In multi-hop question answering (MHQA) tasks, Chain of Thought (CoT) improves the quality of generation by guiding large language models (LLMs) through multi-step reasoning, and Knowledge Graphs (KGs) reduce hallucinations via semantic matching. Outcome Reward Models (ORMs) provide feedback after generating the final answers but fail to evaluate the process for multi-step reasoning. Traditional Process Reward Models (PRMs) evaluate the reasoning process but require costly human annotations or rollout generation. While implicit PRM is trained only with outcome signals and derives step rewards through reward parameterization without explicit annotations, it is more suitable for multi-step reasoning in MHQA tasks. However, existing implicit PRM has only been explored for plain text scenarios. When adapting to MHQA tasks, it cannot handle the graph structure constraints in KGs and capture the potential inconsistency between CoT and KG paths. To address these limitations, we propose the DPRM (Dual Implicit Process Reward Model). It trains two implicit PRMs for CoT and KG reasoning in MHQA tasks. Both PRMs, namely KG-PRM and CoT-PRM, derive step-level rewards from outcome signals via reward parameterization without additional explicit annotations. Among them, KG-PRM uses preference pairs to learn structural constraints from KGs. DPRM further introduces a consistency constraint between CoT and KG reasoning steps, making the two PRMs mutually verify and collaboratively optimize the reasoning paths. We also provide a theoretical demonstration of the derivation of process rewards. Experimental results show that our method outperforms 13 baselines on multiple datasets with up to 16.6% improvement on Hit@1.

Recent generative models have significantly advanced speech restoration tasks, yet their training objectives often misalign with human perceptual preferences, resulting in suboptimal quality. While post-training alignment has proven effective in other generative domains like text and image generation, its application to generative speech restoration remains largely under-explored. This work investigates the challenges of applying preference-based post-training to this task, focusing on how to define a robust preference signal and curate high-quality data to avoid reward hacking. To address these challenges, we propose a multi-metric preference alignment strategy. We construct a new dataset, GenSR-Pref, comprising 80K preference pairs, where each chosen sample is unanimously favored by a complementary suite of metrics covering perceptual quality, signal fidelity, content consistency, and timbre preservation. This principled approach ensures a holistic preference signal. Applying Direct Preference Optimization (DPO) with our dataset, we observe consistent and significant performance gains across three diverse generative paradigms: autoregressive models (AR), masked generative models (MGM), and flow-matching models (FM) on various restoration benchmarks, in both objective and subjective evaluations. Ablation studies confirm the superiority of our multi-metric strategy over single-metric approaches in mitigating reward hacking. Furthermore, we demonstrate that our aligned models can serve as powerful ''data annotators'', generating high-quality pseudo-labels to serve as a supervision signal for traditional discriminative models in data-scarce scenarios like singing voice restoration. Demo Page:https://gensr-pref.github.io

Aligning large language models (LLMs) is a central objective of post-training, often achieved through reward modeling and reinforcement learning methods. Among these, direct preference optimization (DPO) has emerged as a widely adopted technique that fine-tunes LLMs on preferred completions over less favorable ones. While most frontier LLMs do not disclose their curated preference pairs, the broader LLM community has released several open-source DPO datasets, including TuluDPO, ORPO, UltraFeedback, HelpSteer, and Code-Preference-Pairs. However, systematic comparisons remain scarce, largely due to the high computational cost and the lack of rich quality annotations, making it difficult to understand how preferences were selected, which task types they span, and how well they reflect human judgment on a per-sample level. In this work, we present the first comprehensive, data-centric analysis of popular open-source DPO corpora. We leverage the Magpie framework to annotate each sample for task category, input quality, and preference reward, a reward-model-based signal that validates the preference order without relying on human annotations. This enables a scalable, fine-grained inspection of preference quality across datasets, revealing structural and qualitative discrepancies in reward margins. Building on these insights, we systematically curate a new DPO mixture, UltraMix, that draws selectively from all five corpora while removing noisy or redundant samples. UltraMix is 30% smaller than the best-performing individual dataset yet exceeds its performance across key benchmarks. We publicly release all annotations, metadata, and our curated mixture to facilitate future research in data-centric preference optimization.

We propose patching for large language models (LLM) like software versions, a lightweight and modular approach for addressing safety vulnerability. While vendors release improved LLM versions, major releases are costly, infrequent, and difficult to tailor to customer needs, leaving released models with known safety gaps. Unlike full-model fine-tuning or major version updates, our method enables rapid remediation by prepending a compact, learnable prefix to an existing model. This "patch" introduces only 0.003% additional parameters, yet reliably steers model behavior toward that of a safer reference model. Across three critical domains--toxicity mitigation, bias reduction, and harmfulness refusal--policy patches achieve safety improvements comparable to next-generation safety-aligned models while preserving fluency. Our results demonstrate that LLMs can be "patched" much like software, offering vendors and practitioners a practical mechanism for distributing scalable, efficient, and composable safety updates between major model releases. Large language models (LLMs) have achieved remarkable advances in reasoning, generation, and multilingual capabilities (Brown et al., 2020; Wei et al., 2022; Conneau & Lample, 2019). Despite their impressive capabilities, they continue to exhibit serious safety concerns, such as the generation of toxic language (Gehman et al., 2020), biased associations that reinforce stereotypes (Dong et al., 2024), and the production of harmful or dangerous content (Mazeika et al., 2024). Addressing these risks is crucial to the broader challenge of alignment, where models are refined to better align with human values and expectations. Conventional approaches to improving safety rely on alignment techniques such as Reinforcement Learning from Human Feedback (RLHF) (Christiano et al., 2017; Bai et al., 2022; Ouyang et al., 2022), preference-based fine-tuning (Rafailov et al., 2023), domain-specific supervised fine-tuning (Li et al., 2024), etc. While these methods have proven effective, they require substantial computational resources, large-scale data curation, and careful model retraining. In practice, model providers (vendors) often release major updates to their models (major versions) on a fixed schedule, typically once or twice a year.

Preference optimization is a critical post-training technique used to align large language models (LLMs) with human preferences, typically by fine-tuning on ranked response pairs. While methods like Direct Preference Optimization (DPO) have proven effective in English, they often fail to generalize robustly to multilingual settings. We propose a simple yet effective alternative, Confidence-Aware Preference Optimization (CAPO), which replaces DPO's fixed treatment of preference pairs with a dynamic loss scaling mechanism based on a relative reward. By modulating the learning signal according to the confidence in each preference pair, CAPO enhances robustness to noisy or low-margin comparisons, typically encountered in multilingual text. Empirically, CAPO outperforms existing preference optimization baselines by at least 16% in reward accuracy, and improves alignment by widening the gap between preferred and dispreferred responses across languages.

Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters.

In high-stakes scenarios-such as self-harm, legal, or medical queries-LLMs must be both trustworthy and helpful. However, these goals often conflict. We propose priority alignment, a new alignment paradigm that enforces a strict "trustworthy-before-helpful" ordering: optimization of helpfulness is conditioned on first meeting trustworthy thresholds (e.g., harmlessness or honesty). To realize this, we introduce Self-Priority Alignment (SPA)-a fully unsupervised framework that generates diverse responses, self-evaluates them and refines them by the model itself, and applies dual-criterion denoising to remove inconsistency and control variance. From this, SPA constructs lexicographically ordered preference pairs and fine-tunes the model using an uncertainty-weighted alignment loss that emphasizes high-confidence, high-gap decisions. Experiments across multiple benchmarks show that SPA improves helpfulness without compromising safety, outperforming strong baselines while preserving general capabilities. Our results demonstrate that SPA provides a scalable and interpretable alignment strategy for critical LLM applications.

LLMs often require adaptation to domain-specific requirements, a process that can be expensive when relying solely on SFT. We present an empirical study on applying CPO to simulate a post-editing workflow for data-efficient domain adaptation. Our approach synthesizes preference pairs by treating the base model's own raw output as the 'rejected' translation and the human-approved TM entry as the 'chosen' one. This method provides direct feedback on the model's current knowledge, guiding it to align with domain-specific standards. Experiments in English-Brazilian Portuguese and English-Korean show that, by using just 14.7k preference pairs, the model achieves performance close to that of a model trained on 160k+ samples with SFT, demonstrating significant data efficiency. Although we showcase its effectiveness in MT, this application of CPO naturally generalizes to other generative tasks where a model's initial drafts can serve as a contrastive signal against a golden reference.

The alignment of Large Language Models (LLMs) is crucial for ensuring their safety and reliability in practical applications. Direct Preference Optimization (DPO) has emerged as an efficient method that directly optimizes models using preference pairs, significantly reducing resource demands. However, the effectiveness of DPO heavily depends on the data quality, which is frequently compromised by noise. In this work, we propose $γ$-PO, a dynamic target margin preference optimization algorithm that adjust reward margins at the pairwise level. By introducing instance-specific margin calibration, $γ$-PO strategically prioritizes high-confidence pairs (those demonstrating higher reward margins) while suppressing potential noise from ambiguous pairs. Moreover, $γ$-PO is a plug-and-play method, compatible with variants of DPO that rely on reward margin between preference pairs. Across benchmarks such as AlpacaEval2 and Arena-Hard, $γ$-PO achieves an average 4.4\% improvement over other baselines, setting new benchmarks for state-of-the-art performance. Additionally, $γ$-PO requires minimal code changes and has a negligible impact on training efficiency, making it a robust solution for enhancing LLMs alignment. Our codes are available at \href{https://github.com/sunjie279/gammaPO}{https://github.com/sunjie279/gammaPO}.

Humans leverage shared conversational context to become increasingly successful and efficient at communicating over time. One manifestation of this is the formation of ad hoc linguistic conventions, which allow people to coordinate on short, less costly utterances that are understood using shared conversational context. We present a method to train large multimodal models to form conventions, enabling efficient communication. Our approach uses simulated reference games between models, and requires no additional human-produced data. In repeated reference games involving photographs and tangram images, our method enables models to communicate efficiently with people: reducing the message length by up to 41% while increasing success by 15% over the course of the interaction. Human listeners respond faster when interacting with our model that forms conventions. We also show that training based on success or cost alone is insufficient - both are necessary to elicit convention formation.