arXiv subjects

Qinyu Chen

Publications and source records attributed to Qinyu Chen.

At least 19 recordsLinked to original sources

arXiv2026

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup, forcing them to inefficiently simulate retrieval through computation. To address this, we introduce conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embedding for O(1) lookup. By formulating the Sparsity Allocation problem, we uncover a U-shaped scaling law that optimizes the trade-off between neural computation (MoE) and static memory (Engram). Guided by this law, we scale Engram to 27B parameters, achieving superior performance over a strictly iso-parameter and iso-FLOPs MoE baseline. Most notably, while the memory module is expected to aid knowledge retrieval (e.g., MMLU +3.4; CMMLU +4.0), we observe even larger gains in general reasoning (e.g., BBH +5.0; ARC-Challenge +3.7) and code/math domains~(HumanEval +3.0; MATH +2.4). Mechanistic analyses reveal that Engram relieves the backbone's early layers from static reconstruction, effectively deepening the network for complex reasoning. Furthermore, by delegating local dependencies to lookups, it frees up attention capacity for global context, substantially boosting long-context retrieval (e.g., Multi-Query NIAH: 84.2 to 97.0). Finally, Engram establishes infrastructure-aware efficiency: its deterministic addressing enables runtime prefetching from host memory, incurring negligible overhead. We envision conditional memory as an indispensable modeling primitive for next-generation sparse models.

cs.CL↗

arXiv2026

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token dependencies. Furthermore, indiscriminately verifying these extended blocks wastes critical batch capacity on tokens with high rejection risks, severely degrading throughput in high-concurrency serving systems. We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification. To maintain draft quality, DSpark utilizes a semi-autoregressive architecture, coupling a parallel backbone with a lightweight sequential module, to introduce intra-block dependency modeling and mitigate suffix decay. To optimize system efficiency, DSpark employs confidence-scheduled verification, dynamically tailoring the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles. On offline benchmarks across diverse domains, DSpark substantially improves the accepted length over state-of-the-art autoregressive and parallel drafters. When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste. Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60 to 85 percent at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system.

cs.AI↗

arXiv2026

EveLoad: Cognitive Workload Recognition from Event-Based Eye Movements

Cognitive workload monitoring is important for adaptive rehabilitation and assistive interfaces, where task difficulty, pacing, and feedback should be adjusted according to the user's cognitive state to avoid overload and under-challenge. Emerging extended reality and robot-assisted rehabilitation environments provide controllable training tasks, but they require unobtrusive sensing methods that can capture rapid ocular dynamics during interaction. Existing eye-movement-based cognitive workload recognition methods mainly rely on frame-based eye trackers, which often suffer from limited temporal resolution and degraded robustness under rapid eye movements. In contrast, event cameras provide microsecond-level temporal resolution, high dynamic range and low latency, making them suitable for capturing fine-grained ocular dynamics. Many previous studies rely on free-viewing or similar paradigms, where gaze locations can vary across tasks. As a result, models may learn associations between gaze-location distributions and cognitive workload, rather than workload-related eye movement characteristics themselves. In this work, we introduce EveLoad, which, to the best of our knowledge, is the first event-based eye-movement dataset with graded cognitive workload annotations, collected from 20 healthy participants under spatially constrained and task-driven conditions using a controlled N-back-guided fixation paradigm. Based on this dataset, we establish a benchmark for cognitive workload recognition with six workload levels and propose a learning framework that encodes spatiotemporal event representations. Experimental results show that our approach achieves an average subject-specific accuracy of 96.36% and 96.13% under mixed random split evaluation. These results suggest that event-based eye movements may provide a useful sensing pathway for future workload-aware rehabilitation.

cs.LG↗

arXiv2026

When NPUs Are Not Always Faster: A Stage-Level Analysis of Mobile LLM Inference

Deploying large language models (LLMs) on mobile devices increasingly relies on heterogeneous execution, yet no prior study has systematically characterized NPU effectiveness at the operator and pipeline level. We present the first stage-aware, multi-level benchmarking study of mobile LLM inference on a CPU-NPU heterogeneous SoC. We introduce an OPMASK-based controlled pipeline decomposition methodology that isolates communication, quantization, and computation overheads within the NPU execution path. Our results reveal a counter-intuitive stage-level performance reversal: CPUs outperform NPUs in the compute-intensive Prefill stage (up to 1.6x), while NPUs provide only limited acceleration in the memory-bound Decode stage (1.05-1.2x). We further show that scheduling overhead and cross-backend fallback reduce the practical benefits of NPU offloading. For the energy trend, increasing NPU offloading leads to higher energy consumption (up to 51%). Based on these findings, we derive design guidelines for NPU architects targeting on-device LLM inference.

cs.AR↗

arXiv2026

Two-Stage Adaptation for Non-Normative Speech Recognition: Revisiting Speaker-Independent Initialization for Personalization

Personalizing automatic speech recognition (ASR) systems for non-normative speech, such as dysarthric and aphasic speech, is challenging. While speaker-specific fine-tuning (SS-FT) is widely used, it is typically initialized directly from a generic pre-trained model. Whether speaker-independent adaptation provides a stronger initialization prior under such mismatch remains unclear. In this work, we propose a two-stage adaptation framework consisting of speaker-independent fine-tuning (SI-FT) on multi-speaker non-normative data followed by SS-FT, and evaluate it through a controlled comparison with direct SS-FT under identical per-speaker conditions. Experiments on AphasiaBank and UA-Speech with Whisper-Large-v3 and Qwen3-ASR, alongside evaluation on typical-speech datasets TED-LIUM v3 and FLEURS, show that two-stage adaptation consistently improves personalization while maintaining manageable out-of-domain (OOD) trade-offs.

cs.SD↗

arXiv2026

Vib2ECG: A Paired Chest-Lead SCG-ECG Dataset and Benchmark for ECG Reconstruction

Twelve-lead electrocardiography (ECG) is essential for cardiovascular diagnosis, but its long-term acquisition in daily life is constrained by complex and costly hardware. Recent efforts have explored reconstructing ECG from low-cost cardiac vibrational signals such as seismocardiography (SCG), however, due to the lack of a dataset, current methods are limited to limb leads, while clinical diagnosis requires multi-lead ECG, including chest leads. In this work, we propose Vib2ECG, the first paired, multi-channel electro-mechanical cardiac signal dataset, which includes complete twelve-lead ECGs and vibrational signals acquired by inertial measurement units (IMUs) at six chest-lead positions from 17 subjects. Based on this dataset, we also provide a benchmark. Experimental results demonstrate the feasibility of reconstructing electrical cardiac signals at variable locations from vibrational signals using a lightweight 364 K-parameter U-Net. Furthermore, we observe a hallucination phenomenon in the model, where ECG waveforms are generated in regions where no corresponding electrical activity is present. We analyze the causes of this phenomenon and propose potential directions for mitigation. This study demonstrates the feasibility of mobile-device-friendly ECG monitoring through chest-lead ECG prediction from low-cost vibrational signals acquired using IMU sensors. It expands the application of cardiac vibrational signals and provides new insights into the spatial relationship between cardiac electrical and mechanical activities with spatial location variation.

cs.LG↗

arXiv2026

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

General reasoning represents a long-standing and formidable challenge in artificial intelligence. Recent breakthroughs, exemplified by large language models (LLMs) and chain-of-thought prompting, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent upon extensive human-annotated demonstrations, and models' capabilities are still insufficient for more complex problems. Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation. Consequently, the trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields, surpassing its counterparts trained via conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models.

cs.CL↗

arXiv2025

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments.

cs.CL↗

arXiv2025

JaneEye: A 12-nm 2K-FPS 18.9-$μ$J/Frame Event-based Eye Tracking Accelerator

Eye tracking has become a key technology for gaze-based interactions in Extended Reality (XR). However, conventional frame-based eye-tracking systems often fall short of XR's stringent requirements for high accuracy, low latency, and energy efficiency. Event cameras present a compelling alternative, offering ultra-high temporal resolution and low power consumption. In this paper, we present JaneEye, an energy-efficient event-based eye-tracking hardware accelerator designed specifically for wearable devices, leveraging sparse, high-temporal-resolution event data. We introduce an ultra-lightweight neural network architecture featuring a novel ConvJANET layer, which simplifies the traditional ConvLSTM by retaining only the forget gate, thereby halving computational complexity without sacrificing temporal modeling capability. Our proposed model achieves high accuracy with a pixel error of 2.45 on the 3ET+ dataset, using only 17.6K parameters, with up to 1250 Hz event frame rate. To further enhance hardware efficiency, we employ custom linear approximations of activation functions (hardsigmoid and hardtanh) and fixed-point quantization. Through software-hardware co-design, our 12-nm ASIC implementation operates at 400 MHz, delivering an end-to-end latency of 0.5 ms (equivalent to 2000 Frames Per Second (FPS)) at an energy efficiency of 18.9 $μ$J/frame. JaneEye sets a new benchmark in low-power, high-performance eye-tracking solutions suitable for integration into next-generation XR wearables.

eess.SP↗

arXiv2025

Eye Gaze Tells You Where to Compute: Gaze-Driven Efficient VLMs

Vision-Language Models (VLMs) deliver impressive performance in understanding visual content with language instructions. However, redundancy in vision tokens results in the degenerated inference efficiency of VLMs, which hinders real-time use on edge consumer devices such as AR/VR devices. Existing efficiency methods commonly prune visual tokens using learned saliency, sparse attention schedules, or controller policies, but they often require architectural modification or access to intermediate activations. These pipelines add inference-time modules that increase compute and memory and often lead to an accuracy trade-off. Moreover, they also suffer from misalignment between the prompts and the region of interest in the images. Without human guidance, the model may focus on the wrong regions and miss small, high-frequency details when prompts or scenes change. In this paper, we propose GazeVLM, a training-free framework that uses the human eye gaze as a natural supervisory signal to allocate computation where it matters. By extracting gaze-driven regions of interest (ROIs) and optionally combining them with a low-resolution global view, GazeVLM mimics fovea-periphery perception to cut redundant visual tokens while preserving task-relevant details. We evaluate the visual question answering tasks on Qwen2.5-VL-3B/7B on the VOILA-COCO benchmark with human gaze. Quality of the answer is assessed by GPT-4o pairwise judging and a weighted score over coverage, accuracy, details, and fluency. Efficiency is measured by token counts and FLOPs. GazeVLM reduces visual tokens by up to 93.1%, total tokens by up to 59.6%, and FLOPs by 50%, while keeping better answer quality relative to full-resolution baselines. Our results show that aligning model computation with human gaze offers a simple, plug-and-play path toward efficient VLM inference on consumer devices.

cs.CV↗

arXiv2025

AS-ASR: A Lightweight Framework for Aphasia-Specific Automatic Speech Recognition

This paper proposes AS-ASR, a lightweight aphasia-specific speech recognition framework based on Whisper-tiny, tailored for low-resource deployment on edge devices. Our approach introduces a hybrid training strategy that systematically combines standard and aphasic speech at varying ratios, enabling robust generalization, and a GPT-4-based reference enhancement method that refines noisy aphasic transcripts, improving supervision quality. We conduct extensive experiments across multiple data mixing configurations and evaluation settings. Results show that our fine-tuned model significantly outperforms the zero-shot baseline, reducing WER on aphasic speech by over 30% while preserving performance on standard speech. The proposed framework offers a scalable, efficient solution for real-world disordered speech recognition.

eess.AS↗

arXiv2025

EvHand-FPV: Efficient Event-Based 3D Hand Tracking from First-Person View

Hand tracking holds great promise for intuitive interaction paradigms, but frame-based methods often struggle to meet the requirements of accuracy, low latency, and energy efficiency, especially in resource-constrained settings such as Extended Reality (XR) devices. Event cameras provide $μ$s-level temporal resolution at mW-level power by asynchronously sensing brightness changes. In this work, we present EvHand-FPV, a lightweight framework for egocentric First-Person-View 3D hand tracking from a single event camera. We construct an event-based FPV dataset that couples synthetic training data with 3D labels and real event data with 2D labels for evaluation to address the scarcity of egocentric benchmarks. EvHand-FPV also introduces a wrist-based region of interest (ROI) that localizes the hand region via geometric cues, combined with an end-to-end mapping strategy that embeds ROI offsets into the network to reduce computation without explicit reconstruction, and a multi-task learning strategy with an auxiliary geometric feature head that improves representations without test-time overhead. On our real FPV test set, EvHand-FPV improves 2D-AUCp from 0.77 to 0.85 while reducing parameters from 11.2M to 1.2M by 89% and FLOPs per inference from 1.648G to 0.185G by 89%. It also maintains a competitive 3D-AUCp of 0.84 on synthetic data. These results demonstrate accurate and efficient egocentric event-based hand tracking suitable for on-device XR applications. The dataset and code are available at https://github.com/zen5x5/EvHand-FPV.

cs.CV↗

arXiv2025

DeltaLLM: A Training-Free Framework Exploiting Temporal Sparsity for Efficient Edge LLM Inference

Deploying Large Language Models (LLMs) on edge devices remains challenging due to their quadratically increasing computations with the sequence length. Existing studies for dynamic attention pruning are designed for hardware with massively parallel computation capabilities, such as GPUs or TPUs, and aim at long context lengths (e.g., 64K), making them unsuitable for edge scenarios. We present DeltaLLM, a training-free framework that exploits temporal sparsity in attention patterns to enable efficient LLM inference across both the prefilling and decoding stages, on resource-constrained edge devices. DeltaLLM introduces an accuracy- and memory-aware delta matrix construction strategy that introduces temporal sparsity, and a context-aware hybrid attention mechanism that combines full attention in a local context window with delta approximation outside it to increase accuracy. We evaluate our framework on the edge-device-friendly BitNet-b1.58-2B-4T model and Llama3.2-1B-Instruct model across diverse language tasks. The results show that on BitNet, our framework increases the attention sparsity from 0% to 60% during the prefilling stage with slight accuracy improvement on the WG task, and 0% to 57% across both the prefilling and decoding stages, with even higher F1 score from 29.63 to 30.97 on SQuAD-v2 task. On the Llama model, it can also achieve up to 60% sparsity during the prefilling stage and around 57% across both stages with negligible accuracy drop. These results demonstrate that DeltaLLM offers a promising solution for efficient edge deployment, requiring no fine-tuning and seamlessly integrating with existing inference pipelines.

cs.AI↗

arXiv2025

SHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection

Electroencephalography (EEG)-based auditory attention detection (AAD) offers a non-invasive way to enhance hearing aids, but conventional methods rely on too many electrodes, limiting wearability and comfort. This paper presents SHAP-AAD, a two-stage framework that combines DeepSHAP-based channel selection with a lightweight temporal convolutional network (TCN) for efficient AAD using fewer channels.DeepSHAP, an explainable AI technique, is applied to a Convolutional Neural Network (CNN) trained on topographic alpha-power maps to rank channel importance, and the top-k EEG channels are used to train a compact TCN. Experiments on the DTU dataset show that using 32 channels yields comparable accuracy to the full 64-channel setup (79.21% vs. 81.06%) on average. In some cases, even 8 channels can deliver satisfactory accuracy. These results demonstrate the effectiveness of SHAP-AAD in reducing complexity while preserving high detection performance.

eess.SP↗

arXiv2025

TCN-DPD: Parameter-Efficient Temporal Convolutional Networks for Wideband Digital Predistortion

Digital predistortion (DPD) is essential for mitigating nonlinearity in RF power amplifiers, particularly for wideband applications. This paper presents TCN-DPD, a parameter-efficient architecture based on temporal convolutional networks, integrating noncausal dilated convolutions with optimized activation functions. Evaluated on the OpenDPD framework with the DPA_200MHz dataset, TCN-DPD achieves simulated ACPRs of -51.58/-49.26 dBc (L/R), EVM of -47.52 dB, and NMSE of -44.61 dB with 500 parameters and maintains superior linearization than prior models down to 200 parameters, making it promising for efficient wideband PA linearization.

eess.SP↗

arXiv2025

DeltaDPD: Exploiting Dynamic Temporal Sparsity in Recurrent Neural Networks for Energy-Efficient Wideband Digital Predistortion

Digital Predistortion (DPD) is a popular technique to enhance signal quality in wideband RF power amplifiers (PAs). With increasing bandwidth and data rates, DPD faces significant energy consumption challenges during deployment, contrasting with its efficiency goals. State-of-the-art DPD models rely on recurrent neural networks (RNN), whose computational complexity hinders system efficiency. This paper introduces DeltaDPD, exploring the dynamic temporal sparsity of input signals and neuronal hidden states in RNNs for energy-efficient DPD, reducing arithmetic operations and memory accesses while preserving satisfactory linearization performance. Applying a TM3.1a 200MHz-BW 256-QAM OFDM signal to a 3.5 GHz GaN Doherty RF PA, DeltaDPD achieves -50.03 dBc in Adjacent Channel Power Ratio (ACPR), -37.22 dB in Normalized Mean Square Error (NMSE) and -38.52 dBc in Error Vector Magnitude (EVM) with 52% temporal sparsity, leading to a 1.8X reduction in estimated inference power. The DeltaDPD code will be released after formal publication at https://www.opendpd.com.

eess.SP↗

arXiv2025

Event-Based Eye Tracking. 2025 Event-based Vision Workshop

This survey serves as a review for the 2025 Event-Based Eye Tracking Challenge organized as part of the 2025 CVPR event-based vision workshop. This challenge focuses on the task of predicting the pupil center by processing event camera recorded eye movement. We review and summarize the innovative methods from teams rank the top in the challenge to advance future event-based eye tracking research. In each method, accuracy, model size, and number of operations are reported. In this survey, we also discuss event-based eye tracking from the perspective of hardware design.

cs.CV↗

arXiv2025

DeepSeek-V3 Technical Report

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

cs.CL↗