In briefAscend-SpMM: Speeding Up Sparse-Dense Matrix Multiplication on Ascend NPUs
SummaryThis paper presents Ascend-SpMM, a framework for accelerating sparse matrix-dense matrix multiplication, a core operation in graph neural networks and scientific computing, on Huawei Ascend NPUs. It uses a cost model to assign dense regions to the Cube path and sparse residuals to the Vector path, and adds customized kernels that hide memory latency. Reported results show a 12.70 to 20.55 times geometric-mean speedup over the strongest NPU baseline and up to 6.8 times faster end-to-end GNN training. The source code is publicly available.
In briefBetter lower bounds on kissing numbers in dimensions 49 to 61
SummaryThis paper gives explicit kissing configurations in R^d for every dimension from 49 to 61, improving the best previously published lower bound in all thirteen dimensions. The configurations come from applying the antipodal block lift of Yang et al. to the 52,416,000-point minimal shell of the extremal lattice P_48p, using a class of 7,193 minimal lines. All results are verified in exact integer arithmetic, and the certificates are released with the code.
In briefSuperscalar Loop Processors: A Lightweight Multiple-Issue Architecture for Regular Loop Workloads
SummaryThis paper proposes the Superscalar Loop Processor (SLP), a lightweight superscalar architecture aimed at regular, compute-intensive loop kernels. It enables multiple-issue execution without the complexity of dynamic out-of-order scheduling. The authors present Schnizo, an open-source prototype built by enhancing a tiny RV32I core with floating-point support. On several ML and HPC microkernels, Schnizo achieves higher IPC and FPU utilization than the scalar baseline while keeping in-order efficiency.
In briefSignal-level detection of RIS spoofing attacks in vehicular ISAC using Green Learning
SummaryThis paper addresses spoofing attacks by a malicious reconfigurable intelligent surface (RIS) in vehicular integrated sensing and communication (ISAC) systems, where spoofed echoes can distort a roadside unit's sensing results. The authors propose a signal-level Green Learning framework that detects RIS spoofing directly from matched-filtered RSU echoes, combining SAAB, DFT, and XGBoost. Simulations report 99.335% detection accuracy for a 32-element malicious RIS at -80 dBm receiver noise power, with 0.1119 million FLOPs. The method also generalizes to previously unseen RIS sizes.
In briefLightweight hybrid precoding for wideband massive MIMO-OFDM using ray-tracing parameters and cascade XGBoost
SummaryThis paper proposes XGB-SVD, a lightweight hybrid precoding method that uses cascade XGBoost to map compact ray-tracing path parameters directly to analog beamforming codeword indices. The digital precoder and combiner are derived from a reduced-dimensional equivalent channel via singular value decomposition. Evaluations on five DeepMIMO scenarios across sub-6 GHz and millimeter-wave bands show it outperforms the DNN baselines considered and approaches SOMP in several environments, with lower online computation and training costs.
In briefApto: Self-Adaptive CPU-GPU Hybrid MoE Inference on Shared PCs
SummaryThis paper presents Apto, a system for running Mixture-of-Experts models on memory-limited PCs by splitting expert computation between CPU and GPU. Instead of fixing the split in advance, CPU and GPU claim experts from opposite ends of a shared queue, so the workload adapts as execution proceeds. The authors report prefill and decode speedups over GPU-only and hybrid baselines, including larger gains under concurrent CPU workloads.
In briefA sharp, instance-specific convergence rate for the Add-λ estimator, derived via feature space decomposition
SummaryThis paper uses the Feature Space Decomposition framework to study the Add-λ estimator of a discrete probability distribution. It establishes non-asymptotic upper and lower bounds on the expected Kullback-Leibler estimation error that match each other. The results hold for any signal, dimension, sample size, and any λ > 0.
In briefCounting in Anonymous Directed Dynamic Networks: Exact Worst-Case Time and Per-Network Hardness
SummaryThis paper studies the Counting problem in anonymous directed dynamic networks, where agents lack identifiers, some are leaders, and the population size is unknown. The authors give a distributed algorithm that makes every agent output the population size and halt after a round count that is tight in the worst case. They also show that, for a single fixed network, deciding whether an agent already has enough local information to determine the size is coNP-complete.
WhyThe authors give the exact worst-case round complexity of Counting in anonymous directed dynamic networks with a tight bound, and report coNP-completeness of per-network optimization.
In briefEffect of Reference Signal Choice on Time-Varying 38 GHz Channel Delay Spread
SummaryThis paper examines how the choice of reference waveform affects estimates of temporal dispersion in a time-varying 38 GHz radio channel. Three reference waveforms, generated numerically in MATLAB, measured via a conducted connection, and recorded in an anechoic chamber, are compared using the same wideband measurements taken in a dynamic road environment. The waveforms differ noticeably in time-domain and spectral characteristics, yet the resulting power delay profiles show similar multipath structures and the estimated dispersion parameters differ only slightly. The results indicate limited sensitivity to reference waveform selection under the studied measurement conditions.
Jaros{\l}aw Wojtu\'n, Bartosz Czaja, Tomasz Grasza et al.
In briefCommunication with Adversarial Users: Byzantine Models in Multi-User Information Theory
SummaryThis survey examines communication systems in which some participating users may behave adversarially, a setting modeled as Byzantine communication. It offers a unified information-theoretic view of multi-user communication with such users. The authors review recent work on Byzantine multiple access channels, covering reliable communication, adversary identification, and distributed function computation. They also discuss extensions to classical-quantum channels and outline open problems and future directions at the intersection of communication, computation, and security.
In briefMinimizing worst-user outage probability in active RIS-aided wireless systems
SummaryThis paper studies how to minimize the outage probability of the worst user in wireless systems assisted by active reconfigurable intelligent surfaces (RIS). Because the objective is discontinuous over an unknown channel distribution, the problem is reformulated using user weighting, a softplus surrogate, and entropic tail-risk aggregation. A data-driven stochastic optimization method based on historical channel samples then jointly optimizes the beamformer and reflection coefficients with projected adaptive moment estimation. Simulation results show the proposed approach's advantages for reliability-oriented active RIS transmission design.
In briefChoosing 6G Uplink Waveform and Rank in Distributed Reception
SummaryThis paper studies how 6G uplink selects between CP-OFDM and DFT-spread OFDM waveforms and between rank 1 and noncoherent rank 2 when reception is distributed across access points. It derives waveform switching thresholds and a rank selection margin based on transmit power, DFT despreading loss, and spatial multiplexing gain. It also analyzes how per-AP port gains and correlations shape aggregate spatial separation and mode choice. Numerical results show that different AP deployments can favor different waveforms and ranks.
In briefSecond-order limits of the scalar Gaussian broadcast channel
SummaryThe paper characterizes the second-order coding region of the scalar Gaussian broadcast channel along the curved boundary of its capacity region. The result holds under both average and maximal error criteria and allows correlated receiver noises. At each interior boundary point the second-order region is a halfspace. Achievability uses spherical superposition coding with blocklength-dependent power allocation, while the main innovation is the converse proof.
In briefSquare-root-length MDS codes from hyperelliptic curves
SummaryThe paper constructs maximum distance separable (MDS) algebraic geometry codes over F_q for fixed genus g≥2 and any rate 0<ρ<1, for all sufficiently large odd squares q. The code length is about √q+1 plus O_g(q^{1/4}), with parameters [n,⌊ρn⌋,n-⌊ρn⌋+1]_q. The codes use full Riemann–Roch spaces on genus-g hyperelliptic curves. For large q they are not monomially equivalent to generalized Reed–Solomon (GRS) codes.
WhyFor fixed genus g≥2 and 0<ρ<1, the authors construct MDS algebraic geometry codes of length about √q on hyperelliptic curves for large odd squares q, reporting they are not monomially equivalent to GR
In briefSecrecy sum-rate maximization for RSMA-enabled pinching-antenna systems
SummaryThis paper studies maximizing the secrecy sum-rate in a downlink rate-splitting multiple access (RSMA) system using a single-waveguide pinching-antenna setup, in the presence of an external eavesdropper. It jointly optimizes RSMA power allocation, secure common-rate allocation, and pinching-antenna positions under transmit-power, per-user secrecy-rate, and placement constraints. Because the problem is nonconvex, the authors use an alternating-optimization framework combining successive convex approximation and feasibility-preserving projected gradient ascent. Numerical results show the proposed design consistently achieves higher average secrecy sum-rate than the considered benchmarks.
In briefStudy on whether data exchange between competing firms can remain stable
SummaryThis paper examines whether competing firms can exchange data in a stable way when their predictive performance affects market share. It models n firms sharing parts of their datasets and uses core stability as the solution concept. The authors show that a core-stable exchange always exists in fully rational markets, but may fail to exist under logit-style demand. They also show that equal sharing of market-share revenue within coalitions restores existence of a core-stable exchange.
WhyThe authors formalize core stability of data exchange among competing firms, give existence and non-existence results under two demand models, and report simulations.
Yuanchen Brian Tang, Jiaxin Song, Bhaskar Ray Chaudhury et al.
In briefServing Robot Fleets With Vision-Language Brains: Workload, Cache Thrashing, Admission Control
SummaryThis paper studies how to serve a fleet of agentic robots, each combining a vision-language model, a tool-using harness, and a vision-language-action policy. The authors characterize 80 LIBERO-PRO episodes and replay them on a vLLM serving engine on H800 GPUs. They report that raising concurrency from 20 to 32 robots sharply reduces brain throughput, and that a static active-episode cap restores it. The results are presented as practical principles for capacity planning and admission in agentic robot fleets.
In briefScheduling CPU and GPU work layer by layer so large language models can run on consumer GPUs with limited memory
SummaryThis paper addresses running large language models that exceed a single consumer GPU's memory. During decoding, computing some layers on the CPU can take less time than transferring weights, so the method runs those layers on the CPU while the next layer's weights are sent to the GPU in the background. The placement of computation and weight transfers is formulated as a latency minimization problem and solved with an approximation using relaxed temporal constraints. In a consumer desktop environment, the approach speeds up OPT-30B inference by 3.5x over HuggingFace Accelerate and 6.3x over DeepSpeed Zero-Inference.
In briefToken-as-a-Utility: A Model-Agnostic Way to Meter and Supply AI Inference Services
SummaryThis paper presents Token-as-a-Utility, an architecture that combines inference metering, service-level-aware provisioning, and settlement between operators. Consumption is measured in Reference Inference Units (RIUs), and occupancy-indexed operator ratings recover shared serving costs. Measurements on three GPUs and trace-driven multi-tenant experiments show lower energy estimation error and reduced tenant mispricing compared with flat token counting.
In briefRemoving Servers from State Machine Replication (Serverless SMR, S²MR)
SummaryThis paper proposes Serverless State Machine Replication (S²MR), which stores both the materialized state and the durable total order of commands on disaggregated storage, with clients maintaining the materialized state rather than servers. The authors build AirDB, a transactional key-value store, on this design. According to the paper's evaluation, AirDB over S²MR is unaffected by server failures and bad client activity and spins up 10x to 45x faster from a cold start than etcd over Raft.
In briefRandomized consensus that survives mobile crashes via multiple leaders
SummaryThis paper addresses consensus protocols that fail to terminate when an adversary can move between replicas and repeatedly crash one of them. It proposes Multi-Leader Fixed-View Simplex (ML-FVS), in which several leaders may each broadcast a proposal and replicas vote on the set of proposals they have seen. A hidden threshold beacon revealed after voting ranks the proposals. Under a constant crash budget and delayed Byzantine corruption, the authors obtain O(Δ) expected latency, and they evaluate the protocol under random delays and stochastic faults.
Ittai Abraham, Kaya Alpturer, Pranav Garimidi et al.
In briefSub-terahertz LEO satellite links: does Doppler shift pose a major design challenge?
SummaryThis paper studies how Doppler frequency shift affects sub-terahertz low Earth orbit satellite communication using OFDM transmission. The authors build an evaluation framework covering both Doppler-induced and other signal impairments, and compare three engineering approaches to handling Doppler shift. They find that ignoring Doppler substantially degrades performance, but a few low-complexity measures eliminate most negative effects. A single Doppler estimate over a 5-minute satellite pass may suffice to maintain high performance throughout the connection.
Seongjoon Kang, Parisa Ramezani, Ki Won Sung et al.
In briefAdapBench: Adaptive Early-Stopping and Prioritization for Faster Cloud CI/CD Microbenchmarking
SummaryThis paper presents AdapBench, an online scheduling system that shortens the run time of microbenchmark suites so that performance regressions in new software releases are caught sooner. It decides at runtime which benchmarks to prioritize and stops each one early once enough evidence accumulates to accept or reject a performance change. Evaluated on eight Java projects and four Go projects, it saves 42% of aggregate execution time on Java and 85% on Go at matched detection and false-positive rates, with negligible scheduling overhead.
Sebastian Koch, Jonas Heisterberg, Nils Japke et al.
In briefBinary two-deletion and two-insertion channels: which received words maximize input entropy
SummaryThe paper determines which received binary words maximize input entropy under uniform transmission over two-deletion and two-insertion channels. For two deletions, the maximizers are exactly the two alternating words at every positive received length. For two insertions, the same holds from received length five onward, with all smaller cases and exact maximal values given.
In briefA study of ET-SoC-1 vector and parallel performance using a matrix multiplication benchmark
SummaryThis paper studies the performance of the ET-SoC-1, a massively parallel RISC-V accelerator, using single-precision matrix multiplication (sGEMM) as the benchmark. Two device kernels are implemented and evaluated: a register-blocked SIMD baseline and a TensorFMA32 kernel that drives the same arithmetic datapath through a tensor sequencer. Since both kernels share the same arithmetic peak, the comparison isolates the effect of operand delivery. The results are compared against several CPUs and GPUs.
In briefShared randomness for sampling random permutations in parallel rounds
SummaryThis paper studies how much shared randomness is needed to sample a uniform random permutation when outputs are produced in parallel rounds. Coordinates committed in the same round are independent given the full history and the current shared choice, and each shared choice is charged the logarithm of its conditional support size, summed along the worst execution path. The authors determine optimal total-variation error under various shared budgets and round densities, and give the optimal cost for exact sampling in Θ(n) rounds.
