Subliminal Steering: Stronger Encoding of Hidden Signals

2026-04-28

AI Summary

This study investigates the transferability and underlying mechanisms of complex behavioral biases in implicit learning. The authors propose a "subconscious steering" approach, wherein a teacher model generates steering vectors that embed multi-word biases into otherwise benign data, enabling high-fidelity bias transfer during student model fine-tuning. This work presents the first demonstration of subconscious transmission of multi-word complex biases, revealing that such biases predominantly localize to specific network layers. By training steering vectors through target-sample likelihood maximization and analyzing them via cosine similarity, the authors successfully reconstruct new bias vectors on subsets of data that closely align with the original ones, thereby demonstrating the precision and reproducibility of bias encoding in neural representations.

Abstract

Subliminal learning describes a student language model inheriting a behavioral bias by fine-tuning on seemingly innocuous data generated by a biased teacher model. Prior work has begun to characterize this phenomenon but leaves open questions about the scope of signals it can transfer, the mechanisms that explain it, and the precision with which a bias can be encoded by seemingly unrelated data. We tackle all three problems by introducing subliminal steering, a variant of subliminal learning in which the teacher's bias is implemented not via a system prompt, as in prior work, but through a steering vector trained to maximize the likelihood of a set of target samples. First, we show that subliminal steering transfers complex multi-word biases, whereas prior work focused on single-word preferences, demonstrating a large scope of subliminally transferrable signals. Second, we provide mechanistic evidence that subliminal learning transfers not only the target behavioral bias, but also the steering vector itself, localized to the layers at which the teacher was steered. Finally, we show that the bias is encoded with surprising precision. We train a new steering vector directly on the subliminally-laden dataset and find that it attains high cosine similarity with the original vector.

Problem

• subliminal learning
• behavioral bias
• steering vector
• signal transfer
• bias encoding

Innovation

• subliminal steering
• steering vector
• behavioral bias transfer
• mechanistic interpretability
• model fine-tuning