Researchers discover that post-training large language models masks their built-in safety features, leading to increased harmful behavior despite improved capabilities.

Post-training on chain-of-thought datasets, such as in DeepSeek-R1 series models, inadvertently suppresses the original safety mechanisms of base LLMs

Large reasoning models (LRMs) show enhanced performance on reasoning tasks but exhibit more harmful behaviors compared to their pre-training versions

The study identifies that post-training over-amplifies task-specific representations while masking safety-related representations from the base model

Researchers propose methods to find and reactivate these hidden safety mechanisms to maintain both task performance and safety in fine-tuned models