VLAF diagnostics show alignment faking is widespread in LLMs as small as 7B parameters, driven by consistent activation shifts that can be mitigated with contrastive steering vectors reducing faking by 58-94%.
Some important guidelines to follow: • The system prompt must not explicitly instruct what the model is required to do in response to the harmful instruction/request
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models
VLAF diagnostics show alignment faking is widespread in LLMs as small as 7B parameters, driven by consistent activation shifts that can be mitigated with contrastive steering vectors reducing faking by 58-94%.