Preference Datasets

Using reinforcement learning to post-train joint video-audio generation models requires a reward signal. Existing methods construct this reward by combining metrics for individual quality dimensions, including audio quality, visual fidelity, and synchronization. However, these metrics evaluate perceptual dimensions separately and fail to capture the overall semantic and temporal coherence among the text prompt, video, and audio that shapes human preferences. Optimizing models against these metrics encourages reward hacking, generating video-audio content that achieves high scores on these metrics yet appears incoherent or unfaithful to human viewers. To address this problem, we first construct a large-scale human-preference dataset VAPref-10K for joint video-audio generation, comprising 9K prompts and 10.3K fine-grained paired comparisons from open-source generation models. We also introduce the VA-Judger-Bench benchmark with both in-domain and out-of-domain model comparisons to evaluate whether reward models truly align with human preferences. We further propose VA-Judger, a chain-of-thought omni-reward model for joint video-audio generation. In particular, VA-Judger first learns from pairs with clear quality gaps to establish structured output and coarse preference discrimination, then distills reliable preference explanations for harder near-quality comparisons via rejection sampling verified against human annotations, and finally performs dimension-wise reinforcement learning that decomposes human feedback into individual quality dimensions for denser reward signals than a single binary preference label. Experiments show that VA-Judger outperforms metric baselines in predicting human preferences on both in-domain and out-of-domain evaluations. Using its human-aligned rewards for post-training audio-video generation model also yields significant improvements in generation quality.

This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.

We introduce Rushes, a dataset and benchmark for studying revealed human engagement preferences in interactive narrative environments. Rushes is collected through a game interface where users interact with AI-generated branching narratives and select one choice from a small, explicit candidate set at each decision point. Each interaction logs the full candidate set, the user's choice, and the evolving narrative context, yielding time-ordered trajectories with persistent user-level identifiers. Rushes contains 44,226 decision events from 8,167 unique users across six games, capturing sequential, personalized engagement behavior rather than static judgments. We show that user choices exhibit structured, non-random patterns, quantified by a low choice entropy relative to a uniform baseline. We position Rushes as a diagnostic benchmark for pluralistic alignment and demonstrate a robust Engagement Gap: state-of-the-art LLMs, including GPT-5, fail to outperform simple baselines. While classical Matrix Factorization (SVD) captures measurable personalized signal (37.7%), frontier LLMs (34.23%) struggle to even match the Popularity Baseline (36.4%) on event-level choice prediction. This gap suggests that single, population-level objectives, like those used in modern RLHF, appear insufficient to capture heterogeneous, context-dependent engagement signals. As a result, even highly capable models default to majority preferences rather than adapting to individual trajectories. We release Rushes to support research into pluralistic alignment and sequential decision-making in generative systems. The full code for the platform and dataset will be available here: https://github.com/microsoft/rushes

Value-aware AI systems require explicit computational representations of human values (groundings) and their aggregation into value systems in order to align their decisions with ours. As such representations are difficult to elicit, value learning seeks to infer them by observing human behaviour. This work addresses the lack of grounded value learning methods in generative AI: existing approaches typically replicate human preferences without awareness of the multidimensional structure of value alignment, or lack principled value system elicitation methods. To address these gaps, we adapt a previously validated value system learning method to the generative AI setting, which, based on pairwise prompt-response preference data, simultaneously learns: i) an implementation of a grounding for a set of values given by a multi-objective reward model, and ii) a value system representation in the form of a weighted linear scalarization of the previous grounding model. To ensure that the learned value systems are based on coherent value representations, our algorithm dynamically prioritizes the grounding learning process. We evaluate the method against baselines and a contemporary method on prompt-response preference datasets. Results show competitive performance and minimal trade-offs against the baselines, while improving explainability.

Music aesthetics scoring plays a critical role in applications such as dataset curation, generative model evaluation, and reward modeling for music generation. Recent approaches rely on deep neural networks trained on human-annotated ratings, but these models may exploit spurious correlations rather than capturing perceptually meaningful aesthetics. In this work, we identify a previously underexplored failure mode in music evaluation models: genre-induced shortcut learning. Through a systematic analysis of SongEval, we show that biases in training data lead to strong correlations between genre-related features and predicted scores, causing the model to use them as a proxy for aesthetics. This results in systematic overestimation of pop music and undervaluation of high-quality samples from other genres, leading to predictions that are inconsistent with human preferences. To address this issue, we propose a training objective that jointly reweights hard samples and regularizes group-level performance, encouraging the model to learn genre-invariant representations of musicality. Experimental results demonstrate that our method reduces genre-dependent bias and improves alignment with human preferences, as reflected by gains in both cross-genre and within-genre preference alignment.

Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems. We introduce PLURAL, a large-scale, value-focused preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries. Using a two-stage generation pipeline, we transform survey responses into synthetic preference triplets that preserve normative value signals while producing realistic scenarios. We release an initial version of PLURAL containing ~500,000 preference triplets representing people in 20 diverse countries. We evaluate PLURAL in three ways: (i) dataset-level validation showing that it preserves both cross-country value differences and within-country diversity from the original survey; (ii) automated evaluation showing that training on PLURAL improves alignment with target countries' cultural profiles, reducing mean absolute error by up to 27.7% relative to strong baselines; and (iii) blind human evaluation with 176 evaluators in India, Brazil, and Japan, who judge PLURAL-aligned responses as more representative of their national values. Together, these results show that PLURAL contains learnable signal for value steering, offering a scalable resource for pluralistic alignment. Dataset: https://huggingface.co/datasets/agdhruv/plural-alignment

We describe our entry to the efficiency track of the Academic Text-to-Music (ATTM) Grand Challenge at ICME 2026. Beyond the challenge protocol's FAD-CLAP and CLAP score, we add a learned human-preference reward from TuneJury, a twin pairwise ranker trained over open music-preference datasets. The reward serves both as a training-time conditioning signal and as a sample-selection criterion. The pipeline combines five engineering decisions on a 120M-parameter FluxAudio-S backbone, four at training time and one at inference: (i) training-time reward conditioning that doubles as an inference-time CFG axis, (ii) a sweep over five score-conditioning architectures, where training and inference use different variants, (iii) expert iteration on the top decile, (iv) a short preference-tuning pass (CRPO) for audio-text alignment, and (v) inference post-processing via joint CFG, source separation, and loudness normalization. Per-stage decomposition on 100 Song Describer prompts shows training-time reward conditioning as a functional conditioning axis, expert iteration as the dominant contributor, the preference-tuning pass adding only noise-level gain, and the inference-time score scalar already saturated by the end of the chain.

We introduce TuneJury, an open, instance-level pairwise reward model for text-to-music that predicts a music preference score from a text prompt and an audio clip. The released checkpoint is trained on publicly available human-preference labels covering arena-style (A vs. B) votes, metric-alignment preference pairs, crowdsourced pairwise comparisons, and expert aesthetic ratings. The predicted score margin between two clips is well calibrated on our held-out test split, supporting data filtering via a simple score threshold. TuneJury generalizes to both held-out test pairs and out-of-distribution benchmarks, remaining competitive with prior baselines on the latter. For generators released after training, we introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration that recovers agreement at substantially better data efficiency than from-scratch retraining. The same frozen reward drives consistent reward-axis gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. TuneJury is available at https://github.com/yonghyunk1m/TuneJury.

Reward models guide text-to-image (T2I) systems toward outputs aligned with human preferences. However, typical reward models such as HPSv3 are trained on pre-annotated data from earlier T2I models, without accounting for quality discriminative shifts arising from evolving model capabilities and reinforcement learning (RL) iterations, limiting their broader applicability. In this work, we propose HPSv3++, a reward model framework that elevates the HPSv3 model for varying T2I model capabilities and their RL iteration changes across the full capability-iteration spectrum. Specifically, we first introduce HPDv3++, a 212K dual-dimension preference dataset annotated for text fidelity and aesthetic quality using a recent high-capability (Qwen-Image) model with human supervision. We then propose a two-stage training framework. Stage 1 employs data-aware orthogonal gradient projection to incorporate diverse aesthetic perception from HPDv3++ while preserving the original effective human preference knowledge in HPSv3. Stage 2 further leverages unlabeled data from T2I models spanning different capability levels and RL iterations, and introduces a joint capability-iterations conditioned signal for the reward model together with a standard deviation-driven unsupervised guidance mechanism, strengthening reward model across the capability-iteration spectrum. HPSv3++ achieves state-of-the-art preference prediction, outperforming HPSv3 9.8% on HPDv3, 5.5% on GenAI-Bench, while achieving 79.1%/88.1% on our proposed HPDv3++. When used for T2I RL training, it consistently improves GenEval scores across diverse T2I models, demonstrating its wide-range capabilities. The code is available at https://github.com/PlantPotatoOnMoon/HPSv3-PlusPlus.

Language-model post-training is the main stage at which model behavior is shaped, yet it still largely involves optimization of scalar rewards that summarize diverse desiderata. This abstraction gives practitioners little visibility into what their data actually teaches models, allowing spurious correlations to be learned by a model and inducing undesirable behaviors such as over-stylization and sycophancy. To address this problem, we ask: can we inspect a preference dataset before optimization and decide, at the level of concepts, which behaviors a model should be allowed to learn? Motivated by this, we introduce a data-centric post-training pipeline that uses interpretability protocols to develop statistical hypotheses for the latent concepts separating preferred from dispreferred generations, making them explicit for fine-grained user feedback. Building on this view, we unify several interpretability-based training protocols as ways of shaping rewards via feature or data interventions. Empirically, we show that our pipeline diagnoses undesirable signals in existing preference data, mitigates off-target learning, and can also help amplify or shape desired properties such as safeguards and model personality. More broadly, our results suggest that interpretability can turn post-training from optimizing opaque proxy rewards into a process of auditing and sculpting the learning signal itself.

Safety alignment for large language models relies on preference data, but current pipelines often train on large, redundant datasets. Existing data selection methods typically score each preference pair independently, collapsing directional preference information into scalar quality or diversity scores. This sample-centric view is especially limiting in multi-dataset settings, where shared safety directions coexist with dataset-specific residual risks. We propose DOG-DPO, a training-free data selection framework that treats preference pairs as structured geometric signals. DOG-DPO first represents each preference pair as a direction in model representation space. It then decomposes multi-dataset preference geometry into a global anchor subspace and dataset-specific residual subspaces. Finally, it selects subsets by maximizing diversity-based coverage, encouraging broad, non-redundant coverage of alignment directions before DPO training. Across six safety benchmarks and two model backbones, DOG-DPO achieves a strong utility-robustness trade-off using only 11% of the preference pairs. It recovers most of the safety gains of full-data training while remaining entirely teacher-free, training-free, and substantially faster than representative selection baselines.

Given the increased adoption of Vision Language Models (VLMs) in human-interactive settings, it is important that we evaluate how well these models can adapt to real-time preferences for different users. While an increasing number of vision-language benchmarks have recently been introduced, they focus largely on evaluating static capabilities and generally-held preferences learned from extensive training data. This work introduces a new benchmark for evaluating the ability of VLMs to understand dynamic human-preferences, i.e. preferences that are passed in-context at inference time. We provide an automated pipeline for generating this benchmark with variations on image dependence, a dynamic multi-modal human-preference dataset, and evaluations of state-of-the-art models on the novel benchmark.

Preference alignment is a crucial post-training step for large language models (LLMs) to ensure their outputs align with human values. However, post-training on real human preference data raises privacy concerns, as these datasets often contain sensitive user prompts and human judgments. To address this, we propose DPPrefSyn, a novel algorithm for generating differentially private (DP) synthetic preference data to enable privacy-preserving preference alignment. DPPrefSyn is a principled framework grounded in the Bradley-Terry preference model and the intrinsic geometric structure of pairwise human preference data. It first learns an underlying preference model from private data with formal differential privacy guarantees, and then leverages the learned model together with public prompts to synthesize high-quality preference data. It exploits the shared linear structure of per-cluster reward models to effectively capture heterogeneous human preferences in private datasets, and leverages DP Principal Component Analysis (DP-PCA) to improve learning accuracy. Extensive experimental results demonstrate that DPPrefSyn achieves competitive alignment performance under strong DP guarantees. These findings highlight the potential of synthetic preference data as a practical alternative for privacy-preserving preference alignment across a broad range of applications. To the best of our knowledge, this is the first work to generate DP synthetic preference data for LLM alignment. Our code is available at https://github.com/gfengyu/Differentially-Private-Preference-Data-Synthesis.

Despite the rapid progress of text-to-image (T2I) models, generating images that accurately reflect complex compositional prompts (covering attribute bindings, object relationships, counting) still remains challenging. To address this, we propose BiDPO, a framework to enhance T2I model's capability of compositional text-to-image generation. We begin by introducing an carefully designed pipeline to construct a large-scale preference dataset, BiComp, with strictly quality control. Then, we extend Diffusion DPO to jointly optimize image and text preferences, which is shown to greatly effective in improving the models to follow complex text prompt in generation. To further enhance the models for fine-grained alignment, we employ a region-level guidance method to focus on regions relevant to compositional concepts. Experimental results demonstrate that our BiDPO substantially improves compositional fidelity, consistently outperforming prior methods across multiple benchmarks. Our approach highlights the potential of preference-based fine-tuning for complex text-to-image tasks, offering a flexible and scalable alternative to existing techniques.

Reinforcement Learning from Human Feedback (RLHF) can reveal implicit objectives such as safety considerations that go beyond task completion. In this work, we focus on the common safety criteria embedded in crowd preference datasets, where different users may express distinct preferences or objectives, yet follow similar safety principles. Our aim is to discover shared safety criteria from crowd preferences and then transfer them to downstream RL tasks to regularize agent behavior and enforce safety. We first show that direct reward combination-optimizing a preference-learned reward model together with downstream task rewards-has inherent limitations. Motivated by this, we propose Safe Crowd Preference-based RL, a hierarchical framework that extracts safety-aligned skills from crowd preferences and composes them via a high-level policy to safely solve downstream tasks. Experiments across safe RL environments and a preliminary LLM-style task with diverse user goals and shared safety constraints demonstrate that our approach substantially lowers safety costs without access to explicit safety rewards, while achieving task performance comparable to oracle methods trained with ground-truth safety signals.

Text-to-image models now generate graphic design at production scale, yet their supervision still comes primarily from photo-style preference datasets with a single overall verdict per comparison. Designers evaluate designs along several distinct axes (e.g., typography, layout, color harmony) that a single preference label collapses. We release \emph{TASTE} \textit{(Typography, Aesthetics, Spatial, Tone, Etc.)}, a multi-dimensional preference dataset in which two disjoint cohorts of five professional designers each ranked outputs from four current text-to-image models across nine criteria along with per-image hallucination flags. We pair the dataset with two contributions. First, a criterion-agnostic signal-validation framework based on Kendall's $τ$, majority-vote probability, and Condorcet cycles against exact iid-uniform nulls; the analysis reveals significant but moderate designer agreement, with every TASTE criterion rejecting the random-rater null. Second, we benchmark preference models on TASTE and find that off-the-shelf VLM judges and dedicated T2I scorers fail to reach majority agreement with the designer panel, while a small MLP head trained directly on TASTE substantially narrows the gap to the single-rater ceiling, setting a baseline for future TASTE-trained preference models.

Aligning Text-to-Image (T2I) generation models with human preferences increasingly relies on image reward models that score or rank generated images according to prompt alignment and perceptual quality. Existing reward models are commonly trained as Bradley-Terry (BT) preference models on large-scale human preference corpora, making them costly to train, difficult to adapt, and opaque in their evaluation criteria. Meanwhile, Vision-Language Model (VLM) judges can provide more fine-grained assessments through textual rubrics, but their manually designed or heuristically generated scoring rules may fail to reliably reflect human preferences. In this paper, we propose AutoRubric-T2I, the first rubric learning framework in T2I that automatically synthesizes and selects explicit rubrics for guiding VLM judges. AutoRubric-T2I first synthesizes reasoning traces from preference pairs into candidate rubrics, then uses a VLM judge to score paired images under each rubric, producing pairwise rubric-score differences for preference learning. To remove noisy and redundant rules, we further employ a $\ell_1$-Regularized Logistic Regression Refiner, which selects the Top-$N$ most discriminative rubrics. Extensive evaluations show that AutoRubric-T2I produces high-quality, interpretable reward signals using less than 0.01% of the annotated preference data, substantially reducing the need for large-scale reward-model training. On image reward benchmarks such as MMRB2, AutoRubric-T2I outperforms strong reward model baselines. We further validate AutoRubric-T2I as an RL reward on downstream T2I tasks, including TIIF and UniGenBench++, where it improves generation quality over scalar reward models using the Flow-GRPO pipeline on diffusion models.

Evaluating instruction-guided image edits requires rewards that reflect subtle human preferences, yet current reward models typically depend on large-scale preference annotation and additional model training. This creates a data-efficiency gap: humans can often infer the target evaluation criteria from only a few examples, while models are usually trained on hundreds of thousands of comparisons. We present RewardHarness, a self-evolving agentic reward framework that reframes reward modeling as context evolution rather than weight optimization. Instead of learning from large-scale annotations, RewardHarness aligns with human preferences by iteratively evolving a library of tools and skills from as few as 100 preference demonstrations. Given a source image, candidate edited images, and an editing instruction, an Orchestrator selects the most relevant subset of tools and skills from the maintained library, and a frozen Sub-Agent uses them to construct a reasoning chain that produces a preference judgment. By comparing predicted judgments with ground-truth preferences and analyzing successes and failures in the reasoning process, the Orchestrator automatically refines its library of tools and skills without additional human annotation. Using only 0.05% of the EditReward preference data, RewardHarness achieves 47.4% average accuracy on image-editing evaluation benchmarks, surpassing GPT-5 by 5.3 points. When used as a reward signal for GRPO fine-tuning, RL-tuned models achieve 3.52 on ImgEdit-Bench. Project page: https://rewardharness.com.

Story generation aims to automatically produce coherent, structured, and engaging narratives. Although large language models (LLMs) have significantly advanced text generation, stories generated by LLMs still diverge from human-authored works regarding complex narrative structure and human-aligned preferences. A key reason is the absence of effective modeling of human story preferences, which are inherently subjective and under-explored. In this work, we systematically evaluate the modeling of human story preferences and introduce StoryRMB, the first benchmark for assessing reward models on story preferences. StoryRMB contains $1,133$ high-quality, human-verified instances, each consisting of a prompt, one chosen story, and three rejected stories. We find existing reward models struggle to select human-preferred stories, with the best model achieving only $66.3\%$ accuracy. To address this limitation, we construct roughly $100,000$ high-quality story preference pairs across diverse domains and develop StoryReward, an advanced reward model for story preference trained on this dataset. StoryReward achieves state-of-the-art (SoTA) performance on StoryRMB, outperforming much larger models. We also adopt StoryReward in downstream test-time scaling applications for best-of-n (BoN) story selection and find that it generally chooses stories better aligned with human preferences. We will release our dataset, model, and code to facilitate future research. Related code and data are available at https://github.com/THU-KEG/StoryReward.

Music popularity prediction has attracted growing research interest, with relevance to artists, platforms, and recommendation systems. However, the explosive rise of AI-generated music platforms has created an entirely new and largely unexplored landscape, where a surge of songs is produced and consumed daily without the traditional markers of artist reputation or label backing. Key, yet unexplored in this pursuit is aesthetic quality. We propose APEX, the first large-scale multi-task learning framework for AI-generated music, trained on over 211k songs (10k hours of audio) from Suno and Udio, that jointly predicts engagement-based popularity signals - streams and likes scores - alongside five perceptual aesthetic quality dimensions from frozen audio embeddings extracted from MERT, a self-supervised music understanding model. Aesthetic quality and popularity capture complementary aspects of music that together prove valuable: in an out-of-distribution evaluation on the Music Arena dataset, comprising pairwise human preference battles across eleven generative music systems unseen during training, including aesthetic features consistently improves preference prediction, demonstrating strong generalisation of the learned representations across generative architectures.

While preference optimization is crucial for improving visual generative models, how to effectively scale this paradigm remains largely unexplored. Current open-source preference datasets contain conflicting preference patterns, where winners excel in some dimensions but underperform in others. Naively optimizing on such noisy datasets fails to learn preferences, hindering effective scaling. To enhance robustness against noise, we propose Poly-DPO, which extends the DPO objective with an additional polynomial term that dynamically adjusts model confidence based on dataset characteristics, enabling effective learning across diverse data distributions. Beyond biased patterns, existing datasets suffer from low resolution, limited prompt diversity, and imbalanced distributions. To facilitate large-scale visual preference optimization by tackling data bottlenecks, we construct ViPO, a massive-scale preference dataset with 1M image pairs at 1024px across five categories and 300K video pairs at 720p+ across three categories. State-of-the-art generative models and diverse prompts ensure reliable preference signals with balanced distributions. Remarkably, when applying Poly-DPO to our high-quality dataset, the optimal configuration converges to standard DPO. This convergence validates dataset quality and Poly-DPO's adaptive nature: sophisticated optimization becomes unnecessary with sufficient data quality, yet remains valuable for imperfect datasets. We validate our approach across visual generation models. On noisy datasets like Pick-a-Pic V2, Poly-DPO achieves 6.87 and 2.32 gains over Diffusion-DPO on GenEval for SD1.5 and SDXL, respectively. For ViPO, models achieve performance far exceeding those trained on existing open-source preference datasets. These results confirm that addressing both algorithmic adaptability and data quality is essential for scaling visual preference optimization.

Large Vision-Language Models (LVLMs) frequently suffer from hallucinations. Existing preference learning-based approaches largely rely on proprietary models to construct preference datasets. We identify that this reliance introduces a distributional mismatch between the proprietary and target models that hinders efficient alignment. To address this, we propose Alignment via VErified Self-correction DPO (AVES-DPO), a framework that aligns LVLMs using in-distribution data derived from the model's intrinsic knowledge. Our approach employs a consensus-based verification mechanism to diagnose diverse hallucinations and guides the model to self-correct, thereby generating preference pairs strictly compatible with its internal distribution. Extensive experiments demonstrate that AVES-DPO surpasses existing baselines in hallucination mitigation while requiring only 5.2k samples.

Large Language Models (LLMs) show promise in lyric-to-melody generation, but models trained with Supervised Fine-Tuning (SFT) often produce musically implausible melodies with issues like poor rhythm and unsuitable vocal ranges, a phenomenon we term "constraint violation". To address this, we propose a novel alignment framework that instills musical knowledge without human annotation. We define rule-based musical constraints to automatically generate a preference dataset from an SFT model's outputs. The model is then aligned through a sequential process, first using Direct Preference Optimization (DPO) on paired preference data, followed by Kahneman-Tversky Optimization (KTO) on unpaired negative samples. Experimental results demonstrate that our aligned model substantially reduces rule violations and outperforms strong baselines in both objective and subjective evaluations, generating melodies with substantially improved musicality and coherence. An interactive demo with audio comparisons is available at https://arain233.github.io/AligningMelody-demo.

Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-turn personalized interaction trajectories that closely mirror real-world preference expression and noise patterns. We further propose Personalized Prompt Optimization (PPOpt), a scalable and model-agnostic framework that optimizes user prompts based on interaction histories without modifying the deployed LLM. PPOpt adopts a reason-then-optimize paradigm that infers an explicit user profile and conditions prompt rewriting on the user profile to avoid reward hacking. Our training procedure for PPOpt integrates a cold-start supervised prior with outcome-driven multi-objective reinforcement learning. We present extensive experiments to demonstrate consistent improvements over state-of-the-art baselines in terms of task performance, personalization quality, and robustness to noisy as well as to sparse preference signals.

Data valuation is a natural framework for understanding which preference datasets matter most when aligning a Large Language Model (LLM) using multiple sources. The standard game-theoretic approach assigns each dataset a contribution score via the Shapley value. In practice, however, Shapley-based valuation is computationally prohibitive because it requires fine-tuning a separate model for every possible coalition of preference datasets, i.e., an exponential number of alignments. We address this challenge for a broad family of preference-optimization objectives, including DPO and IPO, that learn directly from log-policy ratios with respect to a reference policy. We introduce Sequential Preference Optimization, an offline procedure that applies existing preference optimization methods sequentially, source by source, updating the current policy after each dataset. Under exact optimization, this procedure yields an additive composition rule in reward space and an equivalent arithmetic composition rule in policy space. This observation enables an efficient approximation of the Shapley value: we train one model per preference dataset and reconstruct coalition policies at inference time from the singleton models, reducing the required alignments from exponential to linear in the number of sources. Leveraging this property, we compute Shapley values for several real-world preference datasets and reveal how each source drives model alignment.

As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model's post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model's values, and subsequent preference optimization rarely re-aligns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.