Linear probes recover jailbreak-only answers from aligned models' hidden states, sometimes transfer from base models, and correlate with pairwise preference rankings.
Fine-grained analysis of sentence embeddings using auxiliary prediction tasks
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Linearly Decoding Refused Knowledge in Aligned Language Models
Linear probes recover jailbreak-only answers from aligned models' hidden states, sometimes transfer from base models, and correlate with pairwise preference rankings.