Steering 112 manually identified 'deception features' in Qwen3-4B changed whether the model revealed a hidden word, but the same steering test was used to pick the features.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Transcoders for Investigating Deception in Language Models
Steering 112 manually identified 'deception features' in Qwen3-4B changed whether the model revealed a hidden word, but the same steering test was used to pick the features.