Object removal aims to eliminate target objects specified by a mask while preserving visual consistency with the surrounding regions. Existing methods typically rely on contextual information from surrounding regions. However, in dense scenes where the surrounding regions contain instances visually similar to the removal target, such reliance often leads to semantic interference, resulting in incomplete removal. This problem arises from erroneous information propagation in the attention space, where masked queries tend to align with such instances due to global similarity matching in self-attention. To address this challenge, we propose a Diffusion-based Object Removal framework for dense Scenes, dubbed DORS, built upon a Dynamic Attention Routing mechanism comprising two complementary components: Instance-Filtered Attention (IFA), which suppresses misleading semantic information from similar instances through dynamically constructed mask-guided attention constraints, and Context-Guided Routing (CGR), which dynamically routes complementary scene information to maintain visual consistency. We further introduce DOR-Bench, a benchmark tailored for object removal in dense scenes. Extensive experiments demonstrate that DORS outperforms state-of-the-art methods, particularly in reducing incomplete removal and duplicate artifacts. The code will be available at https://github.com/httang1224/DORS.

Artificial intelligence has advanced scientific discovery, but most AI4Science systems remain fragmented tools that rely on humans to coordinate problem formulation, literature grounding, model use, simulation, validation, and knowledge reuse. This paper presents SCION (Scientific Collaborative Innovation with Agentic Organizational Nexus), an agentic scientific operating system that acts as an organizational nexus. Through a Science Agent serving as a Meta-Harness, SCION connects scientific tasks, tools, agents, artifacts, and memory, transforming research into an executable, auditable, and reusable operational process. At its core is the Research Execution Plan (REP), which compiles high-level scientific intent into staged objectives, dependencies, verification checkpoints, tool requirements, expected artifacts, and fallback conditions. SCION further integrates hierarchical multi-agent execution, profile-driven specialization, selective context construction, governed delegation, and layered epistemic memory to support long-horizon scientific work. We formulate discovery under SCION as Target-conditioned Inverse Search and extend it to hidden-target settings through batch active search under finite experimental budgets. Applications in materials analysis, molecule design, and protein or antibody screening, together with experiments on scientific reading, idea generation, molecule generation, and antibody screening, show that SCION outperforms existing autonomous research-agent baselines, especially in decomposition, verification, refinement, and memory reuse. Overall, SCION shifts AI from isolated tools toward a coordinated operational layer for traceable and reusable scientific innovation.

Multi-object editing aims to modify multiple objects or regions in complex scenes while preserving structural coherence. This task faces significant challenges in scenarios involving overlapping or interacting objects: (1) Inaccurate localization of target objects due to attention misalignment, leading to incomplete or misplaced edits; (2) Attribute-object mismatch, where color or texture changes fail to align with intended regions due to cross-attention leakage, creating semantic conflicts (e.g., color bleeding into non-target areas). Existing methods struggle with these challenges: approaches relying on global cross-attention mechanisms suffer from attention dilution and spatial interference between objects, while mask-based methods fail to bind attributes to geometrically accurate regions due to feature entanglement in multi-object scenarios. To address these limitations, we propose a training-free, inference-stage optimization approach that enables precise localized image manipulation in complex multi-object scenes, named MDE-Edit. MDE-Edit optimizes the noise latent feature in diffusion models via two key losses: Object Alignment Loss (OAL) aligns multi-layer cross-attention with segmentation masks for precise object positioning, and Color Consistency Loss (CCL) amplifies target attribute attention within masks while suppressing leakage to adjacent regions. This dual-loss design ensures localized and coherent multi-object edits. Extensive experiments demonstrate that MDE-Edit outperforms state-of-the-art methods in editing accuracy and visual quality, offering a robust solution for complex multi-object image manipulation tasks. Our code can be accessed at https://github.com/RainbowZhuu/MDE-Edit .

MOSS-Audio is a unified audio-language model for speech, environmental sound, and music understanding, supporting audio captioning, time-aware question answering, timestamped transcription, and audio-grounded reasoning. MOSS-Audio couples a dedicated audio encoder with a modality adapter and a large language model: the encoder produces 12.5 Hz temporal representations, the adapter projects them into the decoder space, and the decoder generates autoregressive text outputs. Two design choices are central to the system: DeepStack cross-layer feature injection, which exposes the decoder to acoustic information from multiple encoder depths, and time markers, which provide explicit temporal cues by inserting timestamp markers into the audio-token stream. At the data level, we design an event-preserving audio annotation pipeline that segments raw audio at coherent event boundaries, applies branch-specific annotation to speech, music, and general audio, and merges the results into unified captions for pretraining. The intermediate branch-specific captions are further retained to support the construction of task-oriented SFT data. The model is pretrained on large-scale audio-language data, with time-aware objectives incorporated to support temporal grounding, and then undergoes multi-stage post-training to enhance instruction following and audio-grounded reasoning. We release 4B and 8B variants in both Instruct and Thinking configurations. MOSS-Audio achieves strong performance across general audio understanding, speech captioning, ASR, and timestamped ASR, positioning it as a promising understanding foundation for future voice agents.

Although pretrained text-to-image (T2I) generation models can produce high-quality images, they often fail to faithfully reflect the semantic intent of complex prompts due to stochastic noise and inherent model limitations. This issue frequently manifests as the model overlooking specific objects or failing to correctly bind attributes to their corresponding entities, a challenge referred to as semantic alignment. Unlike existing approaches that rely on computationally expensive fine-tuning or labor-intensive layout priors, we propose STEDiff, a training-free method designed to enhance semantic representations directly within the text-embedding space. Specifically, we introduce a method that primarily leverages the [EOT] token to strengthen the relevant semantics of sub-sentences and then replaces the corresponding tokens in the original prompt. Furthermore, a novel semantic enhancement loss is incorporated to enforce spatial constraints, ensuring that the semantics of each entity are precisely mapped to their respective image regions. Extensive quantitative and qualitative evaluations on the T2I-CompBench demonstrate that our method notably improves semantic consistency and generation integrity in complex scenarios.

Personalized recommendation systems often suffer from various biases in data-driven paradigms, which can significantly degrade their fairness and generalization. Existing debiasing approaches typically rely on strong assumptions about bias types or purely statistical corrections, limiting their adaptability to unknown and changeable bias patterns, which fail to capture the complexities of human behavior and decision-making. To address this challenge, we propose a novel SVD-based dual adaptive debiasing learning (SDADL) framework. In SDADL, we distinguish between target samples (TS), which are observed positive interactions in training, and flaw samples (FS), which are highly ranked but potentially spurious positives. During training, node embeddings are decomposed via singular value decomposition (SVD) and further refined through dual adaptive learning pathways dedicated to TS and FS, respectively. Notably, the model is not explicitly constrained by the loss on the original embeddings; instead, it adaptively balances dual path learning signals under weak supervision, thereby discovering and mitigating hidden biases in a self-guided manner. Experiments demonstrate that SDADL consistently enhances base GCN models in a plug-and-play form and is effective in both supervised and self-supervised learning scenarios.

This paper presents ChipDiff, a novel staged diffusion framework specifically designed for Chinese ink-wash style transfer. While conventional generative models often struggle to capture the intricate monochromatic tonal layering and “bone-work” (structural brushstrokes) of traditional art, we propose a bifurcated sampling strategy guided by tailored loss gradients to reconcile structural preservation with stylistic refinement. Inspired by the traditional artistic workflow of “brushing-then-inking”, our approach decouples the denoising trajectory into two functional phases. In the first stage, we leverage a combination of content-structure, HED-edge, and semantic-divergence losses to anchor the global layout and latent skeleton. The second stage further refines the synthesis by injecting fine-grained ink-wash textures and anisotropic bleeding effects through multi-scale stylistic representation losses. Unlike prior single-stage methods, ChipDiff provides a mathematically grounded and artistically intuitive mechanism for controlled pattern generation. Extensive experiments, supported by our curated Traditional Chinese Ink-Wash Painting (TCWP) dataset and a comprehensive perception-aligned evaluation protocol, demonstrate that ChipDiff achieves superior performance in stylistic fidelity and structural integrity over state-of-the-art baselines. Project code is available at https://github.com/hengliusky/ChipDiff/.

Rare concept generation focuses on synthesizing customized images conditioned on text prompts that describe objects with unusual attributes. Previous works failed to align the generated images with rare concepts, resulting in incorrect attribute rendering or inconsistent composition of concepts. Such failures, as we observed, stem from the inherent common knowledge bias in the training stage of diffusion models, where objects are strongly associated with their common attributes, making it difficult to break these associations when generating rare concepts. To address such challenges, in this paper, we propose a novel Counterfactual Inference-based Diffusion approach, dubbed CI-Diff. CI-Diff blocks the interference of the model's inherent common knowledge bias and utilizes the Natural Direct Effect to capture the independent influence of the text prompt of rare concepts on image generation so that decoupling the unusual attributes from the rare concepts. To this end, we reformulate the classifier-free guidance mechanism to highlight the atypical attributes. To the best of our knowledge, we are the first to introduce causal inference into the rare concept generation task. Extensive experiments on the RareBench benchmark validate the superiority of CI-Diff over state-of-the-art diffusion models. Our code can be accessed from https://github.com/200204jzy/CI-Diff.

Speaker-Attributed, Time-Stamped Transcription (SATS) aims to transcribe what is said and to precisely determine the timing of each speaker, which is particularly valuable for meeting transcription. Existing SATS systems rarely adopt an end-to-end formulation and are further constrained by limited context windows, weak long-range speaker memory, and the inability to output timestamps. To address these limitations, we present MOSS Transcribe Diarize, a unified multimodal large language model that jointly performs Speaker-Attributed, Time-Stamped Transcription in an end-to-end paradigm. Trained on extensive real wild data and equipped with a 128k context window for up to 90-minute inputs, MOSS Transcribe Diarize scales well and generalizes robustly. Across comprehensive evaluations, it outperforms state-of-the-art commercial systems on multiple public and in-house benchmarks.

Poisoning defense in federated learning (FL) is challenging because the server must make security decisions with limited observability, often relying only on uploaded client updates. The challenge becomes even more pronounced in heterogeneous FL: strong non-identically and independently distributed (Non-IID) data and training diversity can make benign and malicious updates overlap in the observable update space, while adaptive attackers further undermine the reliability of fixed-threshold filtering. To tackle these two coupled challenges—insufficient discriminability and fragile decision stability—we propose a spectrum-feature-aware framework for malicious client detection in FL. Our method maps client updates into the spectral domain and extracts structural energy-distribution features to better separate benign behavior from malicious manipulation. It then combines a Mahalanobis-distance anomaly score with dynamic threshold, allowing filtering decisions to adapt to benign diversity over training rounds. By filtering suspicious updates before aggregation, the framework reduces the cumulative impact of poisoning across rounds. We evaluate the proposed approach on multiple datasets under representative poisoning settings and report results from several perspectives, including detection quality and robustness. Experiments show that our method improves detection effectiveness and attack resistance while preserving task utility, and it exhibits more stable defense behavior in heterogeneous FL scenarios.

Image inpainting has earned substantial progress, owing to the encoder-and-decoder pipeline, which is benefited from the Convolutional Neural Networks (CNNs) with convolutional downsampling to inpaint the masked regions semantically from the known regions within the encoder, coupled with an upsampling process from the decoder for final inpainting output. Recent studies intuitively identify the high-frequency structure and low-frequency texture to be extracted by CNNs from the encoder, and subsequently for a desirable upsampling recovery. However, the existing arts inevitably overlook the information loss for both structure and texture feature maps during the convolutional downsampling process, hence suffer from a non-ideal upsampling output. In this paper, we systematically answer whether and how the structure and texture feature map can mutually help to alleviate the information loss during the convolutional downsampling. Given the structure and texture feature maps, we adopt the statistical normalization and denormalization strategy for the reconstruction guidance during the convolutional downsampling process. The extensive experimental results validate its advantages to the state-of-the-arts over the images from low-to-high resolutions including 256*256 and 512*512, especially holds by substituting all the encoders by ours. Our code is available at https://github.com/htyjers/ConvInpaint-TSGL

Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretrained encoders, semantic distillation, or heterogeneous CNN-based architectures. These designs introduce fixed inductive biases that limit reconstruction fidelity and hinder effective scaling. In this paper, we argue that discrete audio tokenization should be learned fully end-to-end using a homogeneous and scalable architecture. To this end, we first propose CAT (Causal Audio Tokenizer with Transformer), a purely Transformer-based architecture that jointly optimizes the encoder, quantizer, and decoder from scratch for high-fidelity reconstruction. Building on the CAT architecture, we develop MOSS-Audio-Tokenizer, a large-scale audio tokenizer featuring 1.6 billion parameters, pre-trained on 3 million hours of diverse, general audio data. We show that this simple, fully end-to-end approach built from homogeneous, causal Transformer blocks scales gracefully and supports high-fidelity reconstruction across diverse audio domains. Across speech, sound, and music, MOSS-Audio-Tokenizer consistently outperforms prior codecs over a wide range of bitrates, while exhibiting predictable improvements with increased scale. Notably, leveraging the discrete tokens from our model, we develop the first purely autoregressive TTS model that surpasses prior non-autoregressive and cascaded systems. Furthermore, MOSS-Audio-Tokenizer enables competitive ASR performance without auxiliary encoders. Our findings position the CAT architecture as a unified, scalable interface for the next generation of native audio foundation models.

Spoken dialogue systems often rely on cascaded pipelines that transcribe, process, and resynthesize speech. While effective, this design discards paralinguistic cues and limits expressivity. Recent end-to-end methods reduce latency and better preserve these cues, yet still rely on text intermediates, creating a fundamental bottleneck. We present a true speech-to-speech large language model that directly understands and generates speech without relying on text guidance. Our approach combines a modality-based layer-splitting architecture with a frozen pre-training strategy, preserving the reasoning and knowledge of pretrained text LLMs while adding native speech capabilities. Experiments show that our model achieves state-of-the-art results in spoken question answering and delivers comparable speech-to-speech performance relative to existing text-guided systems, while still maintaining competitive text performance. By narrowing the gap between text-guided and direct speech generation, our work establishes a new paradigm for expressive and efficient end-to-end speech interaction. We will release our code and models to support further research in true speech-to-speech foundation models.

Cross-modal hashing (CMH) maps heterogeneous data such as images and text into a shared Hamming space for efficient retrieval. While supervised CMH achieves strong retrieval performance by leveraging label information, it remains highly vulnerable to noisy annotations. Existing noise-handling methods alleviate this issue but generally lack mechanisms to validate the reliability of processed samples, thereby limiting robustness. To address this gap, we propose Robust Deep Cross-modal Hashing via Dual Consensus Learning (RCHDC), which integrates two complementary modules. Label Refinement via Dual Consensus Learning (LRDCL) refines labels through dual-pathway consensus, while the Robust Contrastive Loss Function (RCL) enhances cross-modal discrimination and facilitates robust prototype training. Extensive experiments on multiple datasets demonstrate that RCHDC consistently improves retrieval accuracy and robustness under different noisy conditions, achieving state-of-the-art performance.

Text-to-image synthesis has made significant progress, benefiting from the strong generative capabilities of diffusion models. However, these models struggle to achieve precise text-to-image alignment within cross-attention maps during the denoising process. Existing works primarily focus on inter-subject-token activations (i.e., cross-attention scores) overlap for different subjects, overlooking the intra-subject-token activations scattering issue for identical subjects. In this paper, we propose an Aggregating-and-Isolating cross-attention approach to diffusion models for Text-to-Image synthesis, dubbed AI-T2I. Technically, to address the scattering issue, we devise an aggregation loss to identify and consolidate the scattered intra-token activations, which implicitly helps mitigate the potential overlap issue. Upon that, an isolation loss is further introduced to push the inter-token activations apart, thus fulfilling precise text-to-image alignment. Extensive experiments on various benchmarks demonstrate the superiority of AI-T2I over the state-of-the-art works for text-to-image synthesis. Furthermore, our AI-T2I exhibits excellent generalization across other tasks, e.g., controllable layout generation and personalized generation.

Deep neural networks are susceptible to adversarial examples while suffering from incorrect predictions via imperceptible perturbations. Transfer-based attacks create adversarial examples for surrogate models and transfer these examples to target models under black-box scenarios. Recent studies reveal that adversarial examples in flat loss landscapes exhibit superior transferability to alleviate overfitting on surrogate models. However, the prior arts overlook the influence of perturbation directions, resulting in limited transferability. In this paper, we propose a novel attack method, named Residual Perturbation Attack (ResPA), relying on the residual gradient as the perturbation direction to guide the adversarial examples toward the flat regions of the loss function. Specifically, ResPA conducts an exponential moving average on the input gradients to obtain the first moment as the reference gradient, which encompasses the direction of historical gradients. Instead of heavily relying on the local flatness that stems from the current gradients as the perturbation direction, ResPA further considers the residual between the current gradient and the reference gradient to capture the changes in the global perturbation direction. The experimental results demonstrate the better transferability of ResPA than the existing typical transfer-based attack methods, while the transferability can be further improved by combining ResPA with the current input transformation methods. The code is available at https://github.com/ZezeTao/ResPA.

Referring Video Object Segmentation (RVOS) aims to segment specific objects in videos based on the provided natural language descriptions. As a new supervised visual learning task, achieving RVOS for a given scene requires a substantial amount of annotated data. However, only minimal annotations are usually available for new scenes in realistic scenarios. Another practical problem is that, apart from a single object, multiple objects of the same category coexist in the same scene. Both of these issues may significantly reduce the performance of existing RVOS methods in handling real-world applications. In this paper, we propose a simple yet effective model to address these issues by incorporating a newly designed cross-modal affinity (CMA) module based on a Transformer architecture. The CMA module facilitates the establishment of multi-modal affinity over a limited number of samples, allowing the rapid acquisition of new semantic information while fostering the model’s adaptability to diverse scenarios. Furthermore, we extend our FS-RVOS approach to multiple objects through a new instance sequence matching module over CMA, which filters out all object trajectories with similarity to language features that exceed a matching threshold, thereby achieving few-shot referring multi-object segmentation (FS-RVMOS). To foster research in this field, we establish a new dataset based on currently available datasets, which covers many scenarios in terms of single-object and multi-object data, hence effectively simulating real-world scenes. Extensive experiments and comparative analyses underscore the exceptional performance of our proposed FS-RVOS and FS-RVMOS methods. Our method consistently outperforms existing related approaches through practical performance evaluations and robustness studies, achieving optimal performance on metrics across diverse benchmark tests.

Music recommender systems play a critical role in music streaming platforms by providing users with music that they are likely to enjoy. Recent studies have shown that user emotions can influence users' preferences for music moods. However, existing emotion-aware music recommender systems (EMRSs) explicitly or implicitly assume that users' actual emotional states expressed through identical emotional words are homogeneous. They also assume that users' music mood preferences are homogeneous under the same emotional state. In this article, we propose four types of heterogeneity that an EMRS should account for: emotion heterogeneity across users, emotion heterogeneity within a user, music mood preference heterogeneity across users, and music mood preference heterogeneity within a user. We further propose a Heterogeneity-aware Deep Bayesian Network (HDBN) to model these assumptions. The HDBN mimics a user's decision process of choosing music with four components: personalized prior user emotion distribution modeling, posterior user emotion distribution modeling, user grouping, and Bayesian neural network-based music mood preference prediction. We constructed two datasets, called EmoMusicLJ and EmoMusicLJ-small, to validate our method. Extensive experiments demonstrate that our method significantly outperforms baseline approaches on metrics of HR, Precision, NDCG, and MRR. Ablation studies and case studies further validate the effectiveness of our HDBN. The source code and datasets are available at https://github.com/jingrk/HDBN.