This is a link post for the paper preprint: Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors from the Center on Long-Term Risk.
Selective generalization. Training can teach desired and undesired traits at once. Selective generalization aims to preserve the generalization of desired traits while preventing the generalization of undesired ones. For example, RL environments may teach a model useful capabilities and a propensity to reward hack, and AI developers would like only the capabilities to generalize.
Inoculation adapters (IA) work similarly to inoculation prompting (IP), but instead of eliciting the undesired trait via prompting, we use a LoRA carrying the undesired trait during training. IA improves on IP in:
1. Achieving stronger suppression of undesired traits (e.g., emergent misalignment).
2. Being effective against new capabilities and hard-to-elicit traits, unlike inoculation prompting.
3. Creating substantially fewer surprising backdoors under our probes.
A family of methods. On average, IA outperforms other baselines, such as preventative steering and concept-ablation fine-tuning, in suppressing undesired traits. In terms of retention of the desired trait, (vanilla) IA performs worse than these baselines. We introduce gated IA (GIA) and complementary-gated IA (CGIA), which are in the same family of methods but achieve similar or better retention of the desired trait than the baselines. These variants jointly train gates to attenuate the inoculation adapter.
Undesired-trait expression (y-axis) versus desired-trait expression (x-axis). Lower-right is better. Each point shows the mean across nine normalized selective-generalization setups. IA better suppresses the generalization of the undesired trait. SFT(Safe) trains on an oracle-like, safe-only training dataset that is assumed to be unavailable in our problem setting.
Setup-dependence. Results are strongly setup-dependent, and the 95% confi