Subliminal Learning Is Steering Vector Distillation

• URL: http://arxiv.org/abs/2606.00995v2
• Date: Wed, 03 Jun 2026 16:46:22 GMT
• Title: Subliminal Learning Is Steering Vector Distillation
• Abstract summary: Subliminal learning refers to a student language model acquiring a teacher's traits when fine-tuning.<n>We show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations.
• Score: 14.836506853464542
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Subliminal learning refers to a student language model acquiring a teacher's traits (e.g. a system-prompted preference for owls) when fine-tuned on the teacher's outputs, despite the outputs being semantically unrelated to those traits. It remains poorly understood how data without semantic meaning can transfer specific semantic traits. In this work, we show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations. Across two open-source models, we find that the teacher's system prompt is well approximated by a steering vector, and that the student's behavior is driven by learning an aligned vector over fine-tuning. System prompts that are not well approximated by steering vectors are not subliminally learned. This is a special case of steering vector distillation, in which a student trained on the outputs of a steered teacher learns to imitate that steering. We demonstrate steering vector distillation on a range of semantic and random vectors. Adding a semantic vector to a model's activations can have both model-independent and model-specific (i.e. non-semantic) effects on its behavior, so generated data that is non-semantic can transmit a vector with semantic effects, enabling subliminal learning. This also explains why subliminal learning does not transfer between models. We find that adaptive optimizers are necessary for subliminal learning in language models: activation gradients on steered data carry a small but consistent component along the steering direction, and non-adaptive optimizers impede this by allowing outlier gradients to dominate.

Related papers

• Subliminal Learning is Non-Semantic Distillation [1.2891210250935148]
  Subliminal Learning (SL) is a surprising type of generalization displayed by modern language models.<n>It allows the transfer of a bias or behavior from a teacher model to a student by distilling from seemingly unrelated or random synthetic data.<n>This presents challenges in ensuring AI systems remain predictable and are trained safely.
  arXiv  Detail & Related papers  (2026-08-06T08:18:13Z)
• Causal Interventions on Continuous Variables: A Case Study on Verb Bias in Steering Vectors for In-Context Learning [7.320379735876713]
  We introduce a method for causal intervention on continuous variables.<n>We apply this method to a continuous feature that is well-studied in psycholinguistics, namely verb bias.<n>We show that verb bias is causally represented in steering vectors extracted from large language models.
  arXiv  Detail & Related papers  (2026-05-28T14:09:18Z)
• Subliminal Steering: Stronger Encoding of Hidden Signals [5.13724383217928]
  Subliminal learning describes a student language model inheriting a behavioral bias by fine-tuning on seemingly innocuous data.<n>We introduce subliminal steering, a variant of subliminal learning in which the teacher's bias is implemented not via a system prompt, but through a steering vector trained to maximize the likelihood of a set of target samples.<n>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.
  arXiv  Detail & Related papers  (2026-04-28T15:51:55Z)
• Pando: Do Interpretability Methods Work When Models Won't Explain Themselves? [53.07826484214082]
  We introduce Pando, a model-organism benchmark that breaks this confound via an explanation axis.<n>Pando predicts held-out model decisions from $10$ labeled query-response pairs.<n>When explanations are faithful, black-box elicitation matches or exceeds all white-box methods.
  arXiv  Detail & Related papers  (2026-04-13T06:42:24Z)
• Understanding Unreliability of Steering Vectors in Language Models: Geometric Predictors and the Limits of Linear Approximations [0.0]
  I investigate why steering reliability differs across behaviors and how it is impacted by steering vector training data.<n>I find that higher cosine similarity between training activation differences predicts more reliable steering.<n>I observe that behavior datasets where positive and negative activations are better separated along the steering direction are more reliably steerable.
  arXiv  Detail & Related papers  (2026-02-19T22:37:05Z)
• Improving Instruction-Following in Language Models through Activation Steering [58.876600545898675]
  We derive instruction-specific vector representations from language models and use them to steer models accordingly.<n>We demonstrate how this method can enhance model adherence to constraints such as output format, length, and word inclusion.<n>Our findings demonstrate that activation steering offers a practical and scalable approach for fine-grained control in language generation.
  arXiv  Detail & Related papers  (2024-10-15T08:38:20Z)
• Activation Scaling for Steering and Interpreting Language Models [55.59689963561315]
  We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. We establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa. Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention.
  arXiv  Detail & Related papers  (2024-10-07T12:01:32Z)
• DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving [64.57963116462757]
  State-of-the-art methods usually follow the Teacher-Student' paradigm. Student model only has access to raw sensor data and conducts behavior cloning on the data collected by the teacher model. We propose DriveAdapter, which employs adapters with the feature alignment objective function between the student (perception) and teacher (planning) modules.
  arXiv  Detail & Related papers  (2023-08-01T09:21:53Z)
• Semi-Supervised Semantic Segmentation via Gentle Teaching Assistant [72.4512562104361]
  We argue that the unlabeled data with pseudo labels can facilitate the learning of representative features in the feature extractor. Motivated by this consideration, we propose a novel framework, Gentle Teaching Assistant (GTA-Seg) to disentangle the effects of pseudo labels on feature extractor and mask predictor.
  arXiv  Detail & Related papers  (2023-01-18T07:11:24Z)
• Extracting Latent Steering Vectors from Pretrained Language Models [14.77762401765532]
  We show that latent vectors can be extracted directly from language model decoders without fine-tuning. Experiments show that there exist steering vectors, which, when added to the hidden states of the language model, generate a target sentence nearly perfectly. We find that distances between steering vectors reflect sentence similarity when evaluated on a textual similarity benchmark.
  arXiv  Detail & Related papers  (2022-05-10T19:04:37Z)
• Fair Interpretable Learning via Correction Vectors [68.29997072804537]
  We propose a new framework for fair representation learning centered around the learning of "correction vectors" The corrections are then simply summed up to the original features, and can therefore be analyzed as an explicit penalty or bonus to each feature. We show experimentally that a fair representation learning problem constrained in such a way does not impact performance.
  arXiv  Detail & Related papers  (2022-01-17T10:59:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.