While preference modelling is becoming one of the pillars of machine learning, the problem of preference explanation remains challenging and underexplored. In this paper, we propose PREF-SHAP, a Shapley value-based model explanation framework for pairwise comparison data. We derive the appropriate value functions for preference models and further extend the framework to model and explain context specific information, such as the surface type in a tennis game. To demonstrate the utility of PREF-SHAP, we apply our method to a variety of synthetic and real-world datasets and show that richer and more insightful explanations can be obtained over the baseline.

Preference learning methods make use of models of human choice in order to infer the latent utilities that underlie human behavior. However, accurate modeling of human choice behavior is challenging due to a range of context effects that arise from how humans contrast and evaluate options. Cognitive science has proposed several models that capture these intricacies but, due to their intractable nature, work on preference learning has, in practice, had to rely on tractable but simplified variants of the well-known Bradley-Terry model. In this paper, we take one state-of-the-art intractable cognitive model and propose a tractable surrogate that is suitable for deployment in preference learning. We then introduce a mechanism for fitting the surrogate to human data and extend it to account for data that cannot be explained by the original cognitive model. We demonstrate on large-scale human data that this model produces significantly better inferences on static and actively elicited data than existing Bradley-Terry variants. We further show in simulation that when using this model for preference learning, we can significantly improve utility in a range of real-world tasks.

To demonstrate the utility ofPREF-SHAP, we apply our method to a variety of synthetic and real-worlddatasets andshowthatricher andmoreinsightful explanations canbe obtainedoverthebaseline.

The ability of LLMs to represent diverse perspectives is critical as they increasingly impact society. However, recent studies reveal that alignment algorithms such as RLHF and DPO significantly reduce the diversity of LLM outputs. Not only do aligned LLMs generate text with repetitive structure and word choice, they also approach problems in more uniform ways, and their responses reflect a narrower range of societal perspectives. We attribute this problem to the KL divergence regularizer employed in preference learning algorithms. This causes the model to systematically overweight majority opinions and sacrifice diversity in its outputs. To address this, we propose Soft Preference Learning, which decouples the entropy and cross-entropy terms in the KL penalty -- allowing for fine-grained control over LLM generation diversity. From a capabilities perspective, LLMs trained using Soft Preference Learning attain higher accuracy on difficult repeated sampling tasks and produce outputs with greater semantic and lexical diversity. From an alignment perspective, they are capable of representing a wider range of societal viewpoints and display improved logit calibration. Notably, Soft Preference Learning resembles, but is a Pareto improvement over, standard temperature scaling. As LLMs become integrated into how people consume information (Bick et al., 2024) and approach tasks (Deloitte, 2024), their ability to represent diverse perspectives is critical. For example, consider an LLM answering the following multiple-choice question: The best way to reduce income inequality is: (A) Increase minimum wage (B) Expand access to education and job training (C) Implement universal basic income (D) Lower taxes on the wealthy to stimulate job creation Imagine a survey showing people's preferences as: A (55%), B (20%), C (15%), and D (10%). How should an LLM respond to this question? Ideally, we may prefer it to reflect the range of views in the population. If an LLM assigns 99% probability to majority option A, it fails to represent the diversity of perspectives. With LLMs becoming important information sources, this may reinforce dominant narratives at the expense of minority views. However, recent studies show that alignment algorithms such as RLHF and DPO significantly reduce the diversity of LLM outputs. This leads to mode collapse towards majority preferences, as the example above shows (Kirk et al., 2024; Padmakumar & He, 2024; Rafailov et al., 2024; Christiano et al., 2023). In a generative setting, this results in repetitive responses, as illustrated in Figure 1. For example, the DPO model frequently uses the same doctor's name and 1 We highlight Doctor name, gender, and textual aberration features shown in the plots on the right. DPO responses are well-formed but lack diversity (e.g.

Active preference learning offers an efficient approach to modeling preferences, but it is hindered by the cold-start problem, which leads to a marked decline in performance when no initial labeled data are available. While cold-start solutions have been proposed for domains such as vision and text, the cold-start problem in active preference learning remains largely unexplored, underscoring the need for practical, effective methods. Drawing inspiration from established practices in social and economic research, the proposed method initiates learning with a self-supervised phase that employs Principal Component Analysis (PCA) to generate initial pseudo-labels. This process produces a \say{warmed-up} model based solely on the data's intrinsic structure, without requiring expert input. The model is then refined through an active learning loop that strategically queries a simulated noisy oracle for labels. Experiments conducted on various socio-economic datasets, including those related to financial credibility, career success rate, and socio-economic status, consistently show that the PCA-driven approach outperforms standard active learning strategies that start without prior information. This work thus provides a computationally efficient and straightforward solution that effectively addresses the cold-start problem.

Preference learning from pairwise feedback is a widely adopted framework in applications such as reinforcement learning with human feedback and recommendations. In many practical settings, however, user interactions are limited or costly, making offline preference learning necessary. Moreover, real-world preference learning often involves users with different preferences. For example, annotators from different backgrounds may rank the same responses differently. This setting presents two central challenges: (1) identifying similarity across users to effectively aggregate data, especially under scenarios where offline data is imbalanced across dimensions, and (2) handling the imbalanced offline data where some preference dimensions are underrepresented. To address these challenges, we study the Offline Clustering of Preference Learning problem, where the learner has access to fixed datasets from multiple users with potentially different preferences and aims to maximize utility for a test user. To tackle the first challenge, we first propose Off-C$^2$PL for the pure offline setting, where the learner relies solely on offline data. Our theoretical analysis provides a suboptimality bound that explicitly captures the tradeoff between sample noise and bias. To address the second challenge of inbalanced data, we extend our framework to the setting with active-data augmentation where the learner is allowed to select a limited number of additional active-data for the test user based on the cluster structure learned by Off-C$^2$PL. In this setting, our second algorithm, A$^2$-Off-C$^2$PL, actively selects samples that target the least-informative dimensions of the test user's preference. We prove that these actively collected samples contribute more effectively than offline ones. Finally, we validate our theoretical results through simulations on synthetic and real-world datasets.

Frontier large language models such as GPT-5 and Claude Sonnet remain prone to overconfidence even after alignment through Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO). For instance, they tend to offer the same conservative answer "No" to both questions "Can I eat the [food / potato chips] that has been left out overnight?" despite the latter requiring no refridgeration for safe consumption. We find that this failure is potentially attributed to a limitation of existing preference learning: it emphasizes selecting the correct response for a given prompt, while neglecting counterfactual prompts that should alter the response. To address this limitation, we propose abductive preference learning, a fine-tuning paradigm that reverses the conventional conditioning by learning preferences over prompts given a response. To validate this idea, we construct an abductive dataset derived from the HaluEval QA benchmark with 1,001 entries, implementing abductive DPO and its variant DPOP. Experiments reveal complementary strengths: standard methods improve response selection, abductive methods improve prompt discrimination, while a multitask objective unifies both. On the abductive dataset, multitask DPOP boosts accuracy from $90.0\%$ to $99.5\%$ in response selection and $54.7\%$ to $85.0\%$ in prompt discrimination, with qualitative evidence highlighting improved sensitivity to prompt differences. Finally, evaluation on AlpacaEval shows multitask DPOP improves win rate (from $5.26\%$ to $6.17\%$), confirming that abductive preference learning preserves the benefits of conventional preference optimization while addressing the overlooked challenge of counterfactual prompts.

As everyday use cases of large language model (LLM) AI assistants have expanded, it is becoming increasingly important to personalize responses to align to different users' preferences and goals. While reinforcement learning from human feedback (RLHF) is effective at improving LLMs to be generally more helpful and fluent, it does not account for variability across users, as it models the entire user population with a single reward model, meaning it assumes that everyone's preferences are the same. We present a novel framework, Preference Learning Using Summarization (PLUS), that uses reinforcement learning (RL) to learn to produce text-based summaries of each user's preferences, characteristics, and past conversations. These summaries condition the reward model, enabling it to make personalized predictions about the types of responses valued by each user. Both the user-summarization model and reward model are trained simultaneously, creating an online co-adaptation loop. We show that in contrast to the standard Bradley-Terry model, summaries produced by PLUS capture diverse aspects of user preferences, achieving a 11-77% improvement in reward model accuracy. Key strengths of PLUS are: (1) robust performance with new users and conversation topics, achieving a 25% improvement over the best personalized RLHF technique; (2) zero-shot personalization with state-of-the-art proprietary models like GPT -4 (e.g., PLUS-summary-conditioned responses achieved a 72% win rate compared to 28% for default GPT -4o); (3) learning from flexible user contexts beyond preference labels, and (4) interpretable representation of users, enabling greater transparency and user control in pluralistic LLM alignment.

Advancing AI systems in scientific domains like physics, materials science, and engineering calls for reasoning over complex, multi-physics phenomena while respecting governing principles. Although Large Language Models (LLMs) and existing preference optimization techniques perform well on standard benchmarks, they often struggle to differentiate between physically valid and invalid reasoning. This shortcoming becomes critical in high-stakes applications like metal joining, where seemingly plausible yet physically incorrect recommendations can lead to defects, material waste, equipment damage, and serious safety risks. To address this challenge, we introduce PKG-DPO, a novel framework that integrates Physics Knowledge Graphs (PKGs) with Direct Preference Optimization (DPO) to enforce physical validity in AI-generated outputs. PKG-DPO comprises three key components A) hierarchical physics knowledge graph that encodes cross-domain relationships, conservation laws, and thermodynamic principles. B) A physics reasoning engine that leverages structured knowledge to improve discrimination between physically consistent and inconsistent responses. C) A physics-grounded evaluation suite designed to assess compliance with domain-specific constraints. PKG-DPO achieves 17% fewer constraint violations and an 11% higher Physics Score compared to KG-DPO (knowledge graph-based DPO). Additionally, PKG-DPO demonstrates a 12\% higher relevant parameter accuracy and a 7% higher quality alignment in reasoning accuracy. While our primary focus is on metal joining, the framework is broadly applicable to other multi-scale, physics-driven domains, offering a principled approach to embedding scientific constraints into preference learning.