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
Tuy nhiên, an toàn có thể suy giảm khi đào tạo sau này không bao gồm dữ liệu an toàn, và lợi ích trong an toàn thường đến với chi phí của khả năng chung. Sự căng thẳng này gợi ý rằng an toàn và lý luận chung không nên được coi là mục tiêu đào tạo tách biệt. Thay vào đó, chúng tôi xem an toàn là một khả năng LLM có thể được học thông qua một cấu trúc lý luận chia sẻ với các nhiệm vụ chung. Cả hai yêu cầu mô hình để hiểu ý định hoặc mục tiêu giữa nhiệm vụ, áp dụng kiến thức có liên quan, và lý do về hậu quả tiềm năng. Chúng tôi giới thiệu an toàn giữa đào tạo, một giai đoạn giữa đào tạo trước và sau đào tạo mà cùng nhau phát triển an toàn và lý luận chung thông qua cấu trúc chung này. Được cung cấp an toàn và tiện ích khuyến cáo, mô hình cơ bản tạo ra các tài liệu có cấu
Machine translation (Meta M2M100); the English original follows.
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
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.26809, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange