In briefIndoor radio channel measurements at upper 6 GHz with directional and quasi-omnidirectional antennas
SummaryThis paper reports wideband indoor channel measurements in the upper 6 GHz band using directional and quasi-omnidirectional antenna pairs in a hall and a corridor. It fits path loss models for line-of-sight and non-line-of-sight conditions and characterizes shadow fading, RMS delay spread, and equivalent K factors. Directional links show smaller line-of-sight residuals and delay spread and higher equivalent K factors.
In briefMinimizing estimation error in wireless sensing by optimizing movable antenna array geometry
SummaryThis paper studies movable antenna (MA) arrays that adapt their geometry to reduce angle-of-arrival (AoA) estimation mean squared error (MSE) across arbitrary SNR levels. The authors derive a deterministic MSE approximation that captures both local mainlobe error and outlier error caused by sidelobes, and propose three MA structures suited to different hardware constraints. Numerical results show the approximation closely matches the true MSE and the proposed structures outperform CRB-optimized designs by orders of magnitude in the tested low-to-medium SNR range.
In briefAgentWeave: Efficient Distributed Agent Serving via Flow Decomposition
SummaryThis paper presents AgentWeave, a distributed serving system for agent applications built from LLM calls and tool invocations. It uses an aggregated deployment and a flow decomposition mechanism that splits long agent requests into fine-grained execution units, so short and newly arriving requests are not stalled by long ones. Experiments on four representative agent applications report 1.81x higher throughput, 2.45x lower average latency, and 2.04x lower P90 latency than state-of-the-art baselines.
In briefSpeeding up MoE training by jointly placing experts and coordinating replicas
SummaryTraining sparse mixture-of-experts models across many devices is slowed by uneven workloads and all-to-all communication. This paper presents PlaceMoE, which jointly optimizes expert layout and token-to-replica mapping and overlaps replica gradient synchronization with attention backward passes. On 32-GPU and 64-NPU clusters using Qwen3-VL and DeepSeek-V3, it reports up to 6.94x faster all-to-all communication and 1.74 to 2.33x end-to-end training speedups over VeOmni.
In briefEnergy-aware agent orchestration for deploying services across multiple sites
SummaryThis paper presents an orchestration framework that measures the energy used by LLM reasoning as part of its control loop, across a Kubernetes testbed spanning two European data centers. It compares single-shot tool calling, ReAct, and plan-and-execute strategies on ten natural-language intents. ReAct and plan-and-execute produced executable tool calls but required multiple LLM inferences, and prefix caching reduced mean ReAct GPU energy use by 29.4%.
Rodrigo Moreira, Alisson Chaves, Larissa F. Rodrigues Moreira et al.
In briefMaximizing sum rate in practical RIS-assisted rate-splitting multiple access systems
SummaryThis paper studies how to maximize the multiuser sum rate in reconfigurable intelligent surface (RIS) systems using rate-splitting multiple access (RSMA), accounting for the coupling between RIS phase and amplitude. To handle the resulting non-convex problem, the authors propose a variable separation algorithm that works with continuous phase shifts or any amplitude-phase coupling relationship and can extend to NOMA and OMA systems. Simulations show over 190% gain compared with an ideal-RIS benchmark, and the RSMA scheme significantly outperforms the corresponding NOMA and OMA schemes.
In briefVesta: a queueing-aware scheduler for LLM serving on real workloads
SummaryThis paper presents Vesta, a scheduler for distributed agentic LLM serving that accounts for queueing, cache locality, and KV cache transfer. Each session gets a stable holder for its authoritative KV cache, while each request's executor is chosen independently based on backlog and congestion. On real-world agentic workloads, Vesta improves throughput by 14.0% and reduces mean TTFT by 57.8% over the state-of-the-art scheduler in a PD-colocated deployment. In PD-disaggregated deployments, it reduces mean TTFT by up to 18.1% and P99 TTFT by up to 29.0% at matched throughput.
In briefOptimal receive beam design on the observation axis for RSMA satellite-UAV downlinks
SummaryThis paper studies receiver design for rate-splitting multiple access (RSMA) downlinks from a satellite to a UAV, separating the receiver's observation state from the physical signal. The authors use a physical-observation dual-axis framework and show that, with known line-of-sight response and noise covariance, a noise-weighted matched combiner maximizes every decoding SINR and the minimum user rate. Under directional uncertainty, they characterize the optimal receive direction and mean array gain. For mean rate, they prove the RSMA common-rate water level is jointly concave in receive gains and derive a computable global upper bound.
In briefDeepTelecom: A differentiable framework for simulating wireless channels in changing, mixed environments
SummaryThe paper presents DeepTelecom, a material-aware, high-fidelity multimodal channel dataset and reproducible generation framework for wireless networks with UAVs and reconfigurable intelligent surfaces. A large language model maps scene surfaces to frequency-dependent electromagnetic parameters, and a differentiable GPU-accelerated Sionna ray tracer simulates propagation across several frequency bands. The authors report that in 3D UAV localization, the 90th-percentile error falls from 31.05 m to 2.43 m compared with channel-state-information-only schemes.
In briefByzantine-tolerant single-writer register for GPU-style shared memory
SummaryThis paper studies whether atomic register semantics can be preserved in a GPU-inspired shared memory model when any thread, including the writer, may act Byzantine. It presents CAReg, which uses constant memory with a quantifiable probability of invalid reads, and SnapObject, which keeps a bounded window of authenticated writes to eliminate invalid reads. SnapObject tolerates up to n-1 Byzantine processes through lock-free and wait-free variants. The authors suggest that GPU phase-based, lock-step execution partially compensates for the lack of access control.
Chryssis Georgiou, Sathya Peri, Manaswani Piduguralla et al.
In briefUAV SAR imaging under non-ideal motion in ISAC systems using diffusion posterior sampling
SummaryThis paper studies synthetic aperture radar imaging from drones in low-altitude integrated sensing and communication systems, where non-ideal trajectories and unknown position errors cause severe image defocusing. The authors formulate imaging as an exact linear inverse problem and propose a diffusion posterior sampling framework for super-resolution, combined with joint motion compensation. Simulations report a PSNR above 21.3 dB even under half resolution truncation, and 23.65 dB with a position-error estimation residual of 0.19 cm under unknown motion errors.
In briefTurboProm: A Low-Cost, High-Performance Accelerator for Large-Scale Metrics Monitoring
SummaryLarge cloud and enterprise systems use tools such as Prometheus to monitor metrics like CPU, memory, and network usage, but analyzing these at scale is costly and slow. This paper presents TurboProm, a backwards-compatible drop-in accelerator that analyzes semantic-preserving summaries instead of estimating query results directly from raw time-series data. The authors report significant reductions in query latency and cost compared with state-of-the-art solutions across diverse real and synthetic workloads. The implementation is open source.
In briefLow-latency LDPC decoding using consensus-based residual belief propagation
SummaryThe paper proposes two new decoding algorithms for LDPC codes, C&R-RBP and C&DR-RBP, to reduce the number of iterations needed to reach a target decoding error rate. Both keep the same residual calculation complexity as the original RBP algorithm. C&DR-RBP performs better in early iterations, making it suited to low-latency use. Simulations based on the 5G New Radio standard show both methods outperform competing RBP-based algorithms in frame error rate and bit error rate.
In briefProduct selection under Mallows ranking preferences, with a faster approximation method
SummaryThis paper studies assortment optimization, choosing which products to offer, when customer preferences follow the Mallows model, which describes rankings by their distance from a reference order. The authors show that Mallows choice probabilities can be exactly reproduced by a single collection of mutually independent random utilities, for every possible offered set. Using this representation together with an existing framework for independent random utilities, they derive a polynomial-size finite-support reduction and an approximation scheme running in time O(n^{Õ(1/ε²)}), allowing arbitrary prices, concentration parameters, and modal rankings.
WhyUncovers independent-utility structure in Mallows choice, yielding a polynomial-time approximation scheme that improves on prior quasi-polynomial results with rigorous support.
In briefOrthogonal SC-LDPC codes for insertion, deletion and substitution channels
SummaryThe paper proposes a coding scheme for insertion, deletion and substitution (IDS) channels aimed at short DNA storage strands. It replaces an inner synchronisation code with a spatially coupled LDPC code spread across strands, so each decoded position serves as a synchronisation anchor. The authors bound the achievable rate and show that, for 100-nucleotide strands, the scheme outperforms a rate-matched concatenated construction.
Frederik Walter, Jessica Bariffi, Antonia Wachter-Zeh
In briefSequential detection of when a transmitter switches between PSK modulation schemes, with controlled false alarms and delay
SummaryAdaptive communication transmitters may switch modulation schemes at an unknown time, and receivers must detect the switch quickly without frequent false alarms. The paper studies sequential detection of switches between phase shift keying schemes over an additive white Gaussian noise channel, using CUSUM and MOSUM detectors built from the log-likelihood ratio and its low-SNR approximation. For a switch from QPSK to BPSK, the authors derive the moment-generating and probability density functions of the approximate statistic to relate the threshold to the average run length to false alarm and worst-case detection delay. Monte Carlo simulations show the derived expressions closely match empirical results across thresholds.</summary>
In briefDemand-private coded caching with unequal file sizes and cache sizes: exact trade-off for two users and two files
SummaryThis paper studies coded caching where files and user cache capacities are heterogeneous and each user's demand must stay private from the other users. The authors extend the virtual-user reduction so that non-private schemes can be turned into demand-private ones under heterogeneous cache sizes. For two users and two files with arbitrary file sizes and cache capacities, they exactly characterize the memory-rate trade-off using eight linear inequalities, with achievable schemes and information-theoretic lower bounds.
In briefLimits of covertly serving extra jobs in a queue when Willie sees busy-period statistics
SummaryThe paper studies a queueing system run by Bob that serves legitimate user Willie's jobs, while Bob tries to covertly serve additional Poisson-arriving jobs. Willie receives per-busy-period statistics, either the number of jobs served or the busy-period duration. Under either statistic, the maximum expected number of covert extra jobs scales as the square root of the number of busy periods. The associated scaling constant, called the covert queueing constant, is characterized for each statistic.
Anumula Venkata Sai Sree Sahithi, Arti Yardi, Jithin Ravi
In briefComparing Scheduling Methods for Cloud, Fog, and Edge Computing, with a Selection Guide
SummaryThis review compares methods for scheduling tasks, jobs, and workflows across cloud, fog, and edge computing environments. It covers heuristic, metaheuristic, learning-based, and hybrid techniques, analyzing studies published between 2020 and 2026 by granularity, environment, objectives, constraints, metrics, and tools. The authors describe trade-offs among computational efficiency, optimization capability, adaptability, and implementation complexity across these families. They then propose a decision framework to help select a technique based on workload, environmental conditions, objectives, resource constraints, and computational requirements.
In briefA paper on GPU-resident execution that keeps sensor processing on the device to reduce timing jitter.
SummaryThis paper proposes GPU-resident Execution, a model that keeps sensor input and GPU processing in a long-lived device-resident loop so that no host CPU wakeups are needed. The authors implement it in NVIDIA's Holoscan C++ framework and evaluate it on NVIDIA GPUs and SoCs across several workloads. On a 1080p, 60 Hz camera-to-display application, photon-to-display jitter drops from 4.83 ms to 0.66 ms, and software-processing jitter drops from 1.70 ms to 0.01 ms, compared with a CPU-driven GPU-memory RDMA baseline.
Soham Sinha, Rahul Choudhury, Shekhar Dwivedi et al.
In briefA 1/2-approximation algorithm for Pandora's problem with deadlines
SummaryThis paper studies Pandora's problem with deadlines, a sequential search model where each box has a publicly known deadline for opening and selection. The authors prove that a simple threshold strategy achieves a 1/2-approximation to the optimal expected utility. This improves the previous guarantee, arbitrarily close to 1/4, due to Amanatidis et al. (2025).
In briefAsymmetric Information in Dynamic Lending Games
SummaryThis paper studies a dynamic lending game in which supply is limited but reusable, and borrower arrivals and loan durations are random. It compares how disclosing rivals' aggregate available supply to all, some, or no lenders affects payoffs. The authors find that asymmetric disclosure can raise total lender payoffs while total lending stays the same, and that which lenders receive the information matters.
Kristijan Atanasov, Frederik Mallmann-Trenn, Carmine Ventre
In briefSmoothed Polynomial-Time Algorithms for Deterministic Discounted-Sum and Mean-Payoff Games
SummaryThis paper studies two-player zero-sum games on graphs with discounted-sum and mean-payoff objectives, a problem in NP ∩ coNP whose polynomial-time solvability remains open. The authors analyze a variant of strategy iteration with a new technique. They show it achieves smoothed polynomial complexity for general game graphs, not only ergodic ones. The guarantee is an expected runtime bound, strengthening earlier high-probability results.
WhyA new analysis technique removes the ergodicity restriction and upgrades high-probability to expected-runtime smoothed polynomial bounds, resolving an open problem; impact mainly within game theory an
In briefDistortion of voting to pick several winners when voters give only coarse thresholds
SummaryThis paper studies multiwinner voting, where n voters and m candidates are used to select a committee of size k. Voters report only coarse information based on t utility thresholds rather than exact cardinal utilities. The authors give upper bounds on distortion for deterministic and randomized rules, and norm-dependent lower bounds that match them asymptotically for constant committee sizes.
Mohamad Latifian, Marzieh Nilipour, Alexandros A. Voudouris
In briefComputational Complexity of Autobidding Equilibria in GSP Sponsored Shopping Auctions
SummaryThis paper revisits autobidding equilibria in GSP sponsored shopping auctions, following Dütting et al. It shows that equilibria exist under both budget constraints (ABE-B) and return-on-investment constraints (ABE-R), and that they always admit polynomial-size support over the allocation space. Exact equilibria are shown to be computable in FIXP, approximate equilibria in PPAD, and computing approximate equilibria is PPAD-hard for both models.
