The $K$-armed dueling bandits problem, where the feedback is in the form of noisy pairwise preferences, has been widely studied due its applications in information retrieval, recommendation systems, etc. Motivated by concerns that user preferences/tastes can evolve over time, we consider the problem of _dueling bandits with distribution shifts_. Specifically, we study the recent notion of _significant shifts_ (Suk and Kpotufe, 2022), and ask whether one can design an _adaptive_ algorithm for the dueling problem with $O(\sqrt{K\tilde{L}T})$ dynamic regret,where $\tilde{L}$ is the (unknown) number of significant shifts in preferences. We show that the answer to this question depends on the properties of underlying preference distributions. Firstly, we give an impossibility result that rules out any algorithm with $O(\sqrt{K\tilde{L}T})$ dynamic regret under the well-studied Condorcet and SST classes of preference distributions. Secondly, we show that $\text{SST}\cap \text{STI}$ is the largest amongst popular classes of preference distributions where it is possible to design such an algorithm. Overall, our results provides an almost complete resolution of the above question for the hierarchy of distribution classes.

Active inference proposes expected free energy as an objective for planning and decision-making to adequately balance exploitative and explorative drives in learning agents. The exploitative drive, or what an agent wants to achieve, is formalised as the Kullback-Leibler divergence between a variational probability distribution, updated at each inference step, and a preference probability distribution that indicates what states or observations are more likely for the agent, hence determining the agent's goal in a certain environment. In the literature, the questions of how the preference distribution should be specified and of how a certain specification impacts inference and learning in an active inference agent have been given hardly any attention. In this work, we consider four possible ways of defining the preference distribution, either providing the agents with hard or soft goals and either involving or not goal shaping (i.e., intermediate goals). We compare the performances of four agents, each given one of the possible preference distributions, in a grid world navigation task. Our results show that goal shaping enables the best performance overall (i.e., it promotes exploitation) while sacrificing learning about the environment's transition dynamics (i.e., it hampers exploration).

We study an LLM fine-tuning task for designing reward functions for sequential resource allocation problems in public health, guided by human preferences expressed in natural language. This setting presents a challenging testbed for alignment due to complex and ambiguous objectives and limited data availability. We propose DPO-PRO, a robust fine-tuning algorithm based on Direct Preference Optimization (DPO), which accounts for uncertainty in the preference distribution using a lightweight Distributionally Robust Optimization (DRO) formulation. Unlike prior DRO-based variants, DPO-PRO focuses solely on uncertainty in preferences, avoiding unnecessary conservatism and incurring negligible computational overhead. We evaluate DPO-PRO on a real-world maternal mobile health program operated by the nonprofit organization ARMMAN, as well as on standard alignment benchmarks. Experimental results demonstrate that our method consistently improves robustness to noisy preference signals compared to existing DPO variants. Moreover, DPO-PRO achieves comparable performance to prior self-reflection-based baseline for reward function design, while requiring significantly lower inference-time cost.

Inference-time alignment enables large language models (LLMs) to generate outputs aligned with end-user preferences without further training. Recent post-training methods achieve this by using small guidance models to modify token generation during inference. These methods typically optimize a reward function KL-regularized by the original LLM taken as the reference policy. A critical limitation, however, is their dependence on a pre-trained reward model, which requires fitting to human preference feedback--a potentially unstable process. In contrast, we introduce PITA, a novel framework that integrates preference feedback directly into the LLM's token generation, eliminating the need for a reward model. PITA learns a small preference-based guidance policy to modify token probabilities at inference time without LLM fine-tuning, reducing computational cost and bypassing the pre-trained reward model dependency. The problem is framed as identifying an underlying preference distribution, solved through stochastic search and iterative refinement of the preference-based guidance model. We evaluate PITA across diverse tasks, including mathematical reasoning and sentiment classification, demonstrating its effectiveness in aligning LLM outputs with user preferences.

Direct Preference Optimization (DPO) has become a popular method for fine-tuning large language models (LLMs) due to its stability and simplicity. However, it is also known to be sensitive to noise in the data and prone to overfitting. Recent works have proposed using distributionally robust optimization (DRO) to address potential noise and distributional shift in the data. However, these methods often suffer from excessive conservatism and high computational cost. We propose DPO-PRO (DPO with Preference Robustness), a robust fine-tuning algorithm based on DPO which accounts for uncertainty in the preference distribution through a lightweight DRO formulation. Unlike prior DRO-based variants, DPO-PRO focuses solely on uncertainty in preferences, avoiding unnecessary conservatism and incurring negligible computational overhead. We further show that DPO-PRO is equivalent to a regularized DPO objective that penalizes model overconfidence under weak preference signals. We evaluate DPO-PRO on standard alignment benchmarks and a real-world public health task. Experimental results show that our method consistently improves robustness to noisy preference signals compared to existing DPO variants.

We explore the use of Active Inference (AIF) as a computational user model for spatial pointing, a key problem in Human-Computer Interaction (HCI). We present an AIF agent with continuous state, action, and observation spaces, performing one-dimensional mouse pointing and clicking. We use a simple underlying dynamic system to model the mouse cursor dynamics with realistic perceptual delay. In contrast to previous optimal feedback control-based models, the agent's actions are selected by minimizing Expected Free Energy, solely based on preference distributions over percepts, such as observing clicking a button correctly. Our results show that the agent creates plausible pointing movements and clicks when the cursor is over the target, with similar end-point variance to human users. In contrast to other models of pointing, we incorporate fully probabilistic, predictive delay compensation into the agent. The agent shows distinct behaviour for differing target difficulties without the need to retune system parameters, as done in other approaches. We discuss the simulation results and emphasize the challenges in identifying the correct configuration of an AIF agent interacting with continuous systems.

The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.

Preference optimization methods like DPO have achieved remarkable performance in LLM alignment. However, the evaluation for these methods relies on a single response and overlooks other potential outputs, which could also be generated in real-world applications within this hypothetical space. To address this issue, this paper presents a \textbf{H}ypothesis-based Pr\textbf{E}ference-aware \textbf{A}na\textbf{L}ysis Framework (HEAL), a novel evaluation paradigm that formulates preference alignment as a re-ranking process within hypothesis spaces. The framework incorporates two complementary metrics: ranking accuracy for evaluating ordinal consistency and preference strength correlation for assessing continuous alignment. To facilitate this framework, we develop UniHypoBench, a unified hypothesis benchmark constructed from diverse instruction-response pairs. Through extensive experiments based on HEAL, with a particular focus on the intrinsic mechanisms of preference learning, we demonstrate that current preference learning methods can effectively capture preferences provided by proxy models while simultaneously suppressing negative samples. These findings contribute to preference learning research through two significant avenues. Theoretically, we introduce hypothesis space analysis as an innovative paradigm for understanding preference alignment. Practically, HEAL offers researchers robust diagnostic tools for refining preference optimization methods, while our empirical results identify promising directions for developing more advanced alignment algorithms capable of comprehensive preference capture.

--Traditional recommender systems aim to generate a recommendation list comprising the most relevant or similar items to the user's profile. These approaches can create recommendation lists that omit item genres from the less prominent areas of a user's profile, thereby undermining the user's experience. T o solve this problem, the calibrated recommendation system provides a guarantee of including less representative areas in the recommended list. The calibrated context works with three distributions. The first is from the user's profile, the second is from the candidate items, and the last is from the recommendation list. These distributions are G-dimensional, where G is the total number of genres in the system. This high dimensionality requires a different evaluation method, considering that traditional recommenders operate in a one-dimensional data space. In this sense, we implement fifteen models that help to understand how these distributions are structured. We evaluate the users' patterns in three datasets from the movie domain. The results indicate that the models of outlier detection provide a better understanding of the structures. The calibrated system creates recommendation lists that act similarly to traditional recommendation lists, allowing users to change their groups of preferences to the same degree. Commonly, traditional recommender systems generate recommendations with miscalibration [1]. Miscalibration means that the recommendation lists do not follow the user preferences distribution, instead suggesting items from user's dominant area of interest. It creates an overspecialized recommendation list in which the items from the less dominant area are overwhelmed. This effect puts the user in a filter bubble or an echo chamber problem [2]. For instance, when a specific area dominates the recommended list, the user likely has few other options to interact with, aside from items within that dominant area. Then, the subsequent lists are recommended, with the dominant area becoming more overspecialized. In recent years, calibrated recommendation systems have attracted attention [3]-[8] from the recommender system community to overcome this issue. This type of system demonstrates the capacity to improve several objectives, such as diversity [3], control of popularity bias [4], item coverage [5], precision [6], and the reduction of miscalibration [7]. To illustrate how calibrated recommendation works, consider a scenario: if a user's preferences distribution indicates Corresponding author is Diego Corr ˆ ea da Silva.

Committee-selection problems arise in many contexts and applications, and there has been increasing interest within the social choice research community on identifying which properties are satisfied by different multi-winner voting rules. In this work, we propose a data-driven framework to evaluate how frequently voting rules violate axioms across diverse preference distributions in practice, shifting away from the binary perspective of axiom satisfaction given by worst-case analysis. Using this framework, we analyze the relationship between multi-winner voting rules and their axiomatic performance under several preference distributions. We then show that neural networks, acting as voting rules, can outperform traditional rules in minimizing axiom violations. Our results suggest that data-driven approaches to social choice can inform the design of new voting systems and support the continuation of data-driven research in social choice.