SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{circ}$ Law

2025-07-24

AI Summary

This work addresses the critical challenge of co-evolving safety and intelligence in AI models. We propose SafeLadder, a framework for safety-intelligence co-evolution, and SafeWork-R1, a multimodal reasoning model. Methodologically, we introduce the novel “Safety Insight” mechanism and the AI-45° Law—a principled paradigm guiding joint capability-safety development. Our approach integrates multi-principle verifiers (logical, causal, normative), dual-path inference-time intervention (masking and re-ranking), and deliberative search to enable intrinsically safe reasoning and self-reflection. Experiments demonstrate a 46.54% average improvement on safety benchmarks—outperforming Qwen2.5-VL-72B, GPT-4.1, and Claude Opus 4—while preserving full general-purpose capability with zero degradation. Moreover, the framework exhibits strong cross-architecture generalization across InternVL3, DeepSeek, and Qwen2.5-VL. To our knowledge, this is the first empirical demonstration that safety and intelligence can co-evolve synergistically in large language and multimodal models.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Security and Privacy: Security and privacy of machine learning and AI applications

Abstract

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

Problem

• Develops SafeWork-R1 for coevolving AI safety and intelligence
• Enhances safety reasoning without compromising general capabilities
• Ensures reliability via inference-time interventions and verification

Innovation

• Multimodal reasoning model with coevolving safety
• SafeLadder framework with progressive RL post-training
• Inference-time intervention and deliberative search mechanism