Chengxu Yang

Xinming Wei, Chao Jin, Tuo Dai et al.

Large-scale expert parallelism (EP) is becoming pivotal for training and serving frontier MoE models, but it also amplifies device-level expert load imbalance into compute stragglers, token all-to-all bottlenecks, and activation-memory spikes. Existing balancers redistribute experts periodically based on historical load, which becomes unreliable for production deployments with non-stationary load patterns. We present UltraEP, the first exact-load, real-time balancer for large-EP MoE training and serving prefill on rack-scale nodes (RSNs). Built upon the extended scale-up connectivity of RSNs, UltraEP rebalances every microbatch and layer on critical paths, which requires nontrivial co-design of plan solving and expert replication communication to minimize exposed overhead. To this end, UltraEP eagerly reacts to post-gating load with efficient quota-driven planning, and executes the resulting irregular expert-state transfers with RSN-native persistent tile streaming and relay-based fan-out mitigation. Averaged across MoE models from 106B to 671B parameters in training and prefill, UltraEP achieves 94.3% of the force-balanced ideal throughput, delivering 1.49$\times$ improvement over non-balancing, while reducing the final inter-rank imbalance from 1.30$-$4.01 to 1.01$-$1.04. Additionally, we validate UltraEP's scalability and robustness in production MoE training with 2560 GPUs.

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Zili Zhang, Chengxu Yang, Shenglong Zhang et al.

Training multimodal large language models (MLLMs) is challenged by both model and data heterogeneity. Existing systems redesign the training pipeline to address these challenges, but remain bound by a Pareto frontier between compute and memory efficiency, improving one only at the expense of the other. We present BigMac, a new training pipeline for multimodal LLMs. The core idea of BigMac is to elegantly nest the encoder and generator computation into the original LLM pipeline, forming a dependency-safe nested pipeline structure. With this design, BigMac reduces the activation memory complexity of the encoder and generator to O(1) while keeping the activation memory complexity of the LLM unchanged. At the same time, it achieves the same computational efficiency as the idealized setting with unlimited memory. As a result, BigMac breaks the Pareto frontier between computational efficiency and memory usage, enabling simultaneous optimization of both computation and memory in MLLM training. We evaluate BigMac on multiple MLLMs and training workloads. Experimental results show that BigMac achieves a 1.08$\times$-1.9$\times$ training speedup over baseline systems while maintaining stable memory usage as batch size increases.

Meituan LongCat Team, Bairui Wang, Bayan et al.

We introduce LongCat-Flash-Omni, a state-of-the-art open-source omni-modal model with 560 billion parameters, excelling at real-time audio-visual interaction. By adopting a curriculum-inspired progressive training strategy that transitions from simpler to increasingly complex modality sequence modeling tasks, LongCat-Flash-Omni attains comprehensive multimodal capabilities while maintaining strong unimodal capability. Building upon LongCat-Flash, which adopts a high-performance Shortcut-connected Mixture-of-Experts (MoE) architecture with zero-computation experts, LongCat-Flash-Omni integrates efficient multimodal perception and speech reconstruction modules. Despite its immense size of 560B parameters (with 27B activated), LongCat-Flash-Omni achieves low-latency real-time audio-visual interaction. For training infrastructure, we developed a modality-decoupled parallelism scheme specifically designed to manage the data and model heterogeneity inherent in large-scale multimodal training. This innovative approach demonstrates exceptional efficiency by sustaining over 90% of the throughput achieved by text-only training. Extensive evaluations show that LongCat-Flash-Omni achieves state-of-the-art performance on omni-modal benchmarks among open-source models. Furthermore, it delivers highly competitive results across a wide range of modality-specific tasks, including text, image, and video understanding, as well as audio understanding and generation. We provide a comprehensive overview of the model architecture design, training procedures, and data strategies, and open-source the model to foster future research and development in the community.

Zili Zhang, Yinmin Zhong, Chengxu Yang et al.

Agentic Reinforcement Learning (RL) enables LLMs to solve complex tasks by alternating between a data-collection rollout phase and a policy training phase. During rollout, the agent generates trajectories, i.e., multi-step interactions between LLMs and external tools. Yet, frequent tool calls induce long-tailed trajectory generation that bottlenecks rollouts. This stems from step-centric designs that ignore trajectory context, triggering three system problems for long-tail trajectory generation: queueing delays, interference overhead, and inflated per-token time. We propose Heddle, a trajectory-centric system to optimize the when, where, and how of agentic rollout execution. Heddle integrates three core mechanisms: trajectory-level scheduling using runtime prediction and progressive priority to minimize cumulative queueing; trajectory-aware placement via presorted dynamic programming and opportunistic migration during idle tool call intervals to minimize interference; and trajectory-adaptive resource manager that dynamically tunes model parallelism to accelerate the per-token time of long-tail trajectories while maintaining high throughput for short trajectories. Evaluations across diverse agentic RL workloads demonstrate that Heddle effectively neutralizes the long-tail bottleneck, achieving up to 2.5$\times$ higher end-to-end rollout throughput compared to state-of-the-art baselines.

Chao Jin, Xinming Wei, Yinmin Zhong et al.

Load imbalance is a long-standing challenge in Mixture-of-Experts (MoE) training and is exacerbated in reinforcement learning (RL) for LLMs, where hot experts can shift frequently across micro-batches. Existing MoE training systems rely on historical loads to predict future expert demand, making them less effective under sharp fluctuations. We propose ReLibra, an MoE RL training system that exploits a unique opportunity in RL's rollout-training workflow, routing replay, to enable fine-grained load balancing at micro-batch granularity. Because rollout and training process the same tokens with the same MoE parameters, the token-to-expert routing decisions are known before training starts. Leveraging this information, ReLibra places two MoE load-balancing mechanisms at inter- and intra-batch timescales, matching their communication patterns to hierarchical network bandwidths. At the inter-batch timescale, ReLibra performs expert reordering to redistribute experts for batch-level cross-node balancing; at the intra-batch timescale, it dynamically performs expert replication within a node to absorb micro-batch-level load fluctuations. Experiments on diverse MoE LLMs and RL workloads show that ReLibra improves training throughput by up to 1.6$\times$ over Megatron-LM and by up to 1.2$\times$ over EPLB, even when EPLB is given oracle loads. Moreover, ReLibra remains within 6%-10% of the throughput of an idealized balanced baseline.

Chengxu Yang, Jingling Yuan, Siqi Cai et al.

Hallucinations in large language models (LLMs) are commonly regarded as errors to be minimized. However, recent perspectives suggest that some hallucinations may encode creative or epistemically valuable content, a dimension that remains underquantified in current literature. Existing hallucination detection methods primarily focus on factual consistency, struggling to handle heterogeneous scientific tasks and balance creativity with accuracy. To address these challenges, we propose HIC-Bench, a novel evaluation framework that categorizes hallucinations into Intelligent Hallucinations (IH) and Defective Hallucinations (DH), enabling systematic investigation of their interplay in LLM creativity. HIC-Bench features three core characteristics: (1) Structured IH/DH Assessment. using a multi-dimensional metric matrix integrating Torrance Tests of Creative Thinking (TTCT) metrics (Originality, Feasibility, Value) with hallucination-specific dimensions (scientific plausibility, factual deviation); (2) Cross-Domain Applicability. spanning ten scientific domains with open-ended innovation tasks; and (3) Dynamic Prompt Optimization. leveraging the Dynamic Hallucination Prompt (DHP) to guide models toward creative and reliable outputs. The evaluation process employs multiple LLM judges, averaging scores to mitigate bias, with human annotators verifying IH/DH classifications. Experimental results reveal a nonlinear relationship between IH and DH, demonstrating that creativity and correctness can be jointly optimized. These insights position IH as a catalyst for creativity and reveal the ability of LLM hallucinations to drive scientific innovation.Additionally, the HIC-Bench offers a valuable platform for advancing research into the creative intelligence of LLM hallucinations.

Chengxu Yang, Jingling Yuan, Chuang Hu et al.

Multimodal Large Language Models often suffer from object hallucination. While existing research utilizes attention enhancement and visual retracing, we find these works lack sufficient interpretability regarding attention drift in final model stages. In this paper, we investigate the layer wise evolution of visual features and discover that hallucination stems from deep layer attention regressing toward initial visual noise from early layers. We observe that output reliability depends on acquiring visual anchors at intermediate layers rather than final layers. Based on these insights, we propose CLVA, which stands for Cross-Layer Visual Anchors, a training free method that reinforces critical mid layer features while suppressing regressive noise. This approach effectively pulls deep layer attention back to correct visual regions by utilizing essential anchors captured from attention dynamics. We evaluate our method across diverse architectures and benchmarks, demonstrating outstanding performance without significant increase in computational time and GPU memory.

Tao Tang, Chengxu Yang

The core role of medical images in disease diagnosis makes their quality directly affect the accuracy of clinical judgment. However, due to factors such as low-dose scanning, equipment limitations and imaging artifacts, medical images are often accompanied by non-uniform noise interference, which seriously affects structure recognition and lesion detection. This paper proposes a medical image adaptive denoising model (MI-ND) that integrates multi-scale convolutional and Transformer architecture, introduces a noise level estimator (NLE) and a noise adaptive attention module (NAAB), and realizes channel-spatial attention regulation and cross-modal feature fusion driven by noise perception. Systematic testing is carried out on multimodal public datasets. Experiments show that this method significantly outperforms the comparative methods in image quality indicators such as PSNR, SSIM, and LPIPS, and improves the F1 score and ROC-AUC in downstream diagnostic tasks, showing strong prac-tical value and promotional potential. The model has outstanding benefits in structural recovery, diagnostic sensitivity, and cross-modal robustness, and provides an effective solution for medical image enhancement and AI-assisted diagnosis and treatment.

Tao Tang, Chengxu Yang

Pansharpening refers to the process of integrating a high resolution panchromatic (PAN) image with a lower resolution multispectral (MS) image to generate a fused product, which is pivotal in remote sensing. Despite the effectiveness of CNNs in addressing this challenge, they are inherently constrained by the uniform application of convolutional kernels across all spatial positions, overlooking local content variations. To overcome this issue, we introduce RAPNet, a new architecture that leverages content-adaptive convolution. At its core, RAPNet employs the Receptive-field Adaptive Pansharpening Convolution (RAPConv), designed to produce spatially adaptive kernels responsive to local feature context, thereby enhancing the precision of spatial detail extraction. Additionally, the network integrates the Pansharpening Dynamic Feature Fusion (PAN-DFF) module, which incorporates an attention mechanism to achieve an optimal balance between spatial detail enhancement and spectral fidelity. Comprehensive evaluations on publicly available datasets confirm that RAPNet delivers superior performance compared to existing approaches, as demonstrated by both quantitative metrics and qualitative assessments. Ablation analyses further substantiate the effectiveness of the proposed adaptive components.

Gaofeng Cheng, Haitian Lu, Chengxu Yang et al.

Effectively distinguishing the pronunciation correlations between different written texts is a significant issue in linguistic acoustics. Traditionally, such pronunciation correlations are obtained through manually designed pronunciation lexicons. In this paper, we propose a data-driven method to automatically acquire these pronunciation correlations, called automatic text pronunciation correlation (ATPC). The supervision required for this method is consistent with the supervision needed for training end-to-end automatic speech recognition (E2E-ASR) systems, i.e., speech and corresponding text annotations. First, the iteratively-trained timestamp estimator (ITSE) algorithm is employed to align the speech with their corresponding annotated text symbols. Then, a speech encoder is used to convert the speech into speech embeddings. Finally, we compare the speech embeddings distances of different text symbols to obtain ATPC. Experimental results on Mandarin show that ATPC enhances E2E-ASR performance in contextual biasing and holds promise for dialects or languages lacking artificial pronunciation lexicons.

Chengxu Yang, Qipeng Wang, Mengwei Xu et al.

Federated learning (FL) is an emerging, privacy-preserving machine learning paradigm, drawing tremendous attention in both academia and industry. A unique characteristic of FL is heterogeneity, which resides in the various hardware specifications and dynamic states across the participating devices. Theoretically, heterogeneity can exert a huge influence on the FL training process, e.g., causing a device unavailable for training or unable to upload its model updates. Unfortunately, these impacts have never been systematically studied and quantified in existing FL literature. In this paper, we carry out the first empirical study to characterize the impacts of heterogeneity in FL. We collect large-scale data from 136k smartphones that can faithfully reflect heterogeneity in real-world settings. We also build a heterogeneity-aware FL platform that complies with the standard FL protocol but with heterogeneity in consideration. Based on the data and the platform, we conduct extensive experiments to compare the performance of state-of-the-art FL algorithms under heterogeneity-aware and heterogeneity-unaware settings. Results show that heterogeneity causes non-trivial performance degradation in FL, including up to 9.2% accuracy drop, 2.32x lengthened training time, and undermined fairness. Furthermore, we analyze potential impact factors and find that device failure and participant bias are two potential factors for performance degradation. Our study provides insightful implications for FL practitioners. On the one hand, our findings suggest that FL algorithm designers consider necessary heterogeneity during the evaluation. On the other hand, our findings urge system providers to design specific mechanisms to mitigate the impacts of heterogeneity.