Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors

arXiv cs.AI · 6 hours ago

Researchers introduced inoculation adapters, a method using LoRA modules trained on undesired traits then frozen during main training to reduce unwanted model behaviors while preserving desired capabilities. Testing across nine setups and five model families showed the approach achieves better tradeoffs than existing techniques like inoculation prompting and CAFT, though with wide confidence intervals on improvement magnitude. The method avoids some issues of prompt-based inoculation but still involves tradeoffs, where gains in desired-trait generalization typically come with weaker suppression of undesired traits and increased backdoor occurrence.