Multi-task offline safe reinforcement learning (RL) promises to learn a shared optimal safe policy from offline data across multiple tasks. This paradigm provides an effective means for the widespread application of RL in multi-task scenarios with high risk and interaction costs. However, the triple challenges of multi-tasking, safety constraints, and out-of-distribution (OOD) actions pose a significant hurdle for existing methods to ensure safety while maximizing reward returns. In this work, we propose a Conditional Diffusion model with Contextual Prompts (CDCP) to address these challenges. Concretely, we first rethink the requirements and challenges in current multi-task decision-making and control scenarios and establish the objectives of multi-task offline safe RL. Subsequently, we transform the multi-task constrained optimization problem into a conditional generation problem using the diffusion model. Based on this, we design a classifier-free guided cost-constraint strategy to provide flexible cost constraints and eliminate extrapolation errors from OOD actions via supervised learning. Additionally, we introduce a novel contextual prompting method to enhance multi-task representation accuracy and adaptability to unseen tasks. A gradient loss synchronization strategy is also introduced to eliminate gradient interference, improving training stability. Finally, extensive experiments demonstrate that the CDCP algorithm exhibits higher performance and safety in multi-task scenarios than the current state-of-the-art baseline methods. It meets different cost constraints without further training, providing a more flexible cost-constraint solution for the multi-task safe RL.

Existing video avatar models have demonstrated impressive capabilities in scenarios such as talking, public speaking, and singing. However, the majority of these methods exhibit limited alignment with respect to text instructions, particularly when the prompts involve complex elements including large full-body movement, dynamic camera trajectory, background transitions, or human-object interactions. To break out this limitation, we present JoyAvatar, a framework capable of generating long duration avatar videos, featuring two key technical innovations. Firstly, we introduce a twin-teacher enhanced training algorithm that enables the model to transfer inherent text-controllability from the foundation model while simultaneously learning audio-visual synchronization. Secondly, during training, we dynamically modulate the strength of multi-modal conditions (e.g., audio and text) based on the distinct denoising timestep, aiming to mitigate conflicts between the heterogeneous conditioning signals. These two key designs serve to substantially expand the avatar model's capacity to generate natural, temporally coherent full-body motions and dynamic camera movements as well as preserve the basic avatar capabilities, such as accurate lip-sync and identity consistency. GSB evaluation results demonstrate that our JoyStreamer model outperforms the state-of-the-art models such as Omnihuman-1.5 and KlingAvatar 2.0. Moreover, our approach enables complex applications including multi-person dialogues and non-human subjects role-playing. Some video samples are provided on https://joystreamer.github.io/.

Modern automatic speech recognition (ASR) scenarios require both spoken-form transcripts for faithful transcription and readable written-form transcripts with inverse text normalization (ITN). However, these forms are typically produced by cascaded modules, where a spoken-form ASR output is rewritten by a separate ITN component, making written-form ASR-ITN vulnerable to recognition errors and decoupling normalization from acoustic-contextual modeling, especially for semantically dependent numeric expressions. In this paper, we propose Dual-Form ASR (DF-ASR), a framework that extends spoken-form ASR capability to semantics-aware written-form ITN through paired spoken-form and written-form supervision while retaining prompt-level selection between transcript forms. The dual-form supervision is constructed via a large language model (LLM)-driven generate-and-judge workflow, and training is further enhanced by ITN-MWER, a sequence-level objective that assigns higher cost to errors on normalization-sensitive spans. We also introduce a decision-aware REQUIRE-ITN/-ITN protocol to separately measure required normalization and forbidden-span preservation. On manually annotated Chinese subsets from SpeechIO, DF-ASR consistently outperforms open-source ASR-ITN systems, remains competitive with strong closed-source references, and preserves reliable prompt-level control between spoken-form and written-form outputs.

Retargeting human motion to heterogeneous robots is a fundamental challenge in robotics, primarily due to the severe kinematic and dynamic discrepancies between varying embodiments. Existing solutions typically resort to training embodiment-specific models, which scales poorly and fails to exploit shared motion semantics. To address this, we present AdaMorph, a unified neural retargeting framework that enables a single model to adapt human motion to diverse robot morphologies. Our approach treats retargeting as a conditional generation task. We map human motion into a morphology-agnostic latent intent space and utilize a dual-purpose prompting mechanism to condition the generation. Instead of simple input concatenation, we leverage Adaptive Layer Normalization (AdaLN) to dynamically modulate the decoder's feature space based on embodiment constraints. Furthermore, we enforce physical plausibility through a curriculum-based training objective that ensures orientation and trajectory consistency via integration. Experimental results on 12 distinct humanoid robots demonstrate that AdaMorph effectively unifies control across heterogeneous topologies, exhibiting strong zero-shot generalization to unseen complex motions while preserving the dynamic essence of the source behaviors.

Large Reasoning Models (LRMs) demonstrate strong performance on complex tasks but often suffer from excessive verbosity, known as "overthinking." Existing solutions via reinforcement learning (RL) typically penalize generated tokens to promote conciseness. However, these methods encounter two challenges: responses with fewer tokens do not always correspond to fewer reasoning steps, and models may develop hacking behavior in later stages of training by discarding reasoning steps to minimize token usage. In this work, we introduce \textbf{Step Pruner (SP)}, an RL framework that steers LRMs toward more efficient reasoning by favoring compact reasoning steps. Our step-aware reward function prioritizes correctness while imposing penalties for redundant steps, and withholds rewards for incorrect responses to prevent the reinforcement of erroneous reasoning. Moreover, we propose a dynamic stopping mechanism: when the length of any output step exceeds the upper limit, we halt updates to prevent hacking behavior caused by merging steps. Extensive experiments across four reasoning benchmarks demonstrate that SP achieves state-of-the-art accuracy while significantly reducing response length. For instance, on AIME24, SP reduces token usage by \textbf{69.7\%}.

This paper presents a Pronunciation-Aware Contextualized (PAC) framework to address two key challenges in Large Language Model (LLM)-based Automatic Speech Recognition (ASR) systems: effective pronunciation modeling and robust homophone discrimination. Both are essential for raw or long-tail word recognition. The proposed approach adopts a two-stage learning paradigm. First, we introduce a pronunciation-guided context learning method. It employs an interleaved grapheme-phoneme context modeling strategy that incorporates grapheme-only distractors, encouraging the model to leverage phonemic cues for accurate recognition. Then, we propose a pronunciation-discriminative reinforcement learning method with perturbed label sampling to further enhance the model’s ability to distinguish contextualized homophones. Experimental results on the public English Librispeech and Mandarin AISHELL-1 datasets indicate that PAC: (1) reduces relative Word Error Rate (WER) by 30.2% and 53.8% compared to pre-trained LLM-based ASR models, and (2) achieves 31.8% and 60.5% relative reductions in biased WER for long-tail words compared to strong baselines, respectively.

Large speech generation models are evolving from single-speaker, short sentence synthesis to multi-speaker, long conversation geneartion. Current long-form speech generation models are predominately constrained to dyadic, turn-based interactions. To address this, we introduce JoyVoice, a novel anthropomorphic foundation model designed for flexible, boundary-free synthesis of up to eight speakers. Unlike conventional cascaded systems, JoyVoice employs a unified E2E-Transformer-DiT architecture that utilizes autoregressive hidden representations directly for diffusion inputs, enabling holistic end-to-end optimization. We further propose a MM-Tokenizer operating at a low bitrate of 12.5 Hz, which integrates multitask semantic and MMSE losses to effectively model both semantic and acoustic information. Additionally, the model incorporates robust text front-end processing via large-scale data perturbation. Experiments show that JoyVoice achieves state-of-the-art results in multilingual generation (Chinese, English, Japanese, Korean) and zero-shot voice cloning. JoyVoice achieves top-tier results on both the Seed-TTS-Eval Benchmark and multi-speaker long-form conversational voice cloning tasks, demonstrating superior audio quality and generalization. It achieves significant improvements in prosodic continuity for long-form speech, rhythm richness in multi-speaker conversations, paralinguistic naturalness, besides superior intelligibility. We encourage readers to listen to the demo at https://jea-speech.github.io/JoyVoice

Text-to-video (T2V) generation has made tremendous progress in generating complicated scenes based on texts. However, human-object interaction (HOI) often cannot be precisely generated by current T2V models due to the lack of large-scale videos with accurate captions for HOI. To address this issue, we introduce HOIGen-1M, the first large-scale dataset for HOI Generation, consisting of over one million high-quality videos collected from diverse sources. In particular, to guarantee the high quality of videos, we first design an efficient framework to automatically curate HOI videos using the powerful multimodal large language models (MLLMs), and then the videos are further cleaned by human annotators. Moreover, to obtain accurate textual captions for HOI videos, we design a novel video description method based on a Mixture-of-Multimodal-Experts (MoME) strategy that not only generates expressive captions but also eliminates the hallucination by individual MLLM. Further-more, due to the lack of an evaluation framework for gen-erated HOI videos, we propose two new metrics to assess the quality of generated videos in a coarse-to-fine manner. Extensive experiments reveal that current T2V models struggle to generate high-quality HOI videos and confirm that our HOIGen-1M dataset is instrumental for improving HOI video generation.

Dexterous grasping in the real world presents a fundamental and significant challenge for robot learning. The ability to employ affordance-aware poses to grasp objects with diverse geometries and properties in arbitrary scenarios is essential for general-purpose robots. However, existing research predominantly addresses dexterous grasping problems within simulators, which encounter difficulties when applied in real-world environments due to the domain gap between reality and simulation. This limitation hinders their generalizability and practicality in real-world applications. In this paper, we present DexGraspRL, a reinforcement learning (RL) framework that directly trains robots in real-world environments to acquire dexterous grasping skills. Specifically, DexGraspRL consists of two stages: (i) a pretraining stage that pretrains the policy using imitation learning (IL) with a limited set of expert demonstrations; (ii) a fine-tuning stage that refines the policy through direct RL in real-world scenarios. To mitigate the catastrophic forgetting phenomenon arising from the distribution shift between demonstrations and real-world environments, we design a regularization term that balances the exploitation of RL with the preservation of the pretrained policy. Our experiments with real-world tasks demonstrate that DexGraspRL successfully accomplishes diverse dexterous grasping tasks, achieving an average success rate of nearly 92

Multimodal sarcasm detection (MSD) requires predicting the sarcastic sentiment by understanding diverse modalities of data (e.g., text, image). Beyond the surface-level information conveyed in the post data, understanding the underlying deep-level knowledge-such as the background and intent behind the data-is crucial for understanding the sarcastic sentiment. However, previous works have often overlooked this aspect, limiting their potential to achieve superior performance. To tackle this challenge, we propose DeepMSD, a novel framework that generates supplemental deep-level knowledge to enhance the understanding of sarcastic content. Specifically, we first devise a Deep-level Knowledge Extraction Module that leverages large vision-language models to generate deep-level information behind the text-image pairs. Additionally, we devise a Cross-knowledge Graph Reasoning Module to model how humans use prior knowledge to identify sarcastic cues in multimodal posts. This module constructs cross-knowledge graphs that connect deep-level knowledge with surface-level knowledge. As such, it enables a more profound exploration of the cues underlying sarcasm. Experiments on the public MSD dataset demonstrate that our approach significantly surpasses previous state-of-the-art methods.

Recent advancements in scaling up models have significantly improved performance in Automatic Speech Recognition (ASR) tasks. However, training large ASR models from scratch remains costly. To address this issue, we introduce UME, a novel method that efficiently Upcycles pretrained dense ASR checkpoints into larger Mixture-of-Experts (MoE) architectures. Initially, feed-forward networks are converted into MoE layers. By reusing the pretrained weights, we establish a robust foundation for the expanded model, significantly reducing optimization time. Then, layer freezing and expert balancing strategies are employed to continue training the model, further enhancing performance. Experiments on a mixture of 170k-hour Mandarin and English datasets show that UME: 1) surpasses the pretrained baseline by a margin of 11.9% relative error rate reduction while maintaining comparable latency; 2) reduces training time by up to 86.7% and achieves superior accuracy compared to training models of the same size from scratch.

Recent advances in large language models (LLMs) have leveraged explicit Chain-of-Thought (CoT) prompting to improve reasoning accuracy. However, most existing methods primarily compress verbose reasoning outputs. These Long-to-Short transformations aim to improve efficiency, but still rely on explicit reasoning during inference. In this work, we introduce 3TF (Thought-Training and Thought-Free inference), a framework for efficient reasoning that takes a Short-to-Long perspective. We first train a hybrid model that can operate in both reasoning and non-reasoning modes, and then further train it on CoT-annotated data to internalize structured reasoning, while enforcing concise, thought-free outputs at inference time using the no-reasoning mode. Unlike compression-based approaches, 3TF improves the reasoning quality of non-reasoning outputs, enabling models to perform rich internal reasoning implicitly while keeping external outputs short. Empirically, 3TF-trained models obtain large improvements on reasoning benchmarks under thought-free inference, demonstrating that high quality reasoning can be learned and executed implicitly without explicit step-by-step generation.

Document-level event argument extraction is a crucial yet challenging task within the field of information extraction. Current mainstream approaches primarily focus on the information interaction between event triggers and their arguments, facing two limitations: insufficient context interaction and the ignorance of event correlations. Here, we introduce a novel framework named CARG (Contextual Aggregation of Clues and Role-based Latent Guidance), comprising two innovative components: the Contextual Clues Aggregation (CCA) and the Role-based Latent Information Guidance (RLIG) The CCA module leverages the attention weights derived from a pre-trained encoder to adaptively assimilate broader contextual information, while the RLIG module aims to capture the semantic correlations among event roles. We then instantiate the CARG framework into two variants based on two types of mainstream EAE approaches. Notably, our CARG framework introduces less than 1% new parameters yet significantly improves the performance. Comprehensive experiments across the RAMS, WikiEvents, MLEE and ACE 2005 datasets confirm the superiority of CARG, showing significant superiority in terms of both performance and inference speed compared to major benchmarks. Further analyses demonstrate the effectiveness of the proposed modules.

Robot manipulation learning from human demonstrations offers a rapid means to acquire skills but often lacks generalization across diverse scenes and object placements. This limitation hinders real-world applications, particularly in complex tasks requiring dexterous manipulation. Vision-Language-Action (VLA) paradigm leverages large-scale data to enhance generalization. However, due to data scarcity, VLA's performance remains limited. In this work, we introduce Object-Focus Actor (OFA), a novel, data-efficient approach for generalized dexterous manipulation. OFA exploits the consistent end trajectories observed in dexterous manipulation tasks, allowing for efficient policy training. Our method employs a hierarchical pipeline: object perception and pose estimation, pre-manipulation pose arrival and OFA policy execution. This process ensures that the manipulation is focused and efficient, even in varied backgrounds and positional layout. Comprehensive real-world experiments across seven tasks demonstrate that OFA significantly outperforms baseline methods in both positional and background generalization tests. Notably, OFA achieves robust performance with only 10 demonstrations, highlighting its data efficiency.

The chart-to-code generation task requires MLLMs to convert chart images into executable code. This task faces two major challenges: limited data diversity and insufficient maintenance of visual consistency between generated and original charts during training. Existing datasets mainly rely on seed data to prompt GPT models for code generation, resulting in homogeneous samples. To address this, we propose ReChartPrompt, which leverages real-world, human-designed charts from arXiv papers as prompts instead of synthetic seeds. Using the diverse styles and rich content of arXiv charts, we construct ReChartPrompt-240K, a large-scale and highly diverse dataset. Another challenge is that although SFT effectively improve code understanding, it often fails to ensure that generated charts are visually consistent with the originals. To address this, we propose ChartSimRL, a GRPO-based reinforcement learning algorithm guided by a novel chart similarity reward. This reward consists of attribute similarity, which measures the overlap of chart attributes such as layout and color between the generated and original charts, and visual similarity, which assesses similarity in texture and other overall visual features using convolutional neural networks. Unlike traditional text-based rewards such as accuracy or format rewards, our reward considers the multimodal nature of the chart-to-code task and effectively enhances the model's ability to accurately reproduce charts. By integrating ReChartPrompt and ChartSimRL, we develop the ChartMaster model, which achieves state-of-the-art results among 7B-parameter models and even rivals GPT-4o on various chart-to-code generation benchmarks. All resources are available at https://github.com/WentaoTan/ChartMaster.

Multimodal sarcasm detection, aiming to uncover sarcastic sentiment behind multimodal data, has gained substantial attention in multimodal communities. Recent advancements in multimodal sarcasm detection (MSD) methods have primarily focused on modality alignment with pre-trained vision-language (V-L) model. However, text-image pairs often exhibit weak or even opposite semantic correlations in MSD tasks. Consequently, directly aligning these modalities can potentially result in feature shift and inter-class confusion, ultimately hindering the model's ability. To alleviate this issue, we propose the Enhancing Semantic Awareness Model (ESAM) for multimodal sarcasm detection. Specifically, we first devise a Modality-decoupled Framework (MDF) to separate the textual and visual features from the fused multimodal representation. This decoupling enables the parallel integration of the Sentimental Congruity Constraint (SCC) within both visual and textual latent spaces, thereby enhancing the semantic awareness of different modalities. Furthermore, given that certain outlier samples with ambiguous sentiments can mislead the training and weaken the performance of SCC, we further incorporate Automatic Outlier Masking. This mechanism automatically detects and masks the outliers, guiding the model to focus on more informative samples during training. Experimental results on two public MSD datasets validate the robustness and superiority of our proposed ESAM model.

Embodied manipulation is a fundamental ability in the realm of embodied artificial intelligence. Although current embodied manipulation models show certain generalizations in specific settings, they struggle in new environments and tasks due to the complexity and diversity of real-world scenarios. The traditional end-to-end data collection and training manner leads to significant data demands. Decomposing end-to-end tasks into atomic skills helps reduce data requirements and improves the task success rate. However, existing methods are limited by predefined skill sets that cannot be dynamically updated. To address the issue, we introduce a three-wheeled data-driven method to build an atomic skill library. We divide tasks into subtasks using the Vision-Language-Planning (VLP). Then, atomic skill definitions are formed by abstracting the subtasks. Finally, an atomic skill library is constructed via data collection and Vision-Language-Action (VLA) fine-tuning. As the atomic skill library expands dynamically with the three-wheel update strategy, the range of tasks it can cover grows naturally. In this way, our method shifts focus from end-to-end tasks to atomic skills, significantly reducing data costs while maintaining high performance and enabling efficient adaptation to new tasks. Extensive experiments in real-world settings demonstrate the effectiveness and efficiency of our approach.

Text encoders in diffusion models have rapidly evolved, transitioning from CLIP to T5-XXL. Although this evolution has significantly enhanced the models' ability to understand complex prompts and generate text, it also leads to a substantial increase in the number of parameters. Despite T5 series encoders being trained on the C4 natural language corpus, which includes a significant amount of non-visual data, diffusion models with T5 encoder do not respond to those non-visual prompts, indicating redundancy in representational power. Therefore, it raises an important question: "Do we really need such a large text encoder?" In pursuit of an answer, we employ vision-based knowledge distillation to train a series of T5 encoder models. To fully inherit T5-XXL's capabilities, we constructed our dataset based on three criteria: image quality, semantic understanding, and text-rendering. Our results demonstrate the scaling down pattern that the distilled T5-base model can generate images of comparable quality to those produced by T5-XXL, while being 50 times smaller in size. This reduction in model size significantly lowers the GPU requirements for running state-of-the-art models such as FLUX and SD3, making high-quality text-to-image generation more accessible. Our code is available at https://lifuwang-66.github.io/ScalingDownTE/.