In briefResetting-based designated information distribution
SummaryThis paper studies a network of n nodes under random gossip dynamics, where a resetting mechanism with rate β is applied to a designated subset of r nodes. In the small-β and long-time limit, the expected sum of all node values asymptotically equals n/r times the sum of the initial values of those r nodes. The authors also analyze the speed of convergence and provide an approximate high-probability confidence interval for the expected sum at finite time and finite β.
In briefReference-dependent utility optimization for goal-oriented communication systems
SummaryThis paper proposes a general resource allocation framework that transforms operational rewards through reference-dependent user utilities and evaluates them with expected utility or cumulative prospect theory. Weighted sum-reward maximization, proportional fairness, and α-fairness appear as exact special cases. The paper gives optimality conditions for deterministic orthogonal systems and convergence guarantees for successive convex approximation in nonorthogonal and stochastic systems. Numerical examples show how reference points, utility slopes, channel quality, and budgets shape the active users and allocation policies.
Symeon Vaidanis, Photios A. Stavrou, Marios Kountouris
In briefApplying multi-stage distributed real-time system design to vehicular edge computing
SummaryThis paper addresses vehicular edge computing, where computation-intensive applications must access different resources in a fixed sequence, forming a multi-stage pipeline. The authors model this setting using established multi-stage distributed real-time scheduling techniques. They propose a joint mapping and scheduling heuristic that accounts for vehicle mobility and deadlines, aiming to minimize local energy use while meeting offloaded job deadlines. Simulations show it consistently outperforms a delay-composition baseline and a deadline decomposition approach.
In briefDispatching MoE tokens sooner by offloading NCCL-EP communication during dual-batch overlap
SummaryDual-batch overlap hides mixture-of-experts communication behind another microbatch's computation, but only after dispatch send releases the GPU. The paper evaluates three configurations built on NCCL-EP (GPI, GPI-MC, and GDAKI Free Flow) that reduce the time dispatch send holds the GPU. On a GB300 NVL72 system, all three cut dispatch-send time relative to default GDAKI over most tested operating points. In disaggregated vLLM serving with DBO, they lower median TPOT by 15-25% versus default GDAKI and 16-42% versus DeepEP v1.
Hsin-Chun Yin, Khaled Hamidouche, Ian Ziemba et al.
In briefEvolving Infra: Adaptive LLM Serving for AI Agents at Scale
SummaryThis paper presents Evolving Infra, a serving framework for large language models that adapts at runtime to the changing workloads of AI agents. It separates compute allocation from KV cache management, and uses an online planner and an agent-aware distributed KV manager to balance prefill and decode nodes and guide cache eviction and transfer. Tested on coding, math, and computer-use tasks with up to 256 GPUs, it reduces mean end-to-end latency and P95 time-to-first-token by up to 4.1x and 14x compared with NVIDIA's Dynamo.
In briefOutage and diversity limits of fluid antenna multiple access under correlated Nakagami-m fading
SummaryThis paper analyzes fluid antenna multiple access (FAMA) systems under spatially correlated Nakagami-m fading, providing the first exact and asymptotically tight outage analysis. It models the joint statistics of desired and interference powers across antenna ports and derives an explicit integral expression for outage probability. The achievable diversity order is shown to depend on fading severity and the spatial correlation rank set by the finite aperture rather than on the number of ports. A Gamma-matched closed-form approximation is also introduced and validated by Monte Carlo simulations.
In briefCorrelation-informed soft outputs to reduce BICM information loss in MIMO receivers
SummaryThe paper shows that in MIMO bit-interleaved coded modulation, passing only per-bit soft information between detector and decoder discards correlations among bits within each transmit vector, limiting one-shot receiver performance. The authors propose COIL, a soft-output format that adds pairwise coupling logits for correlated bits, and show how to incorporate them into a belief-propagation channel decoder. In a 4x4 MIMO-OFDM evaluation, a one-shot LMMSE receiver with COIL outperforms a conventional unbiased LMMSE receiver by 1.3 dB and comes within 0.8 dB of four-iteration IDD with a single detection pass.
In briefOptimal Black Hole Search Using Programmable Matter Particles
SummaryThis paper studies the Black Hole Search problem in the context of programmable matter, which consists of tiny mobile particles that can self-arrange. The black hole destroys any particle that visits it without leaving a trace. Particles move in a convex area on a triangular grid, and the goal is to identify a restricted set of vertices containing the black hole while keeping at least one particle alive. The authors present an algorithm using three two-colored particles and show it is Pareto optimal in particle and color count, and asymptotically optimal in number of epochs.
Giuseppe Antonio Di Luna, Alfredo Navarra, Francesco Piselli
In briefSample-efficient noninteractive simulation over finite alphabets
SummaryThe paper studies noninteractive simulation, where two parties without communication generate outputs from independent copies of a finite correlated source. It gives a construction for arbitrary finite output alphabets showing that any simulable joint law can be approximated with a number of samples polynomial in the alphabet sizes and inverse accuracy. For a concrete biased binary source, the authors certify an explicit sample exponent below 19.
In briefSpeeding Up Sparse Equation Solvers Using Graph Computing Accelerators
SummarySolving large sparse linear systems by Gaussian elimination creates extra non-zero entries, so solvers use nested dissection reordering to reduce them. This paper explores using the UpDown accelerator, designed for graph and irregular algorithms, to speed up the METIS nested dissection reordering algorithm. Results from validated simulation and FPGA emulation of small UpDown systems indicate about two orders of magnitude speedup over a single CPU for METIS.
Jiya Su, Fran\c{c}ois-Henry Rouet, Jeff Beisheim et al.
In briefOverlapAdvisor: Choosing Compute-Communication Overlap for Distributed MoE Serving
SummaryThis paper addresses mixture-of-experts (MoE) inference, which relies on All-to-All communication, where no single overlap strategy between computation and communication is best everywhere. The authors propose OverlapAdvisor, which models overlap stages at two granularities and uses a unified performance model to select a strategy for each configuration at launch time. Evaluated on four MoE models with SGLang and vLLM, it improves end-to-end throughput by up to 31.7%.
In briefEstimating link fading and fading shape from downlink SINR in cell-free massive MIMO
SummaryThis paper proposes a method to jointly estimate link-specific large-scale fading coefficients and a common Nakagami-m shape parameter from power and downlink SINR measurements in cell-free massive MIMO. A tractable SINR distribution is derived under global zero-forcing precoding, Nakagami-m fading, and fractional power control. A model-driven neural network is built to solve the resulting maximum-likelihood estimation problem. Simulations show the reconstructed SINR distributions closely match empirical ones.
In briefSpeeding up a new LLM memory module on AWS Trainium using custom on-chip GPSIMD operators
SummaryThe paper addresses Engram, an emerging memory module in large language models that mixes irregular integer work with dense computation. The authors build a toolchain that compiles C++ custom operators for Trainium 3's GPSIMD cores and injects them into NEFF files generated from NKI kernels. Compared with a hand-optimized NKI baseline for DeepSeek-V4.1-Flash, two GPSIMD operators speed up n-gram hashing by 1.79× at decode, table retrieval by up to 4.95×, and the whole Engram computation by up to 1.14×.
In briefBounds and constructions for optimal wide-gap frequency hopping sequences
SummaryThis paper studies wide-gap frequency hopping sequences (WGFHS), which reduce interference to received signals in frequency-hopping systems. It updates upper bounds on the minimum gaps of optimal uniform WGFHS with parameters (v,s,λ;e), where v=ks, k≥3, λ=k, and s is an odd number at least k(k-1)+1. For the cases k≤3, the authors completely prove the existence of such sequences.
In briefWhy Kubernetes can't contain a full lazy-pulling cache, and a bounded design for model serving
SummaryThe paper shows by experiment that the node-level cache used by lazy container-image pulling on Kubernetes can be neither bounded by the platform nor detected by monitoring when it fails. It proposes a design that sets a byte budget on the cache and ensures a failed cache write never causes a failed read, without changing kubelet or the container runtime interface. Results from six campaigns on AWS are reported, and the open-source implementation has been proposed upstream.
WhyThe authors report two impossibility results for lazy-pulling caches on Kubernetes and a byte-budget design keyed to filesystem state, evaluated across six campaigns with an open-source implementation
In briefFixed points of binary belief propagation: finiteness, rigidity, and topological complexity
SummaryThe paper studies belief propagation fixed points of binary Ising models under fixed external fields. It proves that the number of admissible complex fixed points is finite, even at degenerate parameters, and characterizes graphs whose positive response is a single smooth real line. It further shows that, when varying two vertex fields, cubic bipartite graphs can realize every genus, with the minimum vertex count growing logarithmically in the genus. The three main theorems are formally verified in Lean 4 with Mathlib.
In briefKarhunen-Loève transforms of autoregressive windows using matched symmetry groups
SummaryThis paper describes the Karhunen-Loève transform of a window from a stationary autoregressive process of any order through its poles, extending the Algebraic Diversity framework. It analyzes how this transform differs from the type-II discrete cosine transform and bounds the coding-gain loss. It also presents a fast implementation based on the matched symmetry groups.
In briefHow hard is finding competitive prices when agents are paid to take on chores?
SummaryThis paper studies how computationally hard it is to find competitive equilibria in Fisher markets where divisible chores carry disutility. For linear disutilities, exact equilibrium computation is shown to be CLS-complete, even with binary earning requirements. For CES disutilities, finite-precision price computation is also CLS-complete for every fixed exponent above one. The authors additionally give a strongly polynomial exact algorithm for markets whose disutility matrix has rank at most two.
WhyThe authors prove CLS-completeness for chores Fisher market equilibrium and give a strongly polynomial algorithm for rank-two disutility matrices.
In briefCompressing memory-based data while keeping individual items private
SummaryThe paper studies compression of data sources under a privacy constraint: decoding one source symbol must reveal nothing about other symbols beyond that symbol itself. The authors show that for any stationary ergodic source over a finite alphabet, every rate above the entropy rate can be achieved with exact privacy and error-free local decoding on source sequences whose probability tends to one. They also present non-ergodic sources for which private local decoding carries a strictly positive rate penalty.
WhyThe authors extend compatibility of private local decoding with entropy-rate compression from memoryless to stationary ergodic sources and show a strict rate penalty for non-ergodic sources.
Venkat Chandar, Aslan Tchamkerten, Shashank Vatedka
In briefSharp Rates for Recovering Bandlimited Functions from Localized Oversampling
SummaryBandlimited functions, whose frequencies lie in the interval [-δ, δ], can be approximated from a finite set of sample values. The paper shows that the optimal worst-case recovery algorithm has an error decaying exponentially in the number of samples n, and it gives a sharp rate depending on n and δ. It also finds that one-sided sample data can reconstruct such functions with exponentially decaying error, and identifies the optimal rate for this case.
