In briefLinear codes built from zero-divisor graphs of the ring F_p[x]/⟨x^4⟩
SummaryThis paper studies linear codes derived from the zero-divisor graph of the finite local ring R = F_p[x]/⟨x^4⟩ for an odd prime p. Using an equitable partition of Z*(R), it determines parameters of the binary incidence code and constructs induced bipartite subgraph codes achieving dual minimum distance 4. Over F_p, it parameterizes the adjacency and Laplacian codes, derives the Laplacian code's weight distribution via the MacWilliams identity, and establishes its LCD property. It also determines graph automorphism groups and builds two families of Griesmer-optimal linear codes from extremal subgraphs.
In briefSorted mutability as an entropy-based measure: theory and two applications
SummaryThe paper studies sorted mutability, a information-based quantity with an explicit entropic character. The authors show it is directly proportional to data diversity and can be expressed via Shannon entropies, linking compression-based measures to information theory. Unlike regular mutability, sorted mutability removes sequential ordering and characterizes the distribution of observed values. Applications include the two-dimensional ferromagnetic Q-clock model and wind-power production data.
Michel Aguilera, Eugenio E. Vogel, Francisco J. Pe\~na et al.
In briefHintKD: Bandwidth-efficient knowledge distillation for cloud-edge AI inference
SummaryThis paper addresses the heavy wireless bandwidth cost of cloud-edge collaborative inference in 6G networks, where high-dimensional features or soft probability vectors must be transmitted. It proposes HintKD, which compresses the cloud teacher's guidance into a compact discrete hint index via vector quantization, and uses a lightweight FiLM-based adapter on the user equipment to modulate student features. Experiments on the DeepSense 6G beam prediction task report competitive accuracy with O(1) inference-side communication, and MNIST experiments are reported to verify generality.
In briefReinforcement learning for jointly positioning movable antennas in cell-free massive MIMO
SummaryMovable antennas create non-convex minimum-spacing constraints that make joint optimization of antenna positions and transmission strategies hard to solve. The paper proposes GLIIR-HAPPO, a heterogeneous multi-agent reinforcement learning method that splits the problem into coordinated subproblems, embeds geometric constraints in a penalty-based reward, and uses a graph-based critic with role-conditioned federated distillation. Simulations report significant sum-rate gains over state-of-the-art MARL schemes, performance close to a centralized version, and much lower communication overhead.
In briefLachesis: Lifetime-aware placement of AI agent memory across fast memory and high-bandwidth flash
SummaryLarge language model agents accumulate context as KV cache, which consumes substantial memory. High-bandwidth flash offers much larger capacity but has limited write endurance. The paper proposes Lachesis, which places each KV cache segment in HBM or HBF according to its lifetime and frees it once no longer read. In trace-driven simulation, it extends HBF lifetime by 1.19 to 3.13 times over HBM-first placement, reaching 3.3 to 12.2 device-years.
In briefvTen: A tensor-centric framework for verifying specialized AI hardware accelerators
SummaryThe paper argues that a gap between tensor-level software models and signal-level hardware testbenches slows verification of Domain-Specific Accelerators. It proposes vTen, a data-centric framework that separates verification intent from execution, using a declarative DSL and kernel-level batching to reduce host-simulator interactions. On a production-scale 3D U-Net accelerator, vTen reportedly achieves about 2x faster simulation latency and 60.3% lower code complexity than Cocotb.
In brief6G NeXt: Use Cases and Architecture for 6G Split-Computing Network Applications
SummaryThis paper examines two 6G use cases studied by the 6G NeXt research project: holographic communications and a drone anti-collision system. It explains how both impose different requirements on communication and computing infrastructure, since clients are widely distributed and must cooperate, and holographic communication needs high processing power. The authors argue that a high-speed distributed backbone computing infrastructure implementing split computing, tightly integrated with the wireless network, is needed. They elaborate a solution approach and discuss novel technical methods based on an overall communication and computing architecture.
Sergiy Melnyk, Qiuheng Zhou, Hans D. Schotten et al.
In briefOptimal conversion bandwidth for MDS convertible codes when r^F<k^F<r^I
SummaryThis paper studies MDS convertible codes, which let a distributed storage system transform one erasure code into another while both remain MDS. For the parameter range r^F<k^F<r^I, the authors derive new rank inequalities and an improved lower bound on conversion bandwidth, split into three cases. They then give explicit constructions for each case, showing that the bound is tight.
In briefApproximate message passing with discontinuous denoisers and corrected state evolution
SummaryThe paper studies approximate message passing (AMP) with discontinuous denoisers such as hard slicers, quantizers, and hard thresholding, for which classical state evolution analysis does not apply. It shows that the standard Onsager coefficient misses a boundary term from the jumps, and proves that AMP follows state evolution when this term is included, for i.i.d. Gaussian matrices. Simulations with 4- and 8-level PAM show the corrected AMP follows state evolution and approaches Bayes-optimal AMP performance.
Kuranage Roche Rayan Ranasinghe, Takumi Takahashi, Giuseppe Thadeu Freitas de Abreu
In briefGPU acceleration for Awkward Array using Python cuda.compute
SummaryThis paper presents a CUDA execution model for Awkward Array, a Python library used in high-energy physics for nested, variable-length data. It uses the Python CUDA Core Compute Libraries (CCCL) to avoid custom CUDA kernels and fuses multiple operations to reduce kernel launch overhead. Lazy execution builds and optimizes expression graphs before kernel generation. The authors report performance gains over earlier eager GPU execution and note that the design supports user-defined Python code.
In briefGraph-theoretic bounds for broadcasting non-linear function computations
SummaryThis paper studies non-linear function computation broadcast, where a sender with N datasets sends one common message to K users who each hold side information and need a function of the datasets. The authors build a broadcast graph using Körner's characteristic-graph framework and obtain achievable rates for arbitrary K, general source distributions and finite-field demands, without restricting to linear operations. They also derive converse bounds and show, in three-user examples, that non-linear encoding can strictly beat the best linear schemes.
Derya Malak, Vijith Kumar K. P., M. Reza Deylam Salehi
In briefOnline fair division of indivisible items when each item can go only to eligible agents
SummaryThis paper studies fair division of indivisible items where each item may be assigned only to a subset of eligible agents, a model equivalent to restricted assignment in scheduling. The authors define eligibility-aware versions of proportionality and envy-freeness. For identical valuations, they give tight online approximation bounds, showing a sharp separation between goods and chores for proportionality. They also analyze which envy-based guarantees, such as EF1 and EFX, remain achievable offline and online.
WhyIntroduces eligibility-aware fairness notions and proves tight online approximation bounds (Θ(√m), Θ(log m)) for restricted assignment, a clear theoretical contribution with impact mainly within fair
In briefInformation Bottleneck for Stable Variables: A Generalization of the Gaussian Case
SummaryThis paper generalizes the Gaussian information bottleneck problem to jointly stable variables X and Y, where X = Y + A and A is stable and independent of Y. The authors derive optimal linear stable encoders of the form T = kX + N, giving a closed-form solution and critical value. They show that, except in the Gaussian case, such stochastic linear encoders are strictly suboptimal but become asymptotically optimal at large compression rates. The linear solution reduces to the classic Gaussian information bottleneck for jointly Gaussian variables.
In briefSparse kernel mechanisms for local differential privacy on discrete channels
SummaryThis paper studies sparse locally private discrete mechanisms built from a nonnegative kernel and an input-dependent admissible output support. It gives an exact characterization of pure and approximate local differential privacy for this class, showing that the approximate privacy defect splits into support leakage and overlap excess loss. The results are applied to several discrete families, including sparse staircase, sparse randomized response, and Skellam-based mechanisms.
Amirreza Zamani, Parastoo Sadeghi, Mikael Skoglund
In briefPrivate IPFS storage prototype for verifiable digital museum collection data
SummaryThis paper describes a private seven-node IPFS deployment at the Museum für Naturkunde Berlin, used to stage raw scientific acquisition data. Two real acquisition folders were recovered by content identifier, a large machine-learning model-weight folder was ingested, and node telemetry was logged over time. The authors conclude the prototype is a practical institution-local distributed file system for verifiable digital collection objects.
Joachim Snellings, Jasper Funk-Smit, Margot Belot et al.
In briefRewarding Informed Voting in Proof-of-Stake Governance
SummaryThe paper studies reward design for delegated representatives who pay a cost to gather information and vote on a binary governance proposal. It shows that under a non-neutral prior no payment rule makes sincere voting strictly optimal at every positive effort level, and fixed budgets limit accuracy in large symmetric committees. It compares proportional sharing with a conditional rule that rewards only members of a sufficiently large winning coalition, which makes sincere voting strictly optimal at symmetric positive effort. Under suitable reward-cost conditions the conditional rule yields higher effort and decision accuracy, though insincere equilibria may also exist.
In briefNavigation in cities using LEO signals of opportunity with a GLRT-assisted off-grid SBL method
SummaryBecause GNSS signals are unreliable in dense cities, this paper studies positioning using low earth orbit satellite signals of opportunity. It combines a generalized likelihood ratio test with off-grid sparse Bayesian learning to estimate line-of-sight delay and Doppler under multipath and non-Gaussian, correlated noise, then feeds the estimates into an extended Kalman filter for navigation. In simulations for urban settings, the method reduces delay RMSE by up to 77% and position RMSE by 88.5% compared with GLRT-based estimation.
Maedeh Sotoodenia, Mohammad Hossein Kahaei, Alireza Nezamalhosseini
In briefComputational Complexity of Mixed Equilibria in Correlated-Prior First-Price Auctions
SummaryThe paper shows that computing an approximate mixed Bayes-Nash equilibrium in a discrete first-price auction with correlated priors is PPAD-complete. A key intermediate step is proving that computing an approximate Nash equilibrium in a new normal-form game, the hypergraph discrete first-price auction, is also PPAD-complete. The authors note this game may be of independent interest.
In briefFailBench: Testing How Well Distributed AI Training Survives Machine Failures
SummaryThis paper introduces FailBench, an open evaluation tool that tests fault-tolerance methods across seven distributed deep learning training setups. It measures eight crash-recovery mechanisms and a no-protection baseline under single, concurrent, and cascading machine failures on an 8-GPU cluster, across 142 combinations. The authors find no single method is best everywhere, and that overhead varies widely with architecture, from 3.7% to 176% for in-memory replication. They also offer a decision framework for choosing mechanisms and release FailBench publicly.
WhySystematically compares eight fault-tolerance mechanisms across seven parallel architectures with 142 experimental configurations, yielding a practical selection framework for distributed training.
In briefCold-atom Rydberg array with time-space multiplexing for sensitive wireless receivers
SummaryThis paper studies a cold-atom Rydberg array (CoRA) receiver designed to avoid thermal noise, Doppler broadening, and photon shot noise found in hot-vapor atomic receivers. It proposes spatiotemporal multiplexing (STM) for gap-free symbol sampling and a joint binomial particle smoother (JBPS) for signal detection. Numerical results show that STM removes the temporal sampling bottleneck and JBPS lowers the bit error rate. The authors also report a tradeoff: more atoms reduce quantum projection noise but lower the sustainable packet repetition rate.
In briefJoint beamforming for continuous transmissive RIS in multiuser communications
SummaryThis paper studies multiuser downlink communications using a continuous transmissive reconfigurable intelligent surface (CT-RIS), whose passive beamforming is defined by a phase function over a continuous aperture. It jointly optimizes base station precoders and this phase function to maximize the weighted sum rate, using an alternating optimization algorithm. Numerical results show the proposed CT-RIS outperforms the considered benchmarks.
In briefTRANSIT: Transparent Scale-in for Multi-Node LLM Training
SummaryTRANSIT is a framework that lets multi-node model training run on fewer GPUs by transparently using CPU DRAM as an extension of GPU memory. It works through a user-space interposition layer and requires no changes to the application, training framework, scheduler, driver, or operating system. The authors evaluate it on dense and MoE models using up to 64 NVIDIA H100 GPUs over a RoCE network. They report higher per-GPU throughput than TorchTitan, ZeRO-Offload, and ZeRO-Infinity, and over 90% of baseline throughput with 50% fewer GPUs.
Hyungyo Kim, Nicholas Satchanov, Hrishi Shah et al.
In briefNCCL M2N: A Layout- and Topology-Aware Collective for Tensor Resharding
SummaryThis paper presents NCCL M2N, a collective communication primitive for moving distributed tensors between different layouts, deriving transfer regions and a schedule from source and destination meshes. In a single FFN-MoE layer transfer on 256 GB200 GPUs, it reports up to 7.9x speedup over flat direct sends. In a DeepSeek-V3 NeMo-RL experiment, weight-sync time drops from 5.78 s to 2.77 s and step time falls 12.7%.
Kaushik Kandadi, Youngeun Kwon, Sreeram Potluri et al.
In briefMosaic: GPU Sharing with Latency Guarantees via Kernel-Level Interference Prediction
SummaryThis paper presents Mosaic, a kernel-level predictor of interference between workloads sharing a GPU. It combines analytical and lightweight learned models to account for thread-block placement, memory hierarchy contention, and intra-SM resource contention. Across four GPU architectures, Mosaic reduces prediction error by up to an order of magnitude over prior predictors. The authors integrate it into the MosaicSched scheduler, which chooses between full-GPU colocation and SM partitioning to maximize best-effort throughput while keeping p99 latency at or near the SLO target.
In briefProductivity in HPC: A Critical Review of Measurement Methods, from SLOC to AI-Assisted Programming
SummaryThis paper critically reviews how productivity is measured in high-performance computing, tracing methods from source lines of code and static complexity to empirical usability studies and development-time models. It also examines how automated migration and AI-assisted programming weaken the link between code changes and human effort, shifting work toward verification, correction, and performance tuning. The reviewed methods are compared by what they measure, directness, objectivity, reproducibility, and relation to performance and portability. The paper concludes by identifying the main limitations and open challenges in quantified HPC productivity assessment.
In briefHigh-speed FPGA network processing for scientific data flows on a cloud testbed
SummaryThis paper presents an FPGA-based SmartNIC architecture for handling network demands in data-intensive scientific computing. It uses P4-programmable data-plane functions for packet parsing, marking, telemetry, traffic steering, and machine learning inference without host CPU involvement. The framework integrates the SciTags standard and counts packets by SciTag values carried in the IPv6 Flow Label. Experiments on AMD U280 FPGAs in the Open Cloud Testbed show line-rate handling with imperceptible latency while accurately monitoring scientific flows.
Miriam Leeser, Michael Zink, Suranga Handagala et al.
In briefOptimal codes for DNA-storage read counts, and how random codes compare
SummaryDNA storage recovers data by randomly sampling synthesized strands, so the number of reads needed for recovery matters. This paper studies the coverage depth problem: finding a linear code that minimizes the expected number of random column samples needed to span the full information space. The authors derive a general formula for this expectation and prove that the q-ary simplex code and q-ary Hamming code are optimal for the parameters where they exist. They also show that random linear codes have an additive gap from the MDS benchmark that stays bounded in the constant-rate regime, regardless of code length.
WhyDerives a general coverage depth formula and proves optimality of simplex and Hamming codes, a substantive theoretical advance with impact mainly within coding theory.
Roee Gross, Yitzchak Grunbaum, Matteo Bertuzzo et al.
In briefByzantine-tolerant causal unicast with constant message overhead
SummaryThis paper presents a Byzantine-tolerant causal ordering algorithm for point-to-point messages that achieves constant O(1) application message space overhead. It uses a Sender Permission to Send invariant and an Isolated-Buffer Optimistic Model, with per-peer queues to resist state-pinning and head-of-line blocking attacks. The authors formally prove system-wide liveness and a weakened causal safety property called Congestion-Relaxed Causal Delivery.
In briefChoosing Equilibria in Game Learning by Reshaping Learning Dynamics with Metrics
SummaryThis paper studies how state-dependent symmetric positive-definite metrics can change which equilibrium learning dynamics converge to, without moving the game's equilibria. The authors characterize when a rival equilibrium can be destabilized while keeping a designated one stable, and they derive a step-size condition for discrete-time implementation. The method is validated on a continuous-commitment stag hunt and an entropy-regularized iterated prisoner's dilemma, where it selects the desired equilibrium in both.
Karan Mahesh, Runyu Zhang, Michael R. Benjamin et al.
In briefTrading with STARS: Algorithms and Performance Limits for Storage-Constrained Trading
SummaryThis paper studies an online trading problem in which a trader sees i.i.d. prices from a known distribution and must make irrevocable buy, sell, or hold decisions under storage constraints. Performance is measured by regret, the gap between the expected profit of a hindsight-optimal policy and the online algorithm. The authors introduce a parameter β describing how price mass concentrates around the median and derive regret lower bounds depending on it. They propose STARS, which simulates future price scenarios to approximate the value-to-go function, and show it achieves near-optimal regret across a broad range of distributions.
In briefUsing information theory to reconstruct the structure of undeciphered writings
SummaryThe paper extends an earlier framework for recovering geometry and topology from non-random information to undeciphered records such as the Phaistos Disc, Andean khipu, Rongorongo, Indus inscriptions and the Voynich manuscript, with Egyptian layouts and the Arecibo message as controls. It uses algorithmic information dynamics, combining information measures, compression and algorithmic-probability estimates. The authors report structural findings, such as recurrence explained by repetition and khipu knot patterns, and distinguish recoverable organisation from semantic decipherment.
In briefOfan: a destination-aware load-balancing scheme for AI training traffic
SummaryThis paper addresses the problem that existing packet spraying load-balancing algorithms struggle with AI workloads sent at full line rate. The authors show that on a fat tree, the downward path is fixed once an upward path is chosen, so destination-oblivious schemes cannot correct the imbalance they create. They propose Ofan, a switch-based destination-aware scheme, and pOfan, a variant fitting current switch pipelines, with a P4 implementation and an end-to-end FSDP2 evaluation on Llama-3 405B.
WhyProposes destination-aware load balancing, proves $\Theta(\sqrt{m})$ queueing for prior schemes and $O(1)$ for Ofan, with P4 and end-to-end LLM training evaluation.
In briefCUDA-MPQS: A fully GPU-based quadratic sieve that factored RSA-155
SummaryThe paper presents CUDA-MPQS, a self-initializing quadratic sieve in which every stage, from polynomial initialization through sieving, linear algebra and the square root, runs on the GPU. Using it, the authors factored the 512-bit RSA-155 modulus end to end, using 700.6 GPU-hours and about 242 kWh, with 98.4% spent on sieving. They state this is the largest integer factored by the quadratic sieve, exceeding the prior RSA-150 record.
WhyMoving every quadratic sieve stage onto GPUs and factoring RSA-155 is novel; comparisons are thorough, but impact stays within one algorithm family.
In briefRevisiting the Hardness of Dominant-Strategy Mechanism Design (limits on truthful auction approximation with limited communication)
SummaryThe paper studies how much communication dominant-strategy truthful mechanisms need for combinatorial auctions and how well they can approximate optimal welfare. It proves approximation lower bounds for general monotone, XOS, and gross substitutes valuations, answering an open question for gross substitutes. It also gives results for two bidders, including a polynomial-communication universally truthful approximation for submodular valuations.
WhyGives a unified tight lower bound and resolves an open GS question, with new two-bidder approximation results via rigorous proofs.
In briefSymmetric Grassmann formulas for subspace dimension loss
SummaryThis paper studies the dimension defect of finitely many subspaces over an arbitrary field. It derives a symmetric, nonrecursive Grassmann-type formula for the dimension of their sum, expressed through intersections of partial sums. The authors prove the resulting dimension-defect profile is monotone and relate its gaps to discrete curvatures of the average-rank profile.
In briefOne Global Beam Across Many GPUs: High-Throughput Beam Search at Billion-Record Scale
SummaryThis paper shows how multiple GPUs can run a single beam search even when the retained candidate set does not fit on one device. States and candidates stay on the GPUs, while the CPU handles only small control and ancestry records. The authors prove that a distributed pipeline returns the same top-B results as a monolithic search, under stated conditions on candidates, scoring, tie ordering, and one global reduction per key. On eight H200 GPUs, a Cube4 run completed depth 8 at an effective beam width of 2,900,361,216 in 931.266 seconds, a derived rate of about 74.746 million pairs per second.
In briefOrigaMIG: MIG-Aware VM Placement with Local Optimization and Live Migration
SummaryThis paper addresses fragmentation and poor resource use caused by NVIDIA Multi-Instance GPU (MIG) placement rules in cloud data centers. It formulates MIG-aware virtual machine placement as a binary integer linear programming problem that aims to maximize accepted requests, consolidate resources, and reduce migration overhead. The proposed OrigaMIG places arriving requests with a fast greedy rule and invokes the solver on a small neighborhood of physical machines only when needed, executing the result via live migration. In simulations on 128 physical machines with 256 A100 and A30 GPUs, it keeps fewer machines active and migrates 29% to 49% fewer VMs than a stronger baseline policy.
In briefHybrid classical-quantum software solutions to speed up quantum computing adoption
SummaryThis paper examines the software side of hybrid classical-quantum computing, reviewing existing cloud applications, middleware frameworks, and API gateways that let classical and quantum systems work together. It offers practitioners a curated list of algorithms, full-stack libraries, simulators, and cloud providers, and outlines a research agenda and near-term development priorities for hybrid software engineering. The authors argue that because quantum execution depends heavily on classical pre- and post-processing, hybrid computing may become a lasting outcome rather than a temporary workaround.
Advait Deshpande, Luciano Baresi, Arosha K. Bandara
In briefOn the Fragility of Efficiency: Learning Without Coordination When One Player Deviates
SummaryThis paper studies whether players in games can reach efficient, welfare-maximizing outcomes using only their own past actions and payoffs, without communication. The authors test a known welfare-maximizing algorithm when one player deviates, either by playing a fixed action, playing randomly, or strategically switching algorithms. They find that a single deviant can substantially shift long-run outcomes away from efficiency, and that no completely uncoupled algorithm is robust to such deviants in every game.
WhyShows a single deviant can break welfare-maximizing uncoupled learning, and proves no uncoupled algorithm is robust across all games, with substantive theoretical and empirical support.
In briefTighter Upper Bound for Random-Offerer Bilateral Trade via Recursive Hard Instances
SummaryThis paper studies the worst-case efficiency of the random-offerer mechanism in bilateral trade relative to first-best gains from trade. The authors construct independent buyer value and seller cost distributions whose ratio converges to 0.436943488..., improving the previous upper bound of 0.460242308.... The construction combines recursively nested seller distributions with downward transport of buyer values, and the limiting constant is characterized analytically through a recurrence.
WhyRecursive hard instances tighten the random-offerer efficiency bound from 0.4602 to 0.4369, with an analytic characterization of the limit.
In briefLempel-Ziv Parsing with Length-Dependent Windows: Universality and Limits
SummaryLZ77 coders use the same number of bits for every match distance, even though short matches at long distances rarely pay off. This paper studies windows whose reach depends on match length, so a match of length ℓ can refer back at most W(ℓ) symbols. It bounds the cost of this scheme and shows that under suitable conditions greedy parsing keeps LZ77's convergence rates and universality. It also discusses the leading redundancy term and gains from related coding refinements.
WhyProves universality and convergence rates for greedy parsing with length-dependent windows, plus tight redundancy lower bounds; strong theory with impact mainly within information theory.
