As large language models (LLMs) are increasingly used as personalized user assistants, effectively adapting to users' evolving preferences is critical for delivering high-quality personalized responses. While user preferences are often stable in content, their relative strengths shift over time due to changing goals and contexts. Therefore, modeling these dynamic preference strengths can enable finer-grained personalization. However, current methods face two major challenges: (i) limited user feedback makes it difficult to estimate preference strengths accurately, and (ii) natural language ambiguity limits the controllability of preference-guided generation. To address these issues, we propose AdaPA-Agent, a LLM-agent personalization framework that models dynamic preference strengths via Adaptive Preference Arithmetic. First, instead of requiring additional user feedback, AdaPA-Agent employs an alignment-based strength estimation module to estimate the strength of user preferences from the existing user-agent interaction. Then, it guides controllable personalized generation by linearly combining next-token distributions, weighted by the estimated strengths of individual preferences. Experiments on two personalization tasks-conversational recommendation and personalized web interaction-demonstrate that AdaPA-Agent better aligning with users' changing intents, and has achieved over 18.9% and 14.2% improvements compared to ReAct, the widely-used agent framework.

Binary choices, as often used for reinforcement learning from human feedback (RLHF), convey only the direction of a preference. A person may choose apples over oranges and bananas over grapes, but which preference is stronger? Strength is crucial for decision-making under uncertainty and generalization of preference models, but hard to measure reliably. Metadata such as response times and interannotator agreement can serve as proxies for strength, but are often noisy and confounded. We propose ResponseRank to address the challenge of learning from noisy strength signals. Our method uses relative differences in proxy signals to rank responses to pairwise comparisons by their inferred preference strength. To control for systemic variation, we compare signals only locally within carefully constructed strata. This enables robust learning of utility differences consistent with strengthderived rankings while making minimal assumptions about the strength signal. Our contributions are threefold: (1) ResponseRank, a novel method that robustly learns preference strength by leveraging locally valid relative strength signals; (2) empirical evidence of improved sample efficiency and robustness across diverse tasks: synthetic preference learning (with simulated response times), language modeling (with annotator agreement), and RL control tasks (with simulated episode returns); and (3) the Pearson Distance Correlation (PDC), a novel metric that isolates cardinal utility learning from ordinal accuracy.

Although binary choices are simple and widely used, they provide limited information about preference strength. To address this, we leverage human response times, which are inversely related to preference strength, as an additional signal. We propose a computationally efficient method that combines choices and response times to estimate human utility functions, grounded in the EZ diffusion model from psychology. Theoretical and empirical analyses show that for queries with strong preferences, response times complement choices by providing extra information about preference strength, leading to significantly improved utility estimation. We incorporate this estimator into preference-based linear bandits for fixed-budget best-arm identification. Simulations on three real-world datasets demonstrate that using response times significantly accelerates preference learning compared to choice-only approaches.

In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength.

The usage-based constructionist (UCx) approach to language posits that language comprises a network of learned form-meaning pairings (constructions) whose use is largely determined by their meanings or functions, requiring them to be graded and probabilistic. This study investigates whether the internal representations in Large Language Models (LLMs) reflect the proposed function-infused gradience. We analyze representations of the English Double Object (DO) and Prepositional Object (PO) constructions in Pythia-$1.4$B, using a dataset of $5000$ sentence pairs systematically varied by human-rated preference strength for DO or PO. Geometric analyses show that the separability between the two constructions' representations, as measured by energy distance or Jensen-Shannon divergence, is systematically modulated by gradient preference strength, which depends on lexical and functional properties of sentences. That is, more prototypical exemplars of each construction occupy more distinct regions in activation space, compared to sentences that could have equally well have occured in either construction. These results provide evidence that LLMs learn rich, meaning-infused, graded representations of constructions and offer support for geometric measures for representations in LLMs.

Although binary choices are simple and widely used, they provide limited information about preference strength. To address this, we leverage human response times, which are inversely related to preference strength, as an additional signal. We propose a computationally efficient method that combines choices and response times to estimate human utility functions, grounded in the EZ diffusion model from psychology. Theoretical and empirical analyses show that for queries with strong preferences, response times complement choices by providing extra information about preference strength, leading to significantly improved utility estimation. We incorporate this estimator into preference-based linear bandits for fixed-budget best-arm identification.

Reinforcement learning from human feedback (RLHF) (Christiano et al., 2017; Ziegler et al., 2019; Ouyang et al., 2022) has played a pivotal role in aligning large language models (LLMs) (Kenton et al., 2021), enhancing specific capabilities of LLMs in various fields, such as summarization (Stiennon et al., 2020), coding (Gao et al., 2023), and medical assistance (Moor et al., 2023). A crucial component of the RLHF process is the reward model, which serves as the primary mechanism for integrating human preferences and feedback into the learning process (Wang et al., 2024). The reward model guides the optimization procedure of RLHF towards objectives aligned with human preferences (Kaufmann et al., 2023). Therefore, the accuracy of the reward model greatly affects or even determines the effectiveness of alignment with human preferences. Moreover, the direct preference optimization (DPO) method (Rafailov et al., 2024) utilizes LLMs to implicitly represent the reward model through mathematical transformations, bypassing the complex RL optimization phase and focusing solely on the reward modeling phase. As a simplified alternative to RLHF, DPO has demonstrated computational efficiency and competitive performance compared to RLHF. To learn a reward model from human preferences, obtaining high-quality human preference data is crucial (Wang et al., 2024), typically achieved by having human labelers annotate previously collected data consisting of a prompt and a pair of responses (Ouyang et al., 2022; Bai et al., 2022).

Reinforcement learning from human feedback (RLHF) is a prevalent approach to align AI systems with human values by learning rewards from human preference data. Due to various reasons, however, such data typically takes the form of rankings over pairs of trajectory segments, which fails to capture the varying strengths of preferences across different pairs. In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength. By incorporating an adaptive scaling parameter into the loss for each pair, our method increases the flexibility of the reward function. Specifically, it assigns small scaling parameters to pairs with ambiguous preferences, leading to more comparable rewards, and large scaling parameters to those with clear preferences for more distinct rewards. Computationally, our proposed loss function is strictly convex and univariate with respect to each scaling parameter, enabling its efficient optimization through a simple second-order algorithm. Our method is versatile and can be readily adapted to various preference optimization frameworks, including direct preference optimization (DPO). Our experiments with robotic control and natural language generation with large language models (LLMs) show that our method not only improves policy performance but also aligns reward function selection more closely with policy optimization, simplifying the hyperparameter tuning process.

Aligning with human preference datasets has been critical to the success of large language models (LLMs). Reinforcement learning from human feedback (RLHF) employs a costly reward model to provide feedback for on-policy sampling responses. Recently, offline methods that directly fit responses with binary preferences in the dataset have emerged as alternatives. However, existing methods do not explicitly model preference strength information, which is crucial for distinguishing different response pairs. To overcome this limitation, we propose Online Self-Preferring (OSP) language models to learn from self-generated response pairs and self-judged preference strengths. For each prompt and corresponding self-generated responses, we introduce a ranked pairing method to construct multiple response pairs with preference strength information. We then propose the soft-preference cross-entropy loss to leverage such information. Empirically, we demonstrate that leveraging preference strength is crucial for avoiding overfitting and enhancing alignment performance. OSP achieves state-of-the-art alignment performance across various metrics in two widely used human preference datasets. OSP is parameter-efficient and more robust than the dominant online method, RLHF when limited offline data are available and generalizing to out-of-domain tasks. Moreover, OSP language models established by LLMs with proficiency in self-preferring can efficiently self-improve without external supervision.