Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866908957198516224 |
|---|---|
| author | Easley, Eric Farquhar, Sebastian |
| author_facet | Easley, Eric Farquhar, Sebastian |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_10403 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs Easley, Eric Farquhar, Sebastian Machine Learning I.2.7; I.2.6 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. |
| title | Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs |
| topic | Machine Learning I.2.7; I.2.6 |
| url | https://arxiv.org/abs/2604.10403 |