Aaron J. Li

Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language models (LLMs) by mapping them to human-interpretable concept representations. While existing evaluations of SAEs focus on metrics such as the reconstruction-sparsity tradeoff, human (auto-)interpretability, and feature disentanglement, they overlook a critical aspect: the robustness of concept representations to input perturbations. We argue that robustness must be a fundamental consideration for concept representations, reflecting the fidelity of concept labeling. To this end, we formulate robustness quantification as input-space optimization problems and develop a comprehensive evaluation framework featuring realistic scenarios in which adversarial perturbations are crafted to manipulate SAE representations. Empirically, we find that tiny adversarial input perturbations can effectively manipulate concept-based interpretations in most scenarios without notably affecting the outputs of the base LLMs themselves. Overall, our results suggest that SAE concept representations are fragile and may be ill-suited for applications in model monitoring and oversight.

The trustworthiness of Large Language Models (LLMs) refers to the extent to which their outputs are reliable, safe, and ethically aligned, and it has become a crucial consideration alongside their cognitive performance. In practice, Reinforcement Learning From Human Feedback (RLHF) has been widely used to align LLMs with labeled human preferences, but its assumed effect on model trustworthiness hasn’t been rigorously evaluated. To bridge this knowledge gap, this study investigates how models aligned with general-purpose preference data perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Our results demonstrate that RLHF on human preferences doesn’t automatically guarantee trustworthiness, and reverse effects are often observed. Furthermore, we propose to adapt efficient influence function based data attribution methods to the RLHF setting to better understand the influence of fine-tuning data on individual trustworthiness benchmarks, and show its feasibility by providing our estimated attribution scores. Together, our results underscore the need for more nuanced approaches for model alignment from both the data and framework perspectives, and we hope this research will guide the community towards developing language models that are increasingly capable without sacrificing trustworthiness.

In recent years, work has gone into developing deep interpretable methods for image classification that clearly attributes a model’s output to specific features of the data. One such of these methods is the Prototypical Part Network (ProtoPNet), which attempts to classify images based on meaningful parts of the input. While this architecture is able to produce visually interpretable classifications, it often learns to classify based on parts of the image that are not semantically meaningful. To address this problem, we propose the Reward Reweighing, Reselecting, and Retraining (R3) post-processing framework, which performs three additional corrective updates to a pretrained ProtoPNet in an offline and efficient manner. The first two steps involve learning a reward model based on collected human feedback and then aligning the prototypes with human preferences. The final step is retraining, which realigns the base features and the classifier layer of the original model with the updated prototypes. We find that our R3 framework consistently improves both the interpretability and the predictive accuracy of ProtoPNet and its variants.