arXiv subjects

Jiaming Cheng

Publications and source records attributed to Jiaming Cheng.

At least 19 recordsLinked to original sources

arXiv2026

Large Models for Small Devices: Recent Advances and Empirical Analysis of Edge AI Deployment

Running large AI models on resource-constrained edge devices requires model compression to reduce model size and computation. What compresses well, however, need not deploy well. We survey dozens of recent works that report compression results on real hardware and extract practical deployment guidelines from them. Following these guidelines, we deploy compact language and image models on GPU, CPU, and Raspberry Pi platforms across question answering and image segmentation. No single technique wins across tasks. For question answering, Qwen3.5 0.8B reaches 93.85 SQuAD F1 and 92 EM under Q5_K_M GGUF quantization, while structured pruning at the same precision costs 16 F1 at a 1% ratio. For segmentation, the ranking reverses: default quantization leaves parameters and MACs unchanged, whereas pruning cuts model size by nearly 80% at near-constant mIoU. Pruning can even inflate the deployed artifact by 21-49% by breaking k-quant super-block alignment; combined with longer, less format-compliant outputs, this raises Raspberry Pi latency up to 3.4x. Compression can also manufacture the appearance of competence rather than destroy it visibly: one LoRA-recovered variant stays fully parseable and holds 71% strict BoolQ accuracy while sending 97 of 100 predictions to a single class, at 52.6% balanced accuracy. We explain these effects through neural-flow graph analysis and prefill-decode-level latency decomposition, and condense them into task-specific deployment research directions. The right technique depends on the task, the model, and the hardware. Our experiment code and artifacts are open-sourced at https://github.com/Arnavvvkumar/deployment

cs.AI↗

arXiv2026

When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs

Multi-agent LLM systems relay key-value caches instead of text and credit their gains to exchanged "latent thoughts". That credit is a claim about which example's cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender's private information. Where it does, the battery reads ceiling: 100% against 23-25% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system's reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage. A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS's native relay), partial (KVComm's layer subset), and no detected example-specific transfer (C2C's released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.

cs.CR↗

arXiv2026

Robust KV Cache Management for LLM Serving under Output Token Length Uncertainty

KV cache memory is a primary bottleneck in modern LLM serving systems deployed on GPU clusters. A fundamental challenge is that the KV cache must be reserved upon request arrival, while the output token length remains unknown until generation completes. Under-reservation triggers preemption -- forcing termination and recomputation of requests and incurring significant overhead -- whereas over-reservation wastes memory and reduces throughput. This creates a central trade-off between memory efficiency and preemption risk. We present a robust KV cache management framework for LLM serving that jointly optimizes GPU parallelism configuration, KV cache reservation per request class, request routing across heterogeneous serving groups, and prefix caching for shared prompts. The framework incorporates latency SLO constraints and captures the interaction between memory allocation, throughput, and queueing delay. To address output token length uncertainty and workload distribution shift, we develop a Wasserstein distributionally robust optimization (DRO) formulation together with a scalable block coordinate descent algorithm for the resulting mixed-integer problem. Our analysis reveals a critical fractile structure that automatically adapts reservation quantiles to different preemption and memory cost regimes without manual tuning. Trace-driven evaluation on production LLM workloads, including BurstGPT, Azure, and ShareGPT traces, demonstrates up to 56\% lower cost than fixed-quantile reservation baselines while maintaining competitive P99 latency, goodput, and SLO violation rates across diverse operating regimes.

cs.NI↗

arXiv2026

ComHymba: Low-Complexity Domain-Informed Foundation Model for Wireless Communications

Wireless foundation models are a promising route to unify channel reconstruction, sensing, and beam management in future wireless communication systems, but existing designs often inherit LLM-style Transformers with quadratic token complexity and weak integration of propagation priors. This paper proposes ComHymba, a domain-informed wireless foundation model built on an asymmetric masked autoencoder for large-scale self-supervised pre-training on Channel State Information (CSI). ComHymba introduces (i) 3D spatio-temporal-frequency patchification with rotary positional embedding, (ii) domain-informed masking strategies that emulate realistic CSI sparsity and fading patterns, and (iii) a decoupled amplitude--phase weighted objective tailored to channel statistics. Architecturally, we employ Hymba blocks that fuse windowed self-attention with state space models (SSMs), enabling linear-time modeling with respect to the overall channel input size. Experiments on eight downstream tasks spanning channel state information reconstruction, environmental sensing, and beam management show consistent accuracy gains over strong task-specific baselines, together with up to a $3.3\times$ inference speedup versus Transformer backbones. Overall, ComHymba provides a scalable and efficient backbone for AI-native physical-layer intelligence.

eess.SP↗

arXiv2026

Scalable Joint Resource Allocation for SLO-Constrained LLM Inference in Heterogeneous GPU Clouds

Serving large language model (LLM) inference in cloud environments requires jointly optimizing model selection, GPU provisioning, parallelism configuration, and workload routing under latency, accuracy, memory, and budget constraints. While mixed-integer linear programming (MILP) can model this problem, its computational cost limits frequent re-optimization under demand variability. Existing heuristics often optimize individual components separately and may become infeasible when system-wide constraints are enforced. This paper presents a scalable framework for SLO-constrained LLM inference. We formulate the problem as an MILP with a two-phase delay model capturing both prefill and autoregressive decoding under tensor and pipeline parallelism. To solve it efficiently, we develop two constraint-aware heuristics: a Greedy Heuristic (GH) and an Adaptive Greedy Heuristic (AGH). AGH extends GH through multi-start construction, local search, and GPU consolidation. Both methods maintain feasibility through parallelism-aware filtering, cost-based ranking, and adaptive parallelism scaling. Experiments based on the Azure LLM Inference Trace show that GH generates feasible solutions within one second, while AGH achieves near-optimal performance within three seconds and scales to large instances where exact solvers fail to converge. Under out-of-sample stress with up to 1.5x delay and accuracy inflation, AGH degrades gracefully through provisioned headroom, yielding substantially lower cost and SLO violations than cost-minimal MILP solutions. Across synthetic and real Azure workloads, AGH maintains SLO compliance at significantly lower cost than exact MILP solutions. These results demonstrate that high-quality allocations provide substantial robustness to demand variability while enabling rapid adaptation to workload changes.

cs.LG↗

arXiv2025

Adaptive End-to-End Transceiver Design for NextG Pilot-Free and CP-Free Wireless Systems

The advent of artificial intelligence (AI)-native wireless communication is fundamentally reshaping the design paradigm of next-generation (NextG) systems, where intelligent air interfaces are expected to operate adaptively and efficiently in highly dynamic environments. Conventional orthogonal frequency division multiplexing (OFDM) systems rely heavily on pilots and the cyclic prefix (CP), resulting in significant overhead and reduced spectral efficiency. To address these limitations, we propose an adaptive end-to-end (E2E) transceiver architecture tailored for pilot-free and CP-free wireless systems. The architecture combines AI-driven constellation shaping and a neural receiver through joint training. To enhance robustness against mismatched or time-varying channel conditions, we introduce a lightweight channel adapter (CA) module, which enables rapid adaptation with minimal computational overhead by updating only the CA parameters. Additionally, we present a framework that is scalable to multiple modulation orders within a unified model, significantly reducing model storage requirements. Moreover, to tackle the high peak-to-average power ratio (PAPR) inherent to OFDM, we incorporate constrained E2E training, achieving compliance with PAPR targets without additional transmission overhead. Extensive simulations demonstrate that the proposed framework delivers superior bit error rate (BER), throughput, and resilience across diverse channel scenarios, highlighting its potential for AI-native NextG.

eess.SP↗

arXiv2025

Affine Modulation-based Audiogram Fusion Network for Joint Noise Reduction and Hearing Loss Compensation

Hearing aids (HAs) are widely used to provide personalized speech enhancement (PSE) services, improving the quality of life for individuals with hearing loss. However, HA performance significantly declines in noisy environments as it treats noise reduction (NR) and hearing loss compensation (HLC) as separate tasks. This separation leads to a lack of systematic optimization, overlooking the interactions between these two critical tasks, and increases the system complexity. To address these challenges, we propose a novel audiogram fusion network, named AFN-HearNet, which simultaneously tackles the NR and HLC tasks by fusing cross-domain audiogram and spectrum features. We propose an audiogram-specific encoder that transforms the sparse audiogram profile into a deep representation, addressing the alignment problem of cross-domain features prior to fusion. To incorporate the interactions between NR and HLC tasks, we propose the affine modulation-based audiogram fusion frequency-temporal Conformer that adaptively fuses these two features into a unified deep representation for speech reconstruction. Furthermore, we introduce a voice activity detection auxiliary training task to embed speech and non-speech patterns into the unified deep representation implicitly. We conduct comprehensive experiments across multiple datasets to validate the effectiveness of each proposed module. The results indicate that the AFN-HearNet significantly outperforms state-of-the-art in-context fusion joint models regarding key metrics such as HASQI and PESQ, achieving a considerable trade-off between performance and efficiency. The source code and data will be released at https://github.com/deepnetni/AFN-HearNet.

eess.AS↗

arXiv2025

Green-LLM: Optimal Workload Allocation for Environmentally-Aware Distributed Inference

This paper investigates the optimal allocation of large language model (LLM) inference workloads across heterogeneous edge data centers over time. Each data center features on-site renewable generation and faces dynamic electricity prices and spatiotemporal variability in renewable availability. We propose Green-LLM, a lexicographic multi-objective optimization framework that addresses this challenge without requiring manual weight tuning. The proposed model incorporates real-world constraints, including token-dependent processing delay and energy consumption, heterogeneous hardware capabilities, dynamic renewable generation, and spatiotemporal variations in electricity prices and carbon intensity. Unlike existing approaches that optimize individual environmental metrics in isolation, Green-LLM jointly minimizes operational cost, carbon emissions, and delay penalty while enforcing water consumption constraints to ensure both sustainability and quality-of-service requirements. Numerical results demonstrate that Green-LLM achieves significant reductions in carbon emissions and water consumption while maintaining operational costs within 3% of the minimum and ensuring sub-2-second response latency. These findings show that sustainable LLM inference can be achieved without sacrificing service quality or economic efficiency.

cs.NI↗

arXiv2025

Leveraging Local and Global Knowledge Integration with Time-Frequency Calibrated Distillation for Speech Enhancement

In this paper, we propose an intra-set and inter-set recursive fusion framework with time-frequency calibrated knowledge distillation (I$^2$SRF-TFCKD) for SE. Different from previous distillation strategies for SE, the proposed framework fully exploits the time-frequency differential information of speech while facilitating both local information focusing and global knowledge circulation. Firstly, we construct a collaborative distillation paradigm for intra-set and inter-set correlations. Within a correlated set, multi-layer teacher-student features are pairwise matched for calibrated distillation. Subsequently, we generate representative features from each correlated set through recursive fusion to form the fused feature set that enables inter-set knowledge interaction. Secondly, we propose a multi-layer interactive distillation based on dual-stream time-frequency cross-calibration, which calculates the teacher-student similarity calibration weights in the time and frequency domains respectively and performs cross-weighting, thus enabling refined allocation of distillation contributions across different layers according to speech characteristics. The proposed distillation strategy is applied to the dual-path dilated convolutional recurrent network (DPDCRN) that ranked first in the SE track of the L3DAS23 challenge. To evaluate the effectiveness of I$^2$SRF-TFCKD, we conduct experiments on both single-channel and multi-channel SE datasets. Objective evaluations demonstrate that the proposed KD strategy consistently and effectively improves the performance of the low-complexity student model and outperforms other distillation schemes.

cs.SD↗

arXiv2024

Robust Dynamic Edge Service Placement Under Spatio-Temporal Correlated Demand Uncertainty

Edge computing allows Service Providers (SPs) to enhance user experience by placing their services closer to the network edge. Determining the optimal provisioning of edge resources to meet the varying and uncertain demand cost-effectively is a critical task for SPs. This paper introduces a novel two-stage multi-period robust model for edge service placement and workload allocation, aiming to minimize the SP's operating costs while ensuring service quality. The salient feature of this model lies in its ability to enable SPs to utilize dynamic service placement and leverage spatio-temporal correlation in demand uncertainties to mitigate the inherent conservatism of robust solutions. In our model, resource reservation is optimized in the initial stage, preemptively, before the actual demand is disclosed, whereas dynamic service placement and workload allocation are determined in the subsequent stage, following the revelation of uncertainties. To address the challenges posed by integer recourse variables in the second stage of the resulting tri-level adjustable robust optimization problem, we propose a novel iterative, decomposition-based approach, ensuring finite convergence to an exact optimal solution. Extensive numerical results are provided to demonstrate the efficacy of the proposed model and approach.

math.OC↗

arXiv2024

Delay-Aware Robust Edge Network Hardening Under Decision-Dependent Uncertainty

Edge computing promises to offer low-latency and ubiquitous computation to numerous devices at the network edge. For delay-sensitive applications, link delays can have a direct impact on service quality. These delays can fluctuate drastically over time due to various factors such as network congestion, changing traffic conditions, cyberattacks, component failures, and natural disasters. Thus, it is crucial to efficiently harden the edge network to mitigate link delay variation as well as ensure a stable and improved user experience. To this end, we propose a novel robust model for optimal edge network hardening, considering the link delay uncertainty. Departing from the existing literature that treats uncertainties as exogenous, our model incorporates an endogenous uncertainty set to properly capture the impact of hardening and workload allocation decisions on link delays. However, the endogenous set introduces additional complexity to the problem due to the interdependence between decisions and uncertainties. We present two efficient methods to transform the problem into a solvable form. Extensive numerical results are shown to demonstrate the effectiveness of the proposed approach.

cs.NI↗

arXiv2024

Deep Learning for CSI Feedback: One-Sided Model and Joint Multi-Module Learning Perspectives

The use of deep learning (DL) for channel state information (CSI) feedback has garnered widespread attention across academia and industry. The mainstream DL architectures, e.g., CsiNet, deploy DL models on the base station (BS) side and the user equipment (UE) side, which are highly coupled and need to be trained jointly. However, two-sided DL models require collaborations between different network vendors and UE vendors, which entails considerable challenges in order to achieve consensus, e.g., model maintenance and responsibility. Furthermore, DL-based CSI feedback design invokes DL to reduce only the CSI feedback error, whereas jointly optimizing several modules at the transceivers would provide more significant gains. This article presents DL-based CSI feedback from the perspectives of one-sided model and joint multi-module learning. We herein introduce various novel one-sided CSI feedback architectures. In particular, the recently proposed CSI-PPPNet provides a one-sided one-for-all framework, which allows a DL model to deal with arbitrary CSI compression ratios. We review different joint multi-module learning methods, where the CSI feedback module is learned jointly with other modules including channel coding, channel estimation, pilot design and precoding design. Finally, future directions and challenges for DL-based CSI feedback are discussed, from the perspectives of inherent limitations of artificial intelligence (AI) and practical deployment issues.

eess.SP↗

arXiv2024

Two-Stage Distributionally Robust Edge Node Placement Under Endogenous Demand Uncertainty

Edge computing (EC) promises to deliver low-latency and ubiquitous computation to numerous devices at the network edge. This paper aims to jointly optimize edge node (EN) placement and resource allocation for an EC platform, considering demand uncertainty. Diverging from existing approaches treating uncertainties as exogenous, we propose a novel two-stage decision-dependent distributionally robust optimization (DRO) framework to effectively capture the interdependence between EN placement decisions and uncertain demands. The first stage involves making EN placement decisions, while the second stage optimizes resource allocation after uncertainty revelation. We present an exact mixed-integer linear program reformulation for solving the underlying ``min-max-min" two-stage model. We further introduce a valid inequality method to enhance computational efficiency, especially for large-scale networks. Extensive numerical experiments demonstrate the benefits of considering endogenous uncertainties and the advantages of the proposed model and approach.

math.OC↗

arXiv2024

Swin Transformer-Based CSI Feedback for Massive MIMO

For massive multiple-input multiple-output systems in the frequency division duplex (FDD) mode, accurate downlink channel state information (CSI) is required at the base station (BS). However, the increasing number of transmit antennas aggravates the feedback overhead of CSI. Recently, deep learning (DL) has shown considerable potential to reduce CSI feedback overhead. In this paper, we propose a Swin Transformer-based autoencoder network called SwinCFNet for the CSI feedback task. In particular, the proposed method can effectively capture the long-range dependence information of CSI. Moreover, we explore the impact of the number of Swin Transformer blocks and the dimension of feature channels on the performance of SwinCFNet. Experimental results show that SwinCFNet significantly outperforms other DL-based methods with comparable model sizes, especially for the outdoor scenario.

cs.IT↗

arXiv2023

Optimal Workload Allocation for Distributed Edge Clouds With Renewable Energy and Battery Storage

This paper studies an optimal workload allocation problem for a network of renewable energy-powered edge clouds that serve users located across various geographical areas. Specifically, each edge cloud is furnished with both an on-site renewable energy generation unit and a battery storage unit. Due to the discrepancy in electricity pricing and the diverse temporal-spatial characteristics of renewable energy generation, how to optimally allocate workload to different edge clouds to minimize the total operating cost while maximizing renewable energy utilization is a crucial and challenging problem. To this end, we introduce and formulate an optimization-based framework designed for Edge Service Providers (ESPs) with the overarching goal of simultaneously reducing energy costs and environmental impacts through the integration of renewable energy sources and battery storage systems, all while maintaining essential quality-of-service standards. Numerical results demonstrate the effectiveness of the proposed model and solution in maintaining service quality as well as reducing operational costs and emissions. Furthermore, the impacts of renewable energy generation and battery storage on optimal system operations are rigorously analyzed.

math.OC↗

arXiv2023

CrowdCache: A Decentralized Game-Theoretic Framework for Mobile Edge Content Sharing

Mobile edge computing (MEC) is a promising solution for enhancing the user experience, minimizing content delivery expenses, and reducing backhaul traffic. In this paper, we propose a novel privacy-preserving decentralized game-theoretic framework for resource crowdsourcing in MEC. Our framework models the interactions between a content provider (CP) and multiple mobile edge device users (MEDs) as a non-cooperative game, in which MEDs offer idle storage resources for content caching in exchange for rewards. We introduce efficient decentralized gradient play algorithms for Nash equilibrium (NE) computation by exchanging local information among neighboring MEDs only, thus preventing attackers from learning users' private information. The key challenge in designing such algorithms is that communication among MEDs is not fixed and is facilitated by a sequence of undirected time-varying graphs. Our approach achieves linear convergence to the NE without imposing any assumptions on the values of parameters in the local objective functions, such as requiring strong monotonicity to be stronger than its dependence on other MEDs' actions, which is commonly required in existing literature when the graph is directed time-varying. Extensive simulations demonstrate the effectiveness of our approach in achieving efficient resource outsourcing decisions while preserving the privacy of the edge devices.

cs.GT↗

arXiv2023

A Bandit Approach to Online Pricing for Heterogeneous Edge Resource Allocation

Edge Computing (EC) offers a superior user experience by positioning cloud resources in close proximity to end users. The challenge of allocating edge resources efficiently while maximizing profit for the EC platform remains a sophisticated problem, especially with the added complexity of the online arrival of resource requests. To address this challenge, we propose to cast the problem as a multi-armed bandit problem and develop two novel online pricing mechanisms, the Kullback-Leibler Upper Confidence Bound (KL-UCB) algorithm and the Min-Max Optimal algorithm, for heterogeneous edge resource allocation. These mechanisms operate in real-time and do not require prior knowledge of demand distribution, which can be difficult to obtain in practice. The proposed posted pricing schemes allow users to select and pay for their preferred resources, with the platform dynamically adjusting resource prices based on observed historical data. Numerical results show the advantages of the proposed mechanisms compared to several benchmark schemes derived from traditional bandit algorithms, including the Epsilon-Greedy, basic UCB, and Thompson Sampling algorithms.

cs.LG↗

arXiv2023

A Fairness-Aware Attacker-Defender Model for Optimal Edge Network Operation and Protection

While various aspects of edge computing (EC) have been studied extensively, the current literature has overlooked the robust edge network operations and planning problem. To this end, this letter proposes a novel fairness-aware attacker-defender model for optimal edge network operation and hardening against possible attacks and disruptions. The proposed model helps EC platforms identify the set of most critical nodes to be protected to mitigate the impact of failures on system performance. Numerical results show that the proposed solution can not only ensure good service quality but also maintain fairness among different areas during disruptions.

math.OC↗