Publications

Selected

Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering

Xu, W., & Sun, Z. · arXiv:2610.19359 · 2026

PreprintSubmitted to ACL Rolling Review · October 2026

LLM interpretabilityActivation steering
Summary
Introduces a competing-target test for activation steering. In the tested ethical-framework tasks, Kantian and utilitarian vectors shift both target scores similarly, while prompting distinguishes the frameworks; a gain on a named target alone does not establish selective control.
arXiv
Width-to-Depth Ratio Tracks Readout Misalignment in Language Models
Xu, W. · Mechanistic Interpretability Workshop at ICML 2026
Workshop paper
LLM interpretabilityMeasurement
Summary
Studies readout misalignment and the interpretation of information recovered from hidden representations.
A Bayesian account of the phantom lighthouse illusion

Kemp, C., Xu, W., & Cropper, S. J. · Open Mind, 2026

Published
Cognitive modellingVisual perception
Summary
A Bayesian account of a visual illusion, connecting perceptual judgments to assumptions about natural scenes.

All publications

2026

Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering

Xu, W., & Sun, Z. · arXiv:2610.19359 · 2026

PreprintSubmitted to ACL Rolling Review · October 2026

LLM interpretabilityActivation steering
Summary
Introduces a competing-target test for activation steering. In the tested ethical-framework tasks, Kantian and utilitarian vectors shift both target scores similarly, while prompting distinguishes the frameworks; a gain on a named target alone does not establish selective control.
arXiv
Ethical Frameworks in LLMs Route Through a Shared Endorsement Subspace
Xu, W., Rusnak, A., & Kaplan, F. · Mechanistic Interpretability Workshop at ICML 2026
Workshop paper
LLM interpretabilityEthical reasoning
Summary
Examines how ethical-framework representations relate to model judgments, using probing and interventions.
Width-to-Depth Ratio Tracks Readout Misalignment in Language Models
Xu, W. · Mechanistic Interpretability Workshop at ICML 2026
Workshop paper
LLM interpretabilityMeasurement
Summary
Studies readout misalignment and the interpretation of information recovered from hidden representations.
Diachronic Modeling of Tonal Coherence on the Tonnetz Across Classical and Popular Repertoires

Xu, W., Hall, E., & Rohrmeier, M. · Sound and Music Computing Conference (SMC 2026)

Accepted · forthcoming
Computational musicologyTonal structure
Summary
Proposes tonal focus and tonal connection as two partially independent dimensions of tonal coherence on the Tonnetz. 2,800+ pieces across classical (1680–1920) and popular (1950–2020) repertoires occupy distinct regions of this space.

arXiv
Acoustic Overspecification in Electronic Dance Music Taxonomy

Xu, W., Dai, T., Goudet, O., & Wang, X. · Sound and Music Computing Conference (SMC 2026)

Accepted · forthcoming
Computational musicologyMusic classification
Summary
EDM's 35-genre commercial taxonomy is acoustically overspecified. Interpretable features + unsupervised clustering converge to 17–20 natural acoustic families.

arXiv
Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges

Xu, W., Rusnak, A., & Kaplan, F. · Preprint

Earlier preprint
LLM interpretabilityEthical reasoning
Summary
Earlier preprint on ethical-framework probing and its methodological limitations. See also the ICML 2026 workshop paper above for the subsequent work.

arXiv
A Bayesian account of the phantom lighthouse illusion

Kemp, C., Xu, W., & Cropper, S. J. · Open Mind, 2026

Published
Cognitive modellingVisual perception
Summary
Why we see a "phantom beam" in the air some distance from a lighthouse. Explained from natural scene statistics via a Bayesian observer model.

2023

A Bayesian account of two visual illusions involving lighthouse beams

Xu, W., Cropper, S., & Kemp, C. · Proceedings of the 45th Annual Meeting of the Cognitive Science Society

Published
Cognitive modellingVisual perception
Summary
Bent lighthouse beams and the phantom illusion as rational inference under natural image priors. A computational model predicts the phantom appears at ~75 m from the true source.
An active foveated gaze prediction algorithm based on a Bayesian ideal observer

Rashidi, S., Xu, W., Lin, D., Turpin, A., Kulik, L., & Ehinger, K. · Pattern Recognition, 143, 109694

Published
Cognitive modellingVisual perception
Summary
A Bayesian ideal observer adapted for foveated fixation prediction. Competitive on Ehinger and COCO-search18.

DOI