha comenzado a incorporar datos de seguridad en el entrenamiento previo. Sin embargo, la seguridad puede degradarse cuando el entrenamiento posterior ya no incluye datos de seguridad, y los beneficios en seguridad a menudo vienen al coste de la capacidad general. Esta tensión sugiere que la seguridad y el razonamiento general no deben ser tratados como objetivos de entrenamiento separados. En cambio, consideramos la seguridad como una capacidad LLM que puede ser aprendida a través de una estructura de razonamiento compartida con las tareas generales. Teniendo en

Machine translation (Meta M2M100); the English original follows.

Abstract

Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come at the cost of general capability. This tension suggests that safety and general reasoning should not be treated as separate training goals. We instead view safety as an LLM capability that can be learned through a reasoning structure shared with general tasks. Both require the model to understand intent or task goals, apply relevant knowledge, and reason about potential consequences. We introduce safety mid-training, a stage between pre-training and post-training that jointly develops safety and general reasoning through this shared structure. Given safety and utility prompts, the base model generates structured documents and reasoning traces that are then used for mid-training. Safety mid-training reduces average attack success rate (ASR) by 72% to 81% relative to Qwen3-1.7B, 4B, and 8B base models, while also improving general capability. The benefits persist through subsequent post-training. After general-only SFT, mid-trained models largely retain their safety, achieving lower ASR than the baselines without sacrificing general capability. They also retain safety under general-only GRPO without safety rewards, while subsequent RL reaches a better safety-utility frontier when initialized from the mid-trained models. Together, these results suggest that learning safety and general reasoning within a shared mid-training stage can reduce the safety-utility trade-off and make safety more robust to subsequent training.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Task (project management), Computer science, Safety behaviors, Risk analysis (engineering), Work (physics), Knowledge base

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange