提升大模型内在透明度：无需外部模块实现高效监控与自发安全增强｜上海AI Lab & 上交

Core Insights - The article discusses the challenges of AI safety related to large language models (LLMs) and introduces TELLME, a new method aimed at enhancing internal transparency without relying on external monitoring modules [1][2][26]. Group 1: Current Challenges in AI Safety - Concerns about the potential risks associated with LLMs have arisen due to their increasing capabilities [1]. - Existing external monitoring methods are criticized for being unreliable and lacking adaptability, leading to unstable monitoring outcomes [5][6]. - The reliance on "black box" external detectors results in low interpretability and trustworthiness of monitoring results [5]. Group 2: TELLME Methodology - TELLME employs a technique called "representation decoupling" to enhance the internal transparency of LLMs [2]. - The core idea is to clearly separate the internal representations of safe and unsafe behaviors, facilitating more reliable monitoring [3]. - TELLME utilizes contrastive learning to drive the separation of representations, ensuring that similar risks are grouped while dissimilar ones are distanced [7]. Group 3: Experimental Validation - Experiments demonstrate significant improvements in transparency and monitoring capabilities across various scenarios, with clear clustering of different risk behaviors [10][11]. - The method maintains the general capabilities of the model while enhancing safety, proving the effectiveness of the dual constraints designed in TELLME [12]. - Monitoring accuracy increased by 22.3% compared to the original model, showcasing the method's effectiveness [14]. Group 4: Broader Implications - TELLME represents a shift from external monitoring reliance to enhancing the model's own monitorability, leading to higher precision in risk identification [26][27]. - The method shows potential for scalable oversight, suggesting that as model capabilities grow, so too will the effectiveness of TELLME's monitoring [28]. - The approach leads to spontaneous improvements in output safety, indicating a unique mechanism for enhancing model safety [23][28].