arXiv subjects

Chuanyang Zheng

Publications and source records attributed to Chuanyang Zheng.

At least 19 recordsLinked to original sources

arXiv2026

UniPool: Learning Expert-to-Layer Ownership from Brief Global Access

Most Mixture-of-Experts (MoE) transformers preassign each expert to one layer for the whole of training. We study whether a brief global-access phase can learn a compatible expert-to-layer allocation that is then executed privately. UniPool-Lock gives every layer its own router over one global pool, balances aggregate pool usage, and scores experts with a scale-stable NormRouter. After the short ownership-learning phase, it assigns each expert to one layer, locks the disjoint allocation, and continues training as a layer-private MoE. With the same expert-FFN budget and routed expert FLOPs, this recipe lowers held-out loss relative to vanilla MoE by 0.024-0.037 across five dense-equivalent scales from 182M to 1.5B, including -0.0247 at 1.5B after 60B tokens. The ownership-learning phase lasts 2K steps, about 3.3% of training, and post-lock step time is within -0.7% to +2.8% of vanilla MoE in our throughput measurements. Controls locate the gain in full-pool training and in the allocation it produces: a random disjoint allocation fixed at initialization, trained with the same router and losses, matches vanilla, whereas locking the allocation learned during the full-pool phase recovers nearly all of the persistent full-pool gain, and substituting a random allocation at the lock forfeits about half of it. Keeping full-pool access throughout training (UniPool-Full) also improves over vanilla MoE at four scales and outperforms it with only 66.7% (182M) to 50% (469M and 650M) of its expert parameters.

cs.LG↗

arXiv2026

Automatic Rank Allocation for Low-Rank Adaptation in Large Language Models via lp Regularization

Low-rank adaptation (LoRA) has become a popular parameter-efficient fine-tuning method for large language models. A key challenge in LoRA is how to determine the rank of each adaptation matrix, as rank directly controls its capacity and efficiency. Existing adaptive-rank methods typically allocate ranks according to manually designed importance scores, which are not directly derived from an optimization objective. In this work, we propose $\ell_p$-LoRA, a principled rank-allocation method based on $\ell_p$ regularization with $0<p<1$, which is a classical sparsity-inducing technique in signal processing and statistics. Specifically, we regularize the energy of each rank-one LoRA component, encouraging redundant components to vanish while preserving important ones. We derive the corresponding proximal subproblem and reduce the matrix optimization to a two-dimensional problem, leading to an implicit thresholding criterion for identifying redundant components. Experiments on natural language understanding and question-answering tasks demonstrate that the proposed method achieves competitive performance with existing LoRA baselines.

cs.LG↗

arXiv2026

Cubit: Token Mixer with Kernel Ridge Regression

Since its introduction in 2017, the Transformer has become one of the most widely adopted architectures in modern deep learning. Despite extensive efforts to improve positional encoding, attention mechanisms, and feed-forward networks, the core token-mixing mechanism in Transformers remains attention. In this work, we show that the attention module in Transformers can be interpreted as performing Nadaraya-Watson regression, where it computes similarities between tokens and aggregates the corresponding values accordingly. Motivated by this perspective, we propose Cubit, a potential next-generation architecture that leverages Kernel Ridge Regression (KRR), while the vanilla Transformer relies on Nadaraya-Watson regression. Specifically, Cubit modifies the classical attention computation by incorporating the closed-form solution of KRR, combining value aggregation through kernel similarities with normalization via the inverse of the kernel matrix. To improve the training stability, we further propose the Limited-Range Rescale (LRR), which rescales the value layer within a controlled range. We argue that Cubit, as a KRR-based architecture, provides a stronger mathematical foundation than the vanilla Transformer, whose attention mechanism corresponds to Nadaraya-Watson regression. We validate this claim through comprehensive experiments. The experimental results suggest that Cubit may exhibit stronger long-sequence modeling capability. In particular, its performance gain over the Transformer appears to increase as the training sequence length grows.

cs.LG↗

arXiv2026

AdaSpark: Adaptive Sparsity for Efficient Long-Video Understanding

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

cs.CV↗

arXiv2026

ATTS: Asynchronous Test-Time Scaling via Conformal Prediction

Large language models (LLMs) benefit from test-time scaling but are often hampered by high inference latency. Speculative decoding is a natural way to accelerate the scaling process; however, scaling along both the parallel and sequential dimensions poses significant challenges, including substantial memory-bound execution and synchronization overhead. We introduce ATTS (Asynchronous Test-Time Scaling), a statistically guaranteed adaptive scaling framework that follows the hypothesis testing process to address these challenges. By revisiting arithmetic intensity, ATTS identifies synchronization as the primary bottleneck. It enables asynchronous inference through online calibration and proposes an ordinal classification algorithm that supports a three-stage rejection sampling pipeline, scaling along both the sequential and parallel axes. Across experiments on the MATH, AMC23, AIME24, and AIME25 datasets and across multiple draft-target model families, we show that ATTS delivers up to 56.7x speedup in test-time scaling and a 4.14x throughput improvement, while maintaining accurate control of the rejection rate, reducing latency and memory overhead, and incurring no accuracy loss. By scaling both in parallel and sequential dimensions, we enable the 1.5B/70B draft/target model combination to achieve the performance of the state-of-the-art reasoning model o3-mini (high) on the AIME dataset. We have released the code at https://github.com/menik1126/asynchronous-test-time-scaling.

cs.CL↗

arXiv2026

GeoNorm: Unify Pre-Norm and Post-Norm with Geodesic Optimization

The placement of normalization layers, specifically Pre-Norm and Post-Norm, remains an open question in Transformer architecture design. In this work, we rethink these approaches through the lens of manifold optimization, interpreting the outputs of the Feed-Forward Network (FFN) and attention layers as update directions in optimization. Building on this perspective, we introduce GeoNorm, a novel method that replaces standard normalization with geodesic updates on the manifold. Furthermore, analogous to learning rate schedules, we propose a layer-wise update decay for the FFN and attention components. Comprehensive experiments demonstrate that GeoNorm consistently outperforms existing normalization methods in Transformer models. Crucially, GeoNorm can be seamlessly integrated into standard Transformer architectures, achieving performance improvements with negligible additional computational cost.

cs.LG↗

arXiv2026

DQ-LoRe: Dual Queries with Low Rank Approximation Re-ranking for In-Context Learning

Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning. Dual Queries first query LLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question's knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5% to 94.2%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundary of in-context learning and opens up new avenues for addressing complex reasoning challenges. Our code is released at https://github.com/menik1126/DQ-LoRe

cs.CL↗

arXiv2025

SAS: Simulated Attention Score

The attention mechanism is a core component of the Transformer architecture. Various methods have been developed to compute attention scores, including multi-head attention (MHA), multi-query attention, group-query attention and so on. We further analyze the MHA and observe that its performance improves as the number of attention heads increases, provided the hidden size per head remains sufficiently large. Therefore, increasing both the head count and hidden size per head with minimal parameter overhead can lead to significant performance gains at a low cost. Motivated by this insight, we introduce Simulated Attention Score (SAS), which maintains a compact model size while simulating a larger number of attention heads and hidden feature dimension per head. This is achieved by projecting a low-dimensional head representation into a higher-dimensional space, effectively increasing attention capacity without increasing parameter count. Beyond the head representations, we further extend the simulation approach to feature dimension of the key and query embeddings, enhancing expressiveness by mimicking the behavior of a larger model while preserving the original model size. To control the parameter cost, we also propose Parameter-Efficient Attention Aggregation (PEAA). Comprehensive experiments on a variety of datasets and tasks demonstrate the effectiveness of the proposed SAS method, achieving significant improvements over different attention variants.

cs.CL↗

arXiv2025

Benchmarking the Thinking Mode of Multimodal Large Language Models in Clinical Tasks

A recent advancement in Multimodal Large Language Models (MLLMs) research is the emergence of "reasoning MLLMs" that offer explicit control over their internal thinking processes (normally referred as the "thinking mode") alongside the standard "non-thinking mode". This capability allows these models to engage in a step-by-step process of internal deliberation before generating a final response. With the rapid transition to and adoption of these "dual-state" MLLMs, this work rigorously evaluated how the enhanced reasoning processes of these MLLMs impact model performance and reliability in clinical tasks. This paper evaluates the active "thinking mode" capabilities of two leading MLLMs, Seed1.5-VL and Gemini-2.5-Flash, for medical applications. We assessed their performance on four visual medical tasks using VQA-RAD and ROCOv2 datasets. Our findings reveal that the improvement from activating the thinking mode remains marginal compared to the standard non-thinking mode for the majority of the tasks. Their performance on complex medical tasks such as open-ended VQA and medical image interpretation remains suboptimal, highlighting the need for domain-specific medical data and more advanced methods for medical knowledge integration.

cs.CL↗

arXiv2025

Understanding the Mixture-of-Experts with Nadaraya-Watson Kernel

Mixture-of-Experts (MoE) has become a cornerstone in recent state-of-the-art large language models (LLMs). Traditionally, MoE relies on $\mathrm{Softmax}$ as the router score function to aggregate expert output, a designed choice that has persisted from the earliest MoE models to modern LLMs, and is now widely regarded as standard practice. However, the necessity of using $\mathrm{Softmax}$ to project router weights into a probability simplex remains an unchallenged assumption rather than a principled design choice. In this work, we first revisit the classical Nadaraya-Watson regression and observe that MoE shares the same mathematical formulation as Nadaraya-Watson regression. Furthermore, we show that both feed-forward neural network (FFN) and MoE can be interpreted as a special case of Nadaraya-Watson regression, where the kernel function corresponds to the input neurons of the output layer. Motivated by these insights, we propose the \textbf{zero-additional-cost} Kernel Inspired Router with Normalization (KERN), an FFN-style router function, as an alternative to $\mathrm{Softmax}$. We demonstrate that this router generalizes both $\mathrm{Sigmoid}$- and $\mathrm{Softmax}$-based routers. \textbf{Based on empirical observations and established practices in FFN implementation, we recommend the use of $\mathrm{ReLU}$ activation and $\ell_2$-normalization in $\mathrm{KERN}$ router function.} Comprehensive experiments in MoE and LLM validate the effectiveness of the proposed FFN-style router function \methodNorm.

cs.CL↗

arXiv2025

Answer-Consistent Chain-of-thought Reinforcement Learning For Multi-modal Large Langauge Models

Recent advances in large language models (LLMs) have demonstrated that reinforcement learning with verifiable rewards (RLVR) can significantly enhance reasoning abilities by directly optimizing correctness, rather than relying solely on supervised imitation. This paradigm has been extended to multimodal LLMs for complex video and image understanding tasks. However, while outcome-driven RL improves answer accuracy, it can inadvertently decouple the reasoning chain from the final answer, leading to situations where models produce inconsistency between the reasoning trace and final answer. In our experiments on multiple-choice visual question-answering tasks, the standard GRPO method yields only 79.7\% consistency on MMVU between the reasoning steps and the chosen answers, indicating frequent mismatches between answers and reasoning. To this end, we propose Answer-Consistent Reinforcement Learning (ACRE) that modifies the GRPO algorithm with an auxiliary consistency check. After the model generates a chain of thought and an initial answer for a given question, we shuffle the answer options and prompt the model again with the same reasoning trace to predict a second answer. We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct; otherwise, a lower reward is assigned accordingly. This mechanism penalizes reasoning-answer misalignment and discourages the model from relying on spurious patterns, such as option ordering biases. We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2\% and 1.5\% improvement for Video Reasoning and Math Reasoning tasks over the GRPO baseline.

cs.CV↗

arXiv2025

Diagnosing Shoulder Disorders Using Multimodal Large Language Models and Consumer-Grade Cameras

Shoulder disorders, such as frozen shoulder (a.k.a., adhesive capsulitis), are common conditions affecting the health of people worldwide, and have a high incidence rate among the elderly and workers engaged in repetitive shoulder tasks. In regions with scarce medical resources, achieving early and accurate diagnosis poses significant challenges, and there is an urgent need for low-cost and easily scalable auxiliary diagnostic solutions. This research introduces videos captured by consumer-grade devices as the basis for diagnosis, reducing the cost for users. We focus on the innovative application of Multimodal Large Language Models (MLLMs) in the preliminary diagnosis of shoulder disorders and propose a Hybrid Motion Video Diagnosis framework (HMVDx). This framework divides the two tasks of action understanding and disease diagnosis, which are respectively completed by two MLLMs. In addition to traditional evaluation indicators, this work proposes a novel metric called Usability Index by the logical process of medical decision-making (action recognition, movement diagnosis, and final diagnosis). This index evaluates the effectiveness of MLLMs in the medical field from the perspective of the entire medical diagnostic pathway, revealing the potential value of low-cost MLLMs in medical applications for medical practitioners. In experimental comparisons, the accuracy of HMVDx in diagnosing shoulder joint injuries has increased by 79.6\% compared with direct video diagnosis, a significant technical contribution to future research on the application of MLLMs for video understanding in the medical field.

cs.CV↗

arXiv2025

UNComp: Can Matrix Entropy Uncover Sparsity? -- A Compressor Design from an Uncertainty-Aware Perspective

Deploying large language models (LLMs) for long-context inference remains challenging due to their substantial memory and computational demands. While techniques such as Key-Value (KV) cache compression are designed to reduce memory usage, they often neglect the structured sparsity inherent in the relationship between hidden states and their corresponding KV cache. In this work, we explore the role of uncertainty as a potential indicator of sparsity within LLMs. We propose UNComp, an uncertainty-aware framework that leverages truncated matrix entropy to identify areas of low information content, thereby revealing sparsity patterns that can be used for adaptive compression. Unlike traditional methods that apply uniform compression, UNComp dynamically adjusts its approach to compression, guided by uncertainty measures that reflect the importance of various model components. Our analysis shows that sparsity patterns, when derived from uncertainty estimates, can be exploited to reveal special long-range dependencies, such as retrieval heads and retrieval layers. This perspective not only enhances our understanding of how compression can be optimized but also provides new insights into the inherent sparsity of LLMs during long-context inference. By focusing on uncertainty to analyze the sparsity pattern in detail, UNComp reduces the KV cache size to 4.74% of the original, achieves a 6% prefill speedup, and improves throughput by 6.4x - not only delivering strong lossless compression performance, but also validating the effectiveness of the underlying theoretical tool. We release the code at https://github.com/menik1126/UNComp.

cs.CL↗

arXiv2025

SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model

With the widespread adoption of large language models (LLMs) in practical applications, selecting an appropriate model requires balancing not only performance but also operational cost. The emergence of reasoning-capable models has further widened the cost gap between "thinking" (high reasoning) and "non-thinking" (fast, low-cost) modes. In this work, we reveal that approximately 58% of medical questions can be accurately answered by the non-thinking mode alone, without requiring the high-cost reasoning process. This highlights a clear dichotomy in problem complexity and suggests that dynamically routing queries to the appropriate mode based on complexity could optimize accuracy, cost-efficiency, and overall user experience. Based on this, we further propose SynapseRoute, a machine learning-based dynamic routing framework that intelligently assigns input queries to either thinking or non-thinking modes. Experimental results on several medical datasets demonstrate that SynapseRoute not only improves overall accuracy (0.8390 vs. 0.8272) compared to the thinking mode alone but also reduces inference time by 36.8% and token consumption by 39.66%. Importantly, qualitative analysis indicates that over-reasoning on simpler queries can lead to unnecessary delays and even decreased accuracy, a pitfall avoided by our adaptive routing. Finally, this work further introduces the Accuracy-Inference-Token (AIT) index to comprehensively evaluate the trade-offs among accuracy, latency, and token cost.

cs.CL↗

arXiv2025

Logits-Based Finetuning

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research. Traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity. To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation. Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity. This ensures more reliable and effective training. We constructed a large-scale 1.2M logits dataset and trained a series of science-focused models. Experimental results demonstrate that our method achieves significant improvements, with accuracy gains of 18% on Mawps and 22.7% on TabMWP. Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%. Codes are available at https://github.com/dvlab-research/Logits-Based-Finetuning.

cs.LG↗

arXiv2025

Ming-Omni: A Unified Multimodal Model for Perception and Generation

We propose Ming-Omni, a unified multimodal model capable of processing images, text, audio, and video, while demonstrating strong proficiency in both speech and image generation. Ming-Omni employs dedicated encoders to extract tokens from different modalities, which are then processed by Ling, an MoE architecture equipped with newly proposed modality-specific routers. This design enables a single model to efficiently process and fuse multimodal inputs within a unified framework, thereby facilitating diverse tasks without requiring separate models, task-specific fine-tuning, or structural redesign. Importantly, Ming-Omni extends beyond conventional multimodal models by supporting audio and image generation. This is achieved through the integration of an advanced audio decoder for natural-sounding speech and Ming-Lite-Uni for high-quality image generation, which also allow the model to engage in context-aware chatting, perform text-to-speech conversion, and conduct versatile image editing. Our experimental results showcase Ming-Omni offers a powerful solution for unified perception and generation across all modalities. Notably, our proposed Ming-Omni is the first open-source model we are aware of to match GPT-4o in modality support, and we release all code and model weights to encourage further research and development in the community.

cs.AI↗

arXiv2025

ParallelComp: Parallel Long-Context Compressor for Length Extrapolation

Extrapolating ultra-long contexts (text length >128K) remains a major challenge for large language models (LLMs), as most training-free extrapolation methods are not only severely limited by memory bottlenecks, but also suffer from the attention sink, which restricts their scalability and effectiveness in practice. In this work, we propose ParallelComp, a parallel long-context compression method that effectively overcomes the memory bottleneck, enabling 8B-parameter LLMs to extrapolate from 8K to 128K tokens on a single A100 80GB GPU in a training-free setting. ParallelComp splits the input into chunks, dynamically evicting redundant chunks and irrelevant tokens, supported by a parallel KV cache eviction mechanism. Importantly, we present a systematic theoretical and empirical analysis of attention biases in parallel attention-including the attention sink, recency bias, and middle bias-and reveal that these biases exhibit distinctive patterns under ultra-long context settings. We further design a KV cache eviction technique to mitigate this phenomenon. Experimental results show that ParallelComp enables an 8B model (trained on 8K context) to achieve 91.17% of GPT-4's performance under ultra-long contexts, outperforming closed-source models such as Claude-2 and Kimi-Chat. We achieve a 1.76x improvement in chunk throughput, thereby achieving a 23.50x acceleration in the prefill stage with negligible performance loss and pave the way for scalable and robust ultra-long contexts extrapolation in LLMs. We release the code at https://github.com/menik1126/ParallelComp.

cs.CL↗

arXiv2025

Self-Adjust Softmax

The softmax function is crucial in Transformer attention, which normalizes each row of the attention scores with summation to one, achieving superior performances over other alternative functions. However, the softmax function can face a gradient vanishing issue when some elements of the attention scores approach extreme values, such as probabilities close to one or zero. In this paper, we propose Self-Adjust Softmax (SA-Softmax) to address this issue by modifying $softmax(x)$ to $x \cdot softmax(x)$ and its normalized variant $\frac{(x - min(x_{\min},0))}{max(0,x_{max})-min(x_{min},0)} \cdot softmax(x)$. We theoretically show that SA-Softmax provides enhanced gradient properties compared to the vanilla softmax function. Moreover, SA-Softmax Attention can be seamlessly integrated into existing Transformer models to their attention mechanisms with minor adjustments. We conducted experiments to evaluate the empirical performance of Transformer models using SA-Softmax compared to the vanilla softmax function. These experiments, involving models with up to 2.7 billion parameters, are conducted across diverse datasets, language tasks, and positional encoding methods.

cs.CL↗