Searcharxiv⌕ Search

arXiv subjects

Yunhe Wang

Publications and source records attributed to Yunhe Wang.

At least 37 records · Page 2Linked to original sources
arXiv2026

ReThinker: Scientific Reasoning by Rethinking with Guided Reflection and Confidence Control

Expert-level scientific reasoning remains challenging for large language models, particularly on benchmarks such as Humanity's Last Exam (HLE), where rigid tool pipelines, brittle multi-agent coordination, and inefficient test-time scaling often limit performance. We introduce ReThinker, a confidence-aware agentic framework that orchestrates retrieval, tool use, and multi-agent reasoning through a stage-wise Solver-Critic-Selector architecture. Rather than following a fixed pipeline, ReThinker dynamically allocates computation based on model confidence, enabling adaptive tool invocation, guided multi-dimensional reflection, and robust confidence-weighted selection. To support scalable training without human annotation, we further propose a reverse data synthesis pipeline and an adaptive trajectory recycling strategy that transform successful reasoning traces into high-quality supervision. Experiments on HLE, GAIA, and XBench demonstrate that ReThinker consistently outperforms state-of-the-art foundation models with tools and existing deep research systems, achieving state-of-the-art results on expert-level reasoning tasks.

cs.AI↗
arXiv2026

MemoryFormer: Minimize Transformer Computation by Removing Fully-Connected Layers

In order to reduce the computational complexity of large language models, great efforts have been made to to improve the efficiency of transformer models such as linear attention and flash-attention. However, the model size and corresponding computational complexity are constantly scaled up in pursuit of higher performance. In this work, we present MemoryFormer, a novel transformer architecture which significantly reduces the computational complexity (FLOPs) from a new perspective. We eliminate nearly all the computations of the transformer model except for the necessary computation required by the multi-head attention operation. This is made possible by utilizing an alternative method for feature transformation to replace the linear projection of fully-connected layers. Specifically, we first construct a group of in-memory lookup tables that store a large amount of discrete vectors to replace the weight matrix used in linear projection. We then use a hash algorithm to retrieve a correlated subset of vectors dynamically based on the input embedding. The retrieved vectors combined together will form the output embedding, which provides an estimation of the result of matrix multiplication operation in a fully-connected layer. Compared to conducting matrix multiplication, retrieving data blocks from memory is a much cheaper operation which requires little computations. We train MemoryFormer from scratch and conduct extensive experiments on various benchmarks to demonstrate the effectiveness of the proposed model.

cs.CL↗
arXiv2026

An Empirical Study of World Model Quantization

World models learn an internal representation of environment dynamics, enabling agents to simulate and reason about future states within a compact latent space for tasks such as planning, prediction, and inference. However, running world models rely on hevay computational cost and memory footprint, making model quantization essential for efficient deployment. To date, the effects of post-training quantization (PTQ) on world models remain largely unexamined. In this work, we present a systematic empirical study of world model quantization using DINO-WM as a representative case, evaluating diverse PTQ methods under both weight-only and joint weight-activation settings. We conduct extensive experiments on different visual planning tasks across a wide range of bit-widths, quantization granularities, and planning horizons up to 50 iterations. Our results show that quantization effects in world models extend beyond standard accuracy and bit-width trade-offs: group-wise weight quantization can stabilize low-bit rollouts, activation quantization granularity yields inconsistent benefits, and quantization sensitivity is highly asymmetric between encoder and predictor modules. Moreover, aggressive low-bit quantization significantly degrades the alignment between the planning objective and task success, leading to failures that cannot be remedied by additional optimization. These findings reveal distinct quantization-induced failure modes in world model-based planning and provide practical guidance for deploying quantized world models under strict computational constraints. The code will be available at https://github.com/huawei-noah/noah-research/tree/master/QuantWM.

cs.LG↗
arXiv2026

EAQuant: Enhancing Post-Training Quantization for MoE Models via Expert-Aware Optimization

Mixture-of-Experts (MoE) models enable scalable computation and performance in large-scale deep learning but face quantization challenges due to sparse expert activation and dynamic routing. Existing post-training quantization (PTQ) methods fail to address activation outliers, routing instability, and sparse expert calibration, leading to significant performance degradation. To address this, we propose EAQuant, a PTQ framework tailored for MoE architectures. Our method introduces three expert-aware innovations: (1) smoothing aggregation to suppress activation outliers, (2) routing consistency alignment to preserve expert selection post-quantization, and (3) calibration data balance to optimize sparsely activated experts. These strategies collectively enable robust, high-precision quantization of MoE models under ultra-low-bit constraints.Extensive experiments across several extreme quantization settings (e.g., W4A4/W3A4/W3A3/W2A4) demonstrate that EAQuant significantly outperforms existing methods, achieving average accuracy improvements of 1.15 - 13.81% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks and robust performance retention under aggressive quantization. By integrating these innovations, EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression.Our code is available at https://github.com/darren-fzq1/EAQuant.

cs.CL↗
arXiv2026

VersatileFFN: Achieving Parameter Efficiency in LLMs via Adaptive Wide-and-Deep Reuse

The rapid scaling of Large Language Models (LLMs) has achieved remarkable performance, but it also leads to prohibitive memory costs. Existing parameter-efficient approaches such as pruning and quantization mainly compress pretrained models without enhancing architectural capacity, thereby hitting the representational ceiling of the base model. In this work, we propose VersatileFFN, a novel feed-forward network (FFN) that enables flexible reuse of parameters in both width and depth dimensions within a fixed parameter budget. Inspired by the dual-process theory of cognition, VersatileFFN comprises two adaptive pathways: a width-versatile path that generates a mixture of sub-experts from a single shared FFN, mimicking sparse expert routing without increasing parameters, and a depth-versatile path that recursively applies the same FFN to emulate deeper processing for complex tokens. A difficulty-aware gating dynamically balances the two pathways, steering "easy" tokens through the efficient width-wise route and allocating deeper iterative refinement to "hard" tokens. Crucially, both pathways reuse the same parameters, so all additional capacity comes from computation rather than memory. Experiments across diverse benchmarks and model scales demonstrate the effectiveness of the method. The code is available at https://github.com/huawei-noah/noah-research/tree/master/VersatileFFN.

cs.CL↗
arXiv2026

From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs

Diffusion Language Models (DLMs) enable fast generation, yet training large DLMs from scratch is costly. As a practical shortcut, adapting off-the-shelf Auto-Regressive (AR) model weights into a DLM could quickly equip the DLM with strong long-context generation capabilies. Prior "adaptation" attempts either modify logits or randomly grow attention masks to Full-Sequence diffusion, or simply transplant AR weights into a Block-Diffusion recipe, leaving two key questions unaddressed: where is the final destination of adaptation, and how to adapt better? For manifold benefits, we reframe the whole AR-to-DLM adaptation under the Block-Diffusion paradigm, transitioning from block size 1 to the final Block-Diffusion state. Concretely, the principled pathway of adaptation is designed as follows: we keep a context-causal path where causal attention is kept in the prefix, an efficient parallel adaptation procedure where an AR guidance is maintained, and gradual increment of the generation block size for a smoother transition. Built on these components, the adaptation is proved competitive on various models at different scales. With better adaptation, we propose NBDiff-7B that could inherit the long-context modeling and reasoning capabilities, and achieve state-of-the-art performance among the 7B-class DLMs. Codes: https://github.com/YuchuanTian/NBDiff.

cs.CL↗
arXiv2026

Diffusion In Diffusion: Reclaiming Global Coherence in Semi-Autoregressive Diffusion

One of the most compelling features of global discrete diffusion language models is their global bidirectional contextual capability. However, existing block-based diffusion studies tend to introduce autoregressive priors, which, while offering benefits, can cause models to lose this global coherence at the macro level. To regain global contextual understanding while preserving the advantages of the semi-autoregressive paradigm, we propose Diffusion in Diffusion, a 'draft-then-refine' framework designed to overcome the irreversibility and myopia problems inherent in block diffusion models. Our approach first employs block diffusion to generate rapid drafts using small blocks, then refines these drafts through global bidirectional diffusion with a larger bidirectional receptive field. We utilize snapshot confidence remasking to identify the most critical tokens that require modification, and apply mix-scale training to expand the block diffusion model's global capabilities. Empirical results demonstrate that our approach sets a new benchmark for discrete diffusion models on the OpenWebText dataset. Using only 26% of the fine-tuning budget of baseline models, we reduce generative perplexity from 25.7 to 21.9, significantly narrowing the performance gap with autoregressive models.

cs.LG↗
arXiv2026

Deferred Commitment Decoding for Diffusion Language Models

Diffusion language models (DLMs) have recently emerged as a strong alternative to autoregressive models by enabling parallel text generation. To improve inference efficiency and KV-cache compatibility, prior work commonly adopts block-based diffusion, decoding tokens block by block. However, this paradigm suffers from a structural limitation that we term Boundary-Induced Context Truncation (BICT): undecoded tokens near block boundaries are forced to commit without access to nearby future context, even when such context could substantially reduce uncertainty. This limitation degrades decoding certainty and generation quality, especially for tasks requiring precise reasoning, such as mathematical problem solving and code generation. We propose Deferred Commitment Decoding (DCD), a novel, training-free decoding strategy that mitigates this issue. DCD maintains a certainty-aware sliding window over masked tokens, resolving low-uncertainty tokens early while deferring high-uncertainty tokens until sufficient contextual evidence becomes available. Extensive experiments across multiple diffusion language models, benchmarks, and caching configurations show that DCD improves generation accuracy by 1.73% with comparable time on average compared to fixed block-based diffusion methods, with the most significant improvement reaching 16.5%. These results demonstrate that deferring token commitment based on uncertainty is a simple yet effective principle for improving both the quality and efficiency of diffusion language model decoding.

cs.CL↗
arXiv2026

Top 10 Open Challenges Steering the Future of Diffusion Language Model and Its Variants

The paradigm of Large Language Models (LLMs) is currently defined by auto-regressive (AR) architectures, which generate text through a sequential ``brick-by-brick'' process. Despite their success, AR models are inherently constrained by a causal bottleneck that limits global structural foresight and iterative refinement. Diffusion Language Models (DLMs) offer a transformative alternative, conceptualizing text generation as a holistic, bidirectional denoising process akin to a sculptor refining a masterpiece. However, the potential of DLMs remains largely untapped as they are frequently confined within AR-legacy infrastructures and optimization frameworks. In this Perspective, we identify ten fundamental challenges ranging from architectural inertia and gradient sparsity to the limitations of linear reasoning that prevent DLMs from reaching their ``GPT-4 moment''. We propose a strategic roadmap organized into four pillars: foundational infrastructure, algorithmic optimization, cognitive reasoning, and unified multimodal intelligence. By shifting toward a diffusion-native ecosystem characterized by multi-scale tokenization, active remasking, and latent thinking, we can move beyond the constraints of the causal horizon. We argue that this transition is essential for developing next-generation AI capable of complex structural reasoning, dynamic self-correction, and seamless multimodal integration.

cs.CL↗
arXiv2026

U-REPA: Aligning Diffusion U-Nets to ViTs

Representation Alignment (REPA) that aligns Diffusion Transformer (DiT) hidden-states with ViT visual encoders has proven highly effective in DiT training, demonstrating superior convergence properties, but it has not been validated on the canonical diffusion U-Net architecture that shows faster convergence compared to DiTs. However, adapting REPA to U-Net architectures presents unique challenges: (1) different block functionalities necessitate revised alignment strategies; (2) spatial-dimension inconsistencies emerge from U-Net's spatial downsampling operations; (3) space gaps between U-Net and ViT hinder the effectiveness of tokenwise alignment. To encounter these challenges, we propose \textbf{U-REPA}, a representation alignment paradigm that bridges U-Net hidden states and ViT features as follows: Firstly, we propose via observation that due to skip connection, the middle stage of U-Net is the best alignment option. Secondly, we propose upsampling of U-Net features after passing them through MLPs. Thirdly, we observe difficulty when performing tokenwise similarity alignment, and further introduces a manifold loss that regularizes the relative similarity between samples. Experiments indicate that the resulting U-REPA could achieve excellent generation quality and greatly accelerates the convergence speed. With CFG guidance interval, U-REPA could reach $FID<1.5$ in 200 epochs or 1M iterations on ImageNet 256 $\times$ 256, and needs only half the total epochs to perform better than REPA under sd-vae-ft-ema. Codes: https://github.com/YuchuanTian/U-REPA

cs.CV↗
arXiv2025

Nexus: Higher-Order Attention Mechanisms in Transformers

Transformers have achieved significant success across various domains, relying on self-attention to capture dependencies. However, the standard first-order attention mechanism is often limited by a low-rank bottleneck, struggling to capture intricate, multi-hop relationships within a single layer. In this paper, we propose the Nexus, a novel architecture designed to enhance representational power through a recursive framework. Unlike standard approaches that use static linear projections for Queries and Keys, Nexus dynamically refines these representations via nested self-attention mechanisms. Specifically, the Query and Key vectors are themselves outputs of inner attention loops, allowing tokens to aggregate global context and model high-order correlations \textit{prior} to the final attention computation. We enforce a parameter-efficient weight-sharing strategy across recursive steps, ensuring that this enhanced expressivity incurs $\mathcal{O}(1)$ additional parameters. We provide theoretical analysis demonstrating that our method breaks the linear bottleneck of standard attention. Empirically, Nexus outperforms standard Transformers on multiple benchmarks.

cs.CL↗
arXiv2025

ROOT: Robust Orthogonalized Optimizer for Neural Network Training

The optimization of large language models (LLMs) remains a critical challenge, particularly as model scaling exacerbates sensitivity to algorithmic imprecision and training instability. Recent advances in optimizers have improved convergence efficiency through momentum orthogonalization, but suffer from two key robustness limitations: dimensional fragility in orthogonalization precision and vulnerability to outlier-induced noise. To address these robustness challenges, we introduce ROOT, a Robust Orthogonalized Optimizer that enhances training stability through dual robustness mechanisms. First, we develop a dimension-robust orthogonalization scheme using adaptive Newton iterations with fine-grained coefficients tailored to specific matrix sizes, ensuring consistent precision across diverse architectural configurations. Second, we introduce an optimization-robust framework via proximal optimization that suppresses outlier noise while preserving meaningful gradient directions. Extensive experiments demonstrate that ROOT achieves significantly improved robustness, with faster convergence and superior final performance compared to both Muon and Adam-based optimizers, particularly in noisy and non-convex scenarios. Our work establishes a new paradigm for developing robust and precise optimizers capable of handling the complexities of modern large-scale model training. The code will be available at https://github.com/huawei-noah/noah-research/tree/master/ROOT.

cs.LG↗
arXiv2025

ScaleNet: Scaling up Pretrained Neural Networks with Incremental Parameters

Recent advancements in vision transformers (ViTs) have demonstrated that larger models often achieve superior performance. However, training these models remains computationally intensive and costly. To address this challenge, we introduce ScaleNet, an efficient approach for scaling ViT models. Unlike conventional training from scratch, ScaleNet facilitates rapid model expansion with negligible increases in parameters, building on existing pretrained models. This offers a cost-effective solution for scaling up ViTs. Specifically, ScaleNet achieves model expansion by inserting additional layers into pretrained ViTs, utilizing layer-wise weight sharing to maintain parameters efficiency. Each added layer shares its parameter tensor with a corresponding layer from the pretrained model. To mitigate potential performance degradation due to shared weights, ScaleNet introduces a small set of adjustment parameters for each layer. These adjustment parameters are implemented through parallel adapter modules, ensuring that each instance of the shared parameter tensor remains distinct and optimized for its specific function. Experiments on the ImageNet-1K dataset demonstrate that ScaleNet enables efficient expansion of ViT models. With a 2$\times$ depth-scaled DeiT-Base model, ScaleNet achieves a 7.42% accuracy improvement over training from scratch while requiring only one-third of the training epochs, highlighting its efficiency in scaling ViTs. Beyond image classification, our method shows significant potential for application in downstream vision areas, as evidenced by the validation in object detection task.

cs.CV↗
arXiv2025

Revealing the Power of Post-Training for Small Language Models via Knowledge Distillation

The rapid advancement of large language models (LLMs) has significantly advanced the capabilities of artificial intelligence across various domains. However, their massive scale and high computational costs render them unsuitable for direct deployment in resource-constrained edge environments. This creates a critical need for high-performance small models that can operate efficiently at the edge. Yet, after pre-training alone, these smaller models often fail to meet the performance requirements of complex tasks. To bridge this gap, we introduce a systematic post-training pipeline that efficiently enhances small model accuracy. Our post training pipeline consists of curriculum-based supervised fine-tuning (SFT) and offline on-policy knowledge distillation. The resulting instruction-tuned model achieves state-of-the-art performance among billion-parameter models, demonstrating strong generalization under strict hardware constraints while maintaining competitive accuracy across a variety of tasks. This work provides a practical and efficient solution for developing high-performance language models on Ascend edge devices.

cs.CV↗
arXiv2025

LLM Data Selection and Utilization via Dynamic Bi-level Optimization

While large-scale training data is fundamental for developing capable large language models (LLMs), strategically selecting high-quality data has emerged as a critical approach to enhance training efficiency and reduce computational costs. Current data selection methodologies predominantly rely on static, training-agnostic criteria, failing to account for the dynamic model training and data interactions. In this paper, we propose a new Data Weighting Model (DWM) to adjust the weight of selected data within each batch to achieve a dynamic data utilization during LLM training. Specially, to better capture the dynamic data preference of the trained model, a bi-level optimization framework is implemented to update the weighting model. Our experiments demonstrate that DWM enhances the performance of models trained with randomly-selected data, and the learned weighting model can be transferred to enhance other data selection methods and models of different sizes. Moreover, we further analyze how a model's data preferences evolve throughout training, providing new insights into the data preference of the model during training.

cs.LG↗
arXiv2025

Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

This work presents Pangu Embedded, an efficient Large Language Model (LLM) reasoner developed on Ascend Neural Processing Units (NPUs), featuring flexible fast and slow thinking capabilities. Pangu Embedded addresses the significant computational costs and inference latency challenges prevalent in existing reasoning-optimized LLMs. We propose a two-stage training framework for its construction. In Stage 1, the model is finetuned via an iterative distillation process, incorporating inter-iteration model merging to effectively aggregate complementary knowledge. This is followed by reinforcement learning on Ascend clusters, optimized by a latency-tolerant scheduler that combines stale synchronous parallelism with prioritized data queues. The RL process is guided by a Multi-source Adaptive Reward System (MARS), which generates dynamic, task-specific reward signals using deterministic metrics and lightweight LLM evaluators for mathematics, coding, and general problem-solving tasks. Stage 2 introduces a dual-system framework, endowing Pangu Embedded with a "fast" mode for routine queries and a deeper "slow" mode for complex inference. This framework offers both manual mode switching for user control and an automatic, complexity-aware mode selection mechanism that dynamically allocates computational resources to balance latency and reasoning depth. Experimental results on benchmarks including AIME 2024, GPQA, and LiveCodeBench demonstrate that Pangu Embedded with 7B parameters, outperforms similar-size models like Qwen3-8B and GLM4-9B. It delivers rapid responses and state-of-the-art reasoning quality within a single, unified model architecture, highlighting a promising direction for developing powerful yet practically deployable LLM reasoners.

cs.CL↗
arXiv2025

Pangu Pro MoE: Mixture of Grouped Experts for Efficient Sparsity

The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.

cs.CL↗
arXiv2025

SlimLLM: Accurate Structured Pruning for Large Language Models

Large language models(LLMs) have garnered significant attention and demonstrated impressive capabilities in a wide range of applications. However, due to their enormous computational costs, the deployment and application of LLMs are often severely limited. To address this issue, structured pruning is an effective solution to compress the parameters of LLMs. Determining the importance of each sub-module in LLMs and minimizing performance loss are critical issues that need to be carefully addressed in structured pruning. In this paper, we propose an effective and fast structured pruning method named SlimLLM for large language models. For channel and attention head pruning, we evaluate the importance based on the entire channel or head, rather than merely aggregating the importance of individual elements within a sub-module. This approach enables a more holistic consideration of the interdependence among elements within the sub-module. In addition, we design a simple linear regression strategy for the output matrix to quickly recover performance. We also propose layer-based importance ratio to determine the pruning ratio for each layer. Based on the LLaMA benchmark results, our SlimLLM outperforms other methods and achieves state-of-the-art performance.

cs.LG↗