Hard-thresholding next tokens by a safety value function yields higher expected safety, an explicit type-I intervention bound controlled by one threshold, and better empirical safety-helpfulness-similarity trade-offs than Gibbs-style or ARGS steering.
Tradeoffs between alignment and helpfulness in language models
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
fields
cs.LG 3representative citing papers
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
Temporal preference in Qwen3-4B-Instruct-2507 localizes to layers 17–35 (especially L24 attention), has curved residual-stream geometry, is behaviorally unstable, and can be bidirectionally steered.
citing papers explorer
-
Selective Safety Steering via Value-Filtered Decoding
Hard-thresholding next tokens by a safety value function yields higher expected safety, an explicit type-I intervention bound controlled by one threshold, and better empirical safety-helpfulness-similarity trade-offs than Gibbs-style or ARGS steering.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
Temporal Preference Concepts and their Functions in a Large Language Model
Temporal preference in Qwen3-4B-Instruct-2507 localizes to layers 17–35 (especially L24 attention), has curved residual-stream geometry, is behaviorally unstable, and can be bidirectionally steered.