Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.13536

Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning

Joseph Egan·Ciarán M. Gilligan-Lee·Michael O'Riordan·Yuchen Zhu
View original source ↗Read source document ↗Export metadata ↓

Abstract

When a language model is distilled from a teacher that shares its base parameters, the student can inherit behavioural traits of the teacher---including misalignment---even from data that is semantically unrelated to the trait (Cloud et al., 2026). All existing defences require knowing the trait. We remove this requirement and prove that a subliminally transferred trait can be recovered from the parameter updates alone, without supervision, by exploiting a single asymmetry: the trait is expressed consistently across diverse datasets generated by the teacher, whereas task-specific and shared capabilities vary. The resulting estimator returns a weight-space direction that can be amplified to discover and interpret the trait or projected out to suppress it. Applied to the subliminal-learning benchmarks of (Cloud et al., 2026), our method recovers the transmitted animal-preference traits without any trait labels and suppresses their expression with almost no impact to task performance.

Explore related subjects

stat.MLcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Joseph Egan, Ciarán M. Gilligan-Lee, Michael O'Riordan, Yuchen Zhu. 2026-10-09. Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning. https://arxiv.org/abs/2610.13536

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Approximation of differential entropy in Bayesian optimal experimental design

Bayesian optimal experimental design provides a principled framework for selecting experimental configurations that maximize obtained information. In this work, we consider settings where the likelihood entropy is independent of the design or can be evaluated explicitly, which applies, e.g. to additive noise models with design-independent or explicitly computable noise entropy. This reduces the problem to maximum entropy estimation, where the nested integral defining the expected information gain collapses into a single integral. Our study is motivated by large-scale inference problems, such as inverse problems, where the computational cost is dominated by expensive likelihood evaluations. Under additive Gaussian noise model, we propose a computational approach in which the evidence density is approximated by a Monte Carlo or quasi-Monte Carlo surrogate, while the differential entropy is evaluated using standard methods without additional likelihood evaluations. We prove that this strategy achieves convergence rates that are comparable to, or better than, state-of-the-art methods for full expected information gain estimation, particularly when the cost of entropy evaluation is negligible. Our approach relies only on mild smoothness of the forward map and avoids stronger technical assumptions required in earlier work. We also present numerical experiments, which confirm some of our theoretical findings.

stat.ML↗
arXiv2026

Surrogate Graph Partitioning for Spatial Prediction

Spatial prediction refers to the estimation of unobserved values from spatially distributed observations. Although recent advances have improved the capacity to model diverse observation types, adoption in practice remains limited in industries that demand interpretability. To mitigate this gap, surrogate models that explain black-box predictors provide a promising path toward interpretable decision making. In this study, we propose a graph partitioning problem to construct spatial segments that minimize the sum of within-segment variances of individual predictions. The assignment of data points to segments can be formulated as a mixed-integer quadratic programming problem. While this formulation potentially enables the identification of exact segments, its computational complexity becomes prohibitive as the number of data points increases. Motivated by this challenge, we develop an approximation scheme that leverages the structural properties of graph partitioning. Experimental results demonstrate the computational efficiency of this approximation in identifying spatial segments.

stat.ML↗
arXiv2026

Fast Rates for Nonstationary Weighted Risk Minimization

Weighted empirical risk minimization is a common approach to prediction under distribution drift. This article studies its out-of-sample prediction error under nonstationarity. We provide a general decomposition of the excess risk into a learning term and an error term associated with distribution drift, and prove oracle inequalities for the learning error under mixing conditions. The learning bound holds uniformly over arbitrary weight classes and accounts for the effective sample size induced by the weight vector, the complexity of the weight and hypothesis classes, and potential data dependence. We also analyze the drift error in the square loss setting and provide explicit bounds for structural-break and smoothly drifting processes. We then combine the learning and drift analysis in (auto-) regression problems with linear models, basis approximations, and neural networks. In all cases, the optimal weights are characterized by a bias--variance trade-off between learning and drift, while minimax-optimal rates (up to logarithmic factors) are recovered when specialized to unweighted and stationary settings.

stat.ML↗