Pangu Light: Weight Re-Initialization for Pruning and Accelerating LLMs

Large Language Models (LLMs) deliver state-of-the-art capabilities across numerous tasks, but their immense size and inference costs pose significant computational challenges for practical deployment. While structured pruning offers a promising avenue for model compression, existing methods often struggle with the detrimental effects of aggressive, simultaneous width and depth reductions, leading to substantial performance degradation. This paper argues that a critical, often overlooked, aspect in making such aggressive joint pruning viable is the strategic re-initialization and adjustment of remaining weights to improve the model post-pruning training accuracies. We introduce Pangu Light, a framework for LLM acceleration centered around structured pruning coupled with novel weight re-initialization techniques designed to address this ``missing piece''. Our framework systematically targets multiple axes, including model width, depth, attention heads, and RMSNorm, with its effectiveness rooted in novel re-initialization methods like Cross-Layer Attention Pruning (CLAP) and Stabilized LayerNorm Pruning (SLNP) that mitigate performance drops by providing the network a better training starting point. Further enhancing efficiency, Pangu Light incorporates specialized optimizations such as absorbing Post-RMSNorm computations and tailors its strategies to Ascend NPU characteristics. The Pangu Light models consistently exhibit a superior accuracy-efficiency trade-off, outperforming prominent baseline pruning methods like Nemotron and established LLMs like Qwen3 series. For instance, on Ascend NPUs, Pangu Light-32B's 81.6 average score and 2585 tokens/s throughput exceed Qwen3-32B's 80.9 average score and 2225 tokens/s.

The Real Barrier to LLM Agent Usability is Agentic ROI

Large Language Model (LLM) agents represent a promising shift in human-AI interaction, moving beyond passive prompt-response systems to autonomous agents capable of reasoning, planning, and goal-directed action. Despite the widespread application in specialized, high-effort tasks like coding and scientific research, we highlight a critical usability gap in high-demand, mass-market applications. This position paper argues that the limited real-world adoption of LLM agents stems not only from gaps in model capabilities, but also from a fundamental tradeoff between the value an agent can provide and the costs incurred during real-world use. Hence, we call for a shift from solely optimizing model performance to a broader, utility-driven perspective: evaluating agents through the lens of the overall agentic return on investment (Agent ROI). By identifying key factors that determine Agentic ROI--information quality, agent time, and cost--we posit a zigzag development trajectory in optimizing agentic ROI: first scaling up to improve the information quality, then scaling down to minimize the time and cost. We outline the roadmap across different development stages to bridge the current usability gaps, aiming to make LLM agents truly scalable, accessible, and effective in real-world contexts.

Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Pangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUs

We present Pangu Ultra, a Large Language Model (LLM) with 135 billion parameters and dense Transformer modules trained on Ascend Neural Processing Units (NPUs). Although the field of LLM has been witnessing unprecedented advances in pushing the scale and capability of LLM in recent years, training such a large-scale model still involves significant optimization and system challenges. To stabilize the training process, we propose depth-scaled sandwich normalization, which effectively eliminates loss spikes during the training process of deep models. We pre-train our model on 13.2 trillion diverse and high-quality tokens and further enhance its reasoning capabilities during post-training. To perform such large-scale training efficiently, we utilize 8,192 Ascend NPUs with a series of system optimizations. Evaluations on multiple diverse benchmarks indicate that Pangu Ultra significantly advances the state-of-the-art capabilities of dense LLMs such as Llama 405B and Mistral Large 2, and even achieves competitive results with DeepSeek-R1, whose sparse model structure contains much more parameters. Our exploration demonstrates that Ascend NPUs are capable of efficiently and effectively training dense models with more than 100 billion parameters. Our model and system will be available for our commercial customers.

Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs

Given the large volume of side information from different modalities, multimodal recommender systems have become increasingly vital, as they exploit richer semantic information beyond user-item interactions. Recent works highlight that leveraging Graph Convolutional Networks (GCNs) to explicitly model multimodal item-item relations can significantly enhance recommendation performance. However, due to the inherent over-smoothing issue of GCNs, existing models benefit only from shallow GCNs with limited representation power. This drawback is especially pronounced when facing complex and high-dimensional patterns such as multimodal data, as it requires large-capacity models to accommodate complicated correlations. To this end, in this paper, we investigate bypassing GCNs when modeling multimodal item-item relationship. More specifically, we propose a Topology-aware Multi-Layer Perceptron (TMLP), which uses MLPs instead of GCNs to model the relationships between items. TMLP enhances MLPs with topological pruning to denoise item-item relations and intra (inter)-modality learning to integrate higher-order modality correlations. Extensive experiments on three real-world datasets verify TMLP's superiority over nine baselines. We also find that by discarding the internal message passing in GCNs, which is sensitive to node connections, TMLP achieves significant improvements in both training efficiency and robustness against existing models.

Unleashing the Potential of Multi-Channel Fusion in Retrieval for Personalized Recommendations

Recommender systems (RS) are pivotal in managing information overload in modern digital services. A key challenge in RS is efficiently processing vast item pools to deliver highly personalized recommendations under strict latency constraints. Multi-stage cascade ranking addresses this by employing computationally efficient retrieval methods to cover diverse user interests, followed by more precise ranking models to refine the results. In the retrieval stage, multi-channel retrieval is often used to generate distinct item subsets from different candidate generators, leveraging the complementary strengths of these methods to maximize coverage. However, forwarding all retrieved items overwhelms downstream rankers, necessitating truncation. Despite advancements in individual retrieval methods, multi-channel fusion, the process of efficiently merging multi-channel retrieval results, remains underexplored. We are the first to identify and systematically investigate multi-channel fusion in the retrieval stage. Current industry practices often rely on heuristic approaches and manual designs, which often lead to suboptimal performance. Moreover, traditional gradient-based methods like SGD are unsuitable for this task due to the non-differentiable nature of the selection process. In this paper, we explore advanced channel fusion strategies by assigning systematically optimized weights to each channel. We utilize black-box optimization techniques, including the Cross Entropy Method and Bayesian Optimization for global weight optimization, alongside policy gradient-based approaches for personalized merging. Our methods enhance both personalization and flexibility, achieving significant performance improvements across multiple datasets and yielding substantial gains in real-world deployments, offering a scalable solution for optimizing multi-channel fusion in retrieval.

All Roads Lead to Rome: Unveiling the Trajectory of Recommender Systems Across the LLM Era

Recommender systems (RS) are vital for managing information overload and delivering personalized content, responding to users' diverse information needs. The emergence of large language models (LLMs) offers a new horizon for redefining recommender systems with vast general knowledge and reasoning capabilities. Standing across this LLM era, we aim to integrate recommender systems into a broader picture, and pave the way for more comprehensive solutions for future research. Therefore, we first offer a comprehensive overview of the technical progression of recommender systems, particularly focusing on language foundation models and their applications in recommendation. We identify two evolution paths of modern recommender systems -- via list-wise recommendation and conversational recommendation. These two paths finally converge at LLM agents with superior capabilities of long-term memory, reflection, and tool intelligence. Along these two paths, we point out that the information effectiveness of the recommendation is increased, while the user's acquisition cost is decreased. Technical features, research methodologies, and inherent challenges for each milestone along the path are carefully investigated -- from traditional list-wise recommendation to LLM-enhanced recommendation to recommendation with LLM agents. Finally, we highlight several unresolved challenges crucial for the development of future personalization technologies and interfaces and discuss the future prospects.

Retrieval-Oriented Knowledge for Click-Through Rate Prediction

Click-through rate (CTR) prediction plays an important role in personalized recommendations. Recently, sample-level retrieval-based models (e.g., RIM) have achieved remarkable performance by retrieving and aggregating relevant samples. However, their inefficiency at the inference stage makes them impractical for industrial applications. To overcome this issue, this paper proposes a universal plug-and-play Retrieval-Oriented Knowledge (ROK) framework. Specifically, a knowledge base, consisting of a retrieval-oriented embedding layer and a knowledge encoder, is designed to preserve and imitate the retrieved & aggregated representations in a decomposition-reconstruction paradigm. Knowledge distillation and contrastive learning methods are utilized to optimize the knowledge base, and the learned retrieval-enhanced representations can be integrated with arbitrary CTR models in both instance-wise and feature-wise manners. Extensive experiments on three large-scale datasets show that ROK achieves competitive performance with the retrieval-based CTR models while reserving superior inference efficiency and model compatibility.

M-scan: A Multi-Scenario Causal-driven Adaptive Network for Recommendation

We primarily focus on the field of multi-scenario recommendation, which poses a significant challenge in effectively leveraging data from different scenarios to enhance predictions in scenarios with limited data. Current mainstream efforts mainly center around innovative model network architectures, with the aim of enabling the network to implicitly acquire knowledge from diverse scenarios. However, the uncertainty of implicit learning in networks arises from the absence of explicit modeling, leading to not only difficulty in training but also incomplete user representation and suboptimal performance. Furthermore, through causal graph analysis, we have discovered that the scenario itself directly influences click behavior, yet existing approaches directly incorporate data from other scenarios during the training of the current scenario, leading to prediction biases when they directly utilize click behaviors from other scenarios to train models. To address these problems, we propose the Multi-Scenario Causal-driven Adaptive Network M-scan). This model incorporates a Scenario-Aware Co-Attention mechanism that explicitly extracts user interests from other scenarios that align with the current scenario. Additionally, it employs a Scenario Bias Eliminator module utilizing causal counterfactual inference to mitigate biases introduced by data from other scenarios. Extensive experiments on two public datasets demonstrate the efficacy of our M-scan compared to the existing baseline models.

D2K: Turning Historical Data into Retrievable Knowledge for Recommender Systems

A vast amount of user behavior data is constantly accumulating on today's large recommendation platforms, recording users' various interests and tastes. Preserving knowledge from the old data while new data continually arrives is a vital problem for recommender systems. Existing approaches generally seek to save the knowledge implicitly in the model parameters. However, such a parameter-centric approach lacks scalability and flexibility -- the capacity is hard to scale, and the knowledge is inflexible to utilize. Hence, in this work, we propose a framework that turns massive user behavior data to retrievable knowledge (D2K). It is a data-centric approach that is model-agnostic and easy to scale up. Different from only storing unary knowledge such as the user-side or item-side information, D2K propose to store ternary knowledge for recommendation, which is determined by the complete recommendation factors -- user, item, and context. The knowledge retrieved by target samples can be directly used to enhance the performance of any recommendation algorithms. Specifically, we introduce a Transformer-based knowledge encoder to transform the old data into knowledge with the user-item-context cross features. A personalized knowledge adaptation unit is devised to effectively exploit the information from the knowledge base by adapting the retrieved knowledge to the target samples. Extensive experiments on two public datasets show that D2K significantly outperforms existing baselines and is compatible with a major collection of recommendation algorithms.