The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features

Aligning large language models is critical for their usability and safety. However, the prevailing approach of Reinforcement Learning from Human Feedback (RLHF) induces diffuse, opaque parameter changes, making it difficult to discern what the model has internalized. Hence, we introduce Feature Steering with Reinforcement Learning (FSRL), a transparent alignment framework that trains a lightweight adapter to steer behavior by modulating interpretable features from a Sparse Autoencoder (SAE). First, we demonstrate that FSRL is an effective method for preference optimization and is comparable with current RLHF methods. We then perform mechanistic analysis on the trained adapter, and find that its policy systematically promotes style features over explicit alignment concepts, suggesting that the preference optimization process rewards stylistic presentation as a proxy for quality. Ultimately, we hope that FSRL provides a tool for both interpretable model control and diagnosing the internal mechanisms of alignment.

Evaluating Uncertainty in Deep Gaussian Processes

Reliable uncertainty estimates are crucial in modern machine learning. Deep Gaussian Processes (DGPs) and Deep Sigma Point Processes (DSPPs) extend GPs hierarchically, offering promising methods for uncertainty quantification grounded in Bayesian principles. However, their empirical calibration and robustness under distribution shift relative to baselines like Deep Ensembles remain understudied. This work evaluates these models on regression (CASP dataset) and classification (ESR dataset) tasks, assessing predictive performance (MAE, Accu- racy), calibration using Negative Log-Likelihood (NLL) and Expected Calibration Error (ECE), alongside robustness under various synthetic feature-level distribution shifts. Results indicate DSPPs provide strong in-distribution calibration leveraging their sigma point approximations. However, compared to Deep Ensembles, which demonstrated superior robustness in both per- formance and calibration under the tested shifts, the GP-based methods showed vulnerabilities, exhibiting particular sensitivity in the observed metrics. Our findings underscore ensembles as a robust baseline, suggesting that while deep GP methods offer good in-distribution calibration, their practical robustness under distribution shift requires careful evaluation. To facilitate reproducibility, we make our code available at https://github.com/matthjs/xai-gp.