Safety Pre-training & Post-training Measures

Training data filters & unlearning

Filter out data that could be hazardous, such as bioweapon and gain-of-function-related knowledge. While currently less successful, unlearning techniques could also be applied to make hazardous knowledge more difficult for users to access.

Further research into foundational approaches such as Safety-By-Design and Quantitative Safety Guarantees

Safety-By-Design focuses on integrating safety principles into model architecture and training processes from the outset, reducing the potential for harmful capabilities to emerge. Quantitative Safety Guarantees aim to provide measurable, mathematically grounded assurances that risks remain below predefined thresholds, enhancing trust in model behavior across diverse scenarios. These approaches strengthen the foundation for safe AI deployment, complementing existing safeguards and addressing evolving challenges in high-risk contexts.

Safety alignment training against harmful instructions

Through alignment training (e.g., RLHF/RLAIF) and red-team-driven fine-tuning, enhance the model's ability to recognize and refuse high-risk content related to violence, weapon development, etc.

Embedding safety values and behavioral constraints

Inject constraints aligned with values like honesty and controllability during training to ensure the model adheres to human intent in complex scenarios

Real-time monitoring of reasoning processes

Introduce automated chain-of-thought monitoring to identify anomalies or potentially malicious behaviors during reasoning, to help detect deceptive, conspiratorial, or manipulative outputs.

Enhancing interpretability and formal verification

Use techniques like neural network reverse engineering to analyze internal mechanisms and identify risks; combine with formal verification methods to mathematically validate critical behaviors, increasing trustworthiness.

Restricting dangerous capability formation

Employ unlearning and capability boundary control to suppress the development of abilities tied to high-risk tasks without significantly undermining general performance.

Differentiated fine-tuning strategies

Design targeted fine-tuning paths based on risk levels and application scenarios to improve the model's safety adaptability in specific contexts.

Enhancing model anomaly detection capabilities

Train the model to be sensitive to anomalous behaviors, enabling it to automatically halt execution or issue alerts when high-risk instructions are triggered.