A Unified Understanding and Evaluation of Steering Methods

Steering methods provide a practical approach to controlling large language models by applying steering vectors to intermediate activations, guiding outputs toward desired behaviors while avoiding retraining. Despite their growing importance, the field lacks a unified understanding and consistent evaluation across tasks and datasets, hindering progress. This paper introduces a unified framework for analyzing and evaluating steering methods, formalizing their core principles and offering theoretical insights into their effectiveness. Through comprehensive empirical evaluations on multiple-choice and open-ended text generation tasks, we validate these insights, identifying key factors that influence performance and demonstrating the superiority of certain methods. Our work bridges theoretical and practical perspectives, offering actionable guidance for advancing the design, optimization, and deployment of steering methods in LLMs.

On the Generalization of Preference Learning with DPO

Large language models (LLMs) have demonstrated remarkable capabilities but often struggle to align with human preferences, leading to harmful or undesirable outputs. Preference learning, which trains models to distinguish between preferred and non-preferred responses based on human feedback, has become a crucial component for ensuring that LLMs align with human values. Despite the widespread adoption in real-world systems, a thorough theoretical understanding of the generalization guarantees for these models remain lacking. This paper bridges that gap by introducing a new theoretical framework to analyze the generalization guarantees of models trained with direct preference optimization (DPO). While existing generalization theory often focuses on overparameterized models achieving near-optimal loss or models independent of the training process, our framework rigorously assesses how well models generalize after a finite number of gradient steps, reflecting real-world LLM training practices. By analyzing the reward margin associated with each sample and its trajectory throughout training, we can effectively bound the generalization error. We derive learning guarantees showing that, under specific conditions, models trained with DPO can correctly discern preferred responses on unseen data with high probability. These insights are empirically validated on contemporary LLMs, underscoring the practical relevance of our theoretical findings.

Understanding the Learning Dynamics of Alignment with Human Feedback

Aligning large language models (LLMs) with human intentions has become a critical task for safely deploying models in real-world systems. While existing alignment approaches have seen empirical success, theoretically understanding how these methods affect model behavior remains an open question. Our work provides an initial attempt to theoretically analyze the learning dynamics of human preference alignment. We formally show how the distribution of preference datasets influences the rate of model updates and provide rigorous guarantees on the training accuracy. Our theory also reveals an intricate phenomenon where the optimization is prone to prioritizing certain behaviors with higher preference distinguishability. We empirically validate our findings on contemporary LLMs and alignment tasks, reinforcing our theoretical insights and shedding light on considerations for future alignment approaches. Disclaimer: This paper contains potentially offensive text; reader discretion is advised.

Evaluating the Utility of Model Explanations for Model Development

One of the motivations for explainable AI is to allow humans to make better and more informed decisions regarding the use and deployment of AI models. But careful evaluations are needed to assess whether this expectation has been fulfilled. Current evaluations mainly focus on algorithmic properties of explanations, and those that involve human subjects often employ subjective questions to test human's perception of explanation usefulness, without being grounded in objective metrics and measurements. In this work, we evaluate whether explanations can improve human decision-making in practical scenarios of machine learning model development. We conduct a mixed-methods user study involving image data to evaluate saliency maps generated by SmoothGrad, GradCAM, and an oracle explanation on two tasks: model selection and counterfactual simulation. To our surprise, we did not find evidence of significant improvement on these tasks when users were provided with any of the saliency maps, even the synthetic oracle explanation designed to be simple to understand and highly indicative of the answer. Nonetheless, explanations did help users more accurately describe the models. These findings suggest caution regarding the usefulness and potential for misunderstanding in saliency-based explanations.