Safety
Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs
arXiv:2604.10403v1 Announce Type: new Abstract: We address jailbreaks, backdoors, and unlearning for large language models (LLMs). Unlike prior work, which trains LLMs based on their actions when give
arXiv:2604.10403v1 Announce Type: new Abstract: We address jailbreaks, backdoors, and unlearning for large language models (LLMs). Unlike prior work, which trains LLMs based on their actions when given malign instructions, our method specifically trains the model to change how it interprets instructions. Our method, Latent Instruction Representation Alignment (LIRA), greatly improves generalization. We further boost generalization through an internally adversarial training algorithm. Our methods block over 99% of PEZ jailbreak attacks; remove a challenging insecure code backdoor; and achieve optimal forgetting on WMDP cyber with negligible loss of benign capabilities.
Related
- What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
- Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism
- Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model
- Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs
- Geometric Properties of the Voronoi Tessellation in Latent Semantic Manifolds of Large Language Models
Source: arXiv cs.LG | 2026-04-14