RM-R1: Reward Modeling as Reasoning

Reward modeling is essential for aligning large language models (LLMs) with human preferences, especially through reinforcement learning from human feedback (RLHF). To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. However, existing RMs either produce opaque scalar scores or directly generate the prediction of a preferred answer, making them struggle to integrate natural language critiques, thus lacking interpretability. Inspired by recent advances of long chain-of-thought (CoT) on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning capabilities into reward modeling significantly enhances RM's interpretability and performance. In this work, we introduce a new class of generative reward models -- Reasoning Reward Models (ReasRMs) -- which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. The training consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. RM-R1 improves LLM rollouts by self-generating reasoning traces or chat-specific rubrics and evaluating candidate responses against them. Empirically, our models achieve state-of-the-art or near state-of-the-art performance of generative RMs across multiple comprehensive reward model benchmarks, outperforming much larger open-weight models (e.g., Llama3.1-405B) and proprietary ones (e.g., GPT-4o) by up to 13.8%. Beyond final performance, we perform thorough empirical analysis to understand the key ingredients of successful ReasRM training. To facilitate future research, we release six ReasRM models along with code and data at https://github.com/RM-R1-UIUC/RM-R1.

Leveraging Pretrained Diffusion Models for Zero-Shot Part Assembly

3D part assembly aims to understand part relationships and predict their 6-DoF poses to construct realistic 3D shapes, addressing the growing demand for autonomous assembly, which is crucial for robots. Existing methods mainly estimate the transformation of each part by training neural networks under supervision, which requires a substantial quantity of manually labeled data. However, the high cost of data collection and the immense variability of real-world shapes and parts make traditional methods impractical for large-scale applications. In this paper, we propose first a zero-shot part assembly method that utilizes pre-trained point cloud diffusion models as discriminators in the assembly process, guiding the manipulation of parts to form realistic shapes. Specifically, we theoretically demonstrate that utilizing a diffusion model for zero-shot part assembly can be transformed into an Iterative Closest Point (ICP) process. Then, we propose a novel pushing-away strategy to address the overlap parts, thereby further enhancing the robustness of the method. To verify our work, we conduct extensive experiments and quantitative comparisons to several strong baseline methods, demonstrating the effectiveness of the proposed approach, which even surpasses the supervised learning method. The code has been released on https://github.com/Ruiyuan-Zhang/Zero-Shot-Assembly.

CAE-DFKD: Bridging the Transferability Gap in Data-Free Knowledge Distillation

Data-Free Knowledge Distillation (DFKD) enables the knowledge transfer from the given pre-trained teacher network to the target student model without access to the real training data. Existing DFKD methods focus primarily on improving image recognition performance on associated datasets, often neglecting the crucial aspect of the transferability of learned representations. In this paper, we propose Category-Aware Embedding Data-Free Knowledge Distillation (CAE-DFKD), which addresses at the embedding level the limitations of previous rely on image-level methods to improve model generalization but fail when directly applied to DFKD. The superiority and flexibility of CAE-DFKD are extensively evaluated, including: \textit{\textbf{i.)}} Significant efficiency advantages resulting from altering the generator training paradigm; \textit{\textbf{ii.)}} Competitive performance with existing DFKD state-of-the-art methods on image recognition tasks; \textit{\textbf{iii.)}} Remarkable transferability of data-free learned representations demonstrated in downstream tasks.

Nexus-Gen: A Unified Model for Image Understanding, Generation, and Editing

Unified multimodal large language models (MLLMs) aim to integrate multimodal understanding and generation abilities through a single framework. Despite their versatility, existing open-source unified models exhibit performance gaps against domain-specific architectures. To bridge this gap, we present Nexus-Gen, a unified model that synergizes the language reasoning capabilities of LLMs with the image synthesis power of diffusion models. To align the embedding space of the LLM and diffusion model, we conduct a dual-phase alignment training process. (1) The autoregressive LLM learns to predict image embeddings conditioned on multimodal inputs, while (2) the vision decoder is trained to reconstruct high-fidelity images from these embeddings. During training the LLM, we identified a critical discrepancy between the autoregressive paradigm's training and inference phases, where error accumulation in continuous embedding space severely degrades generation quality. To avoid this issue, we introduce a prefilled autoregression strategy that prefills input sequence with position-embedded special tokens instead of continuous embeddings. Through dual-phase training, Nexus-Gen has developed the integrated capability to comprehensively address the image understanding, generation and editing tasks. All models, datasets, and codes are published at https://github.com/modelscope/Nexus-Gen.git to facilitate further advancements across the field.

ISDrama: Immersive Spatial Drama Generation through Multimodal Prompting

Multimodal immersive spatial drama generation focuses on creating continuous multi-speaker binaural speech with dramatic prosody based on multimodal prompts, with potential applications in AR, VR, and others. This task requires simultaneous modeling of spatial information and dramatic prosody based on multimodal inputs, with high data collection costs. To the best of our knowledge, our work is the first attempt to address these challenges. We construct MRSDrama, the first multimodal recorded spatial drama dataset, containing binaural drama audios, scripts, videos, geometric poses, and textual prompts. Then, we propose ISDrama, the first immersive spatial drama generation model through multimodal prompting. ISDrama comprises these primary components: 1) Multimodal Pose Encoder, based on contrastive learning, considering the Doppler effect caused by moving speakers to extract unified pose information from multimodal prompts. 2) Immersive Drama Transformer, a flow-based mamba-transformer model that generates high-quality drama, incorporating Drama-MOE to select proper experts for enhanced prosody and pose control. We also design a context-consistent classifier-free guidance strategy to coherently generate complete drama. Experimental results show that ISDrama outperforms baseline models on objective and subjective metrics. The demos and dataset are available at https://aaronz345.github.io/ISDramaDemo.

Versatile Framework for Song Generation with Prompt-based Control

Song generation focuses on producing controllable high-quality songs based on various prompts. However, existing methods struggle to generate vocals and accompaniments with prompt-based control and proper alignment. Additionally, they fall short in supporting various tasks. To address these challenges, we introduce VersBand, a multi-task song generation framework for synthesizing high-quality, aligned songs with prompt-based control. VersBand comprises these primary models: 1) VocalBand, a decoupled model, leverages the flow-matching method for generating singing styles, pitches, and mel-spectrograms, allowing fast, high-quality vocal generation with style control. 2) AccompBand, a flow-based transformer model, incorporates the Band-MOE, selecting suitable experts for enhanced quality, alignment, and control. This model allows for generating controllable, high-quality accompaniments aligned with vocals. 3) Two generation models, LyricBand for lyrics and MelodyBand for melodies, contribute to the comprehensive multi-task song generation system, allowing for extensive control based on multiple prompts. Experimental results demonstrate that VersBand performs better over baseline models across multiple song generation tasks using objective and subjective metrics. Audio samples are available at https://aaronz345.github.io/VersBandDemo.

AdaQual-Diff: Diffusion-Based Image Restoration via Adaptive Quality Prompting

Restoring images afflicted by complex real-world degradations remains challenging, as conventional methods often fail to adapt to the unique mixture and severity of artifacts present. This stems from a reliance on indirect cues which poorly capture the true perceptual quality deficit. To address this fundamental limitation, we introduce AdaQual-Diff, a diffusion-based framework that integrates perceptual quality assessment directly into the generative restoration process. Our approach establishes a mathematical relationship between regional quality scores from DeQAScore and optimal guidance complexity, implemented through an Adaptive Quality Prompting mechanism. This mechanism systematically modulates prompt structure according to measured degradation severity: regions with lower perceptual quality receive computationally intensive, structurally complex prompts with precise restoration directives, while higher quality regions receive minimal prompts focused on preservation rather than intervention. The technical core of our method lies in the dynamic allocation of computational resources proportional to degradation severity, creating a spatially-varying guidance field that directs the diffusion process with mathematical precision. By combining this quality-guided approach with content-specific conditioning, our framework achieves fine-grained control over regional restoration intensity without requiring additional parameters or inference iterations. Experimental results demonstrate that AdaQual-Diff achieves visually superior restorations across diverse synthetic and real-world datasets.

CSMF: Cascaded Selective Mask Fine-Tuning for Multi-Objective Embedding-Based Retrieval

Multi-objective embedding-based retrieval (EBR) has become increasingly critical due to the growing complexity of user behaviors and commercial objectives. While traditional approaches often suffer from data sparsity and limited information sharing between objectives, recent methods utilizing a shared network alongside dedicated sub-networks for each objective partially address these limitations. However, such methods significantly increase the model parameters, leading to an increased retrieval latency and a limited ability to model causal relationships between objectives. To address these challenges, we propose the Cascaded Selective Mask Fine-Tuning (CSMF), a novel method that enhances both retrieval efficiency and serving performance for multi-objective EBR. The CSMF framework selectively masks model parameters to free up independent learning space for each objective, leveraging the cascading relationships between objectives during the sequential fine-tuning. Without increasing network parameters or online retrieval overhead, CSMF computes a linearly weighted fusion score for multiple objective probabilities while supporting flexible adjustment of each objective's weight across various recommendation scenarios. Experimental results on real-world datasets demonstrate the superior performance of CSMF, and online experiments validate its significant practical value.

Enhanced Battery Capacity Estimation in Data-Limited Scenarios through Swarm Learning

Data-driven methods have shown potential in electric-vehicle battery management tasks such as capacity estimation, but their deployment is bottlenecked by poor performance in data-limited scenarios. Sharing battery data among algorithm developers can enable accurate and generalizable data-driven models. However, an effective battery management framework that simultaneously ensures data privacy and fault tolerance is still lacking. This paper proposes a swarm battery management system that unites a decentralized swarm learning (SL) framework and credibility weight-based model merging mechanism to enhance battery capacity estimation in data-limited scenarios while ensuring data privacy and security. The effectiveness of the SL framework is validated on a dataset comprising 66 commercial LiNiCoAlO2 cells cycled under various operating conditions. Specifically, the capacity estimation performance is validated in four cases, including data-balanced, volume-biased, feature-biased, and quality-biased scenarios. Our results show that SL can enhance the estimation accuracy in all data-limited cases and achieve a similar level of accuracy with central learning where large amounts of data are available.

FedEPA: Enhancing Personalization and Modality Alignment in Multimodal Federated Learning

Federated Learning (FL) enables decentralized model training across multiple parties while preserving privacy. However, most FL systems assume clients hold only unimodal data, limiting their real-world applicability, as institutions often possess multimodal data. Moreover, the lack of labeled data further constrains the performance of most FL methods. In this work, we propose FedEPA, a novel FL framework for multimodal learning. FedEPA employs a personalized local model aggregation strategy that leverages labeled data on clients to learn personalized aggregation weights, thereby alleviating the impact of data heterogeneity. We also propose an unsupervised modality alignment strategy that works effectively with limited labeled data. Specifically, we decompose multimodal features into aligned features and context features. We then employ contrastive learning to align the aligned features across modalities, ensure the independence between aligned features and context features within each modality, and promote the diversity of context features. A multimodal feature fusion strategy is introduced to obtain a joint embedding. The experimental results show that FedEPA significantly outperforms existing FL methods in multimodal classification tasks under limited labeled data conditions.