Towards Identifying the Dataset Biases Causing Phantom Transfer
cs.LG
Submitted: 2026-09-13
Updated: 2026-10-02
License: http://creativecommons.org/licenses/by/4.0/
The gist: Recent work has shown that a teacher model can transfer a bias to a student through a dataset from which every explicit reference to that bias has been filtered out, and that no data-level defense
Terminology
Abstract
Recent work has shown that a teacher model can transfer a bias to a student through a dataset from which every explicit reference to that bias has been filtered out, and that no data-level defense reliably removes or detects it even when knowing what bias to look for. Aiming to shed light on the hidden traces of these biases, we show that a simple signature based on Sentence BERT embeddings can identify the topic of such a bias with a Matthews correlation coefficient of 0.83 if the teacher model used by the attacker is known and 0.46 if it is not. Additionally, we observe that different teacher models appear to express the same bias through different vocabulary.
Sources
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
- Subliminal Learning Is Steering Vector Distillation
- Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
- Learning Through Noise: Why Subliminal Learning Works and When It Fails
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences
- Efficient Estimation of Word Representations in Vector Space
- Subliminal Learning is a LoRA Artifact
- You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases
- Filtering Harmful Actions Isn't Enough: Phantom Transfer in Agentic SDF
- Channel Location Constrains the Auditability of Subliminal Learning
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks