In briefError-correcting codes with high girth built from higher-degree permutation polynomials
SummaryThis paper studies quasi-cyclic LDPC codes, whose girth and minimum distance are limited by the commutativity of their circulant permutation matrices. The authors build protograph-based codes from higher-degree permutation polynomials, focusing on a family of non-commuting binomial permutation polynomials. Simulations show better decoding performance than constructions based on affine permutation polynomials.
Tasawwar Hussain, Anthony G\'omez-Fonseca, and Tefjol Pllaha
In briefMultiple Access in Pinching-Antenna Systems: Survey and Impairment Evaluation
SummaryThis paper surveys multiple access techniques for pinching-antenna systems, which attach dielectric pinches to a waveguide to create nearby line-of-sight radiation points. It organizes fifty-six works into a unified taxonomy by spatial degrees of freedom, showing users on one waveguide can be separated only by time, frequency, power, or guided mode. The authors identify impairments that disrupt this separation and evaluate time, frequency, power, and space-domain access under identical models, finding the degradation depends on the method.
In briefRandomization Outperforms Deterministic Adaptive Parallel Sampling
SummaryThe paper studies parallel samplers that reveal each coordinate once and draw each batch independently from exact conditional marginals, under a three-round cap. For an explicit binary family, it shows randomization beats every deterministic adaptive policy: the best deterministic forward KL divergence grows linearly with dimension, while a uniform mixture of four fixed schedules is exponentially smaller. Selecting the schedule needs only two fair random bits. A finite-sum certificate confirms this separation already at dimension 389.
In briefVenusRL: A Fully Disaggregated Agentic RL System with Priority Scheduling
SummaryThis paper presents VenusRL, a fully disaggregated reinforcement learning system for training LLM agents through multi-turn tool interactions. It targets two bottlenecks: training throughput limited by the slowest trajectories, and tool sandboxes that are over-provisioned in memory. It uses priority-aware scheduling and a memory-aware page-sharing pool to address them. According to the paper, it achieves up to 4.24x end-to-end training speedup over baselines and cuts environment cost by up to 89%.
Mingjun Zhang (1), Yucheng Li (2), Menghao Zhang (2) et al.
In briefAnalysis of how well a multi-rate version of NOMA (a radio access method that shares one channel among users by power) keeps data errors low when there are many users and different symbol speeds
SummaryThis paper analytically studies the bit error rate of downlink multi-rate NOMA over Nakagami-m fading channels for any number of users and symbol rates. It derives an exact error expression for the first detected user and closed-form approximations for the others, assuming perfect successive interference cancellation. A symbol-pattern-dependent SINR analysis explains why MR-NOMA outperforms conventional single-rate NOMA, and an average BER minimization for power allocation is solved numerically and checked by simulation.
In briefHow prompt wording shapes LLMs' tendency to follow default options in group-decision games
SummaryThis paper tests whether large language models follow default values pre-filled in an interface when making choices in two one-shot social dilemmas: a common-pool resource extraction game and a threshold public-good game. Pre-filled defaults pull choices toward the default value, but the strength of this pull varies widely with wording, default type, and action-space granularity. The authors conclude that evaluating agent behaviour without controlling the choice architecture can miss a key source of variation.
In briefAFORE: Reconfiguring Experts Efficiently in Attention-FFN Disaggregated MoE Serving
SummaryThis paper addresses efficient serving of Mixture-of-Experts models, where uneven expert loads can slow the feed-forward stage in attention-FFN disaggregation. The authors present AFORE, which uses the upcoming microbatch expert distribution to plan expert migrations and overlaps those migrations with computation of earlier microbatches. Evaluated on a 110B-parameter MoE model over four dynamic workloads, AFORE improves output throughput by 10.1-17.6% and reduces P95 inter-token latency by 7.1-9.5% compared with the strongest baseline.
In briefClosing the Prediction Gap: Completing Machine Shape to Match Real Hardware Behavior
SummaryThis paper completes a device-shape description used to partition work across heterogeneous, multi-institution hardware ensembles without communication. It fills in five previously unmeasured fields, including memory capacity, power draw, execution unit capability, compiler directive configuration, and memory reservation overhead, and adds a Boolean reached(d) to explain an NVIDIA fp16 power shift. The fields are measured on five devices (A100, H100, V100, MI100, Max 1550), with reservation overhead ranging from about 0.004 to 0.030 and a residual of negative 9.8 percent for the excluded Triton autotuner.
In briefPaNGEA: A GPU-Based Parallel Method for Finding Good Solutions in Mixed-Integer Optimization
SummaryThis paper introduces PaNGEA, a primal heuristic designed to run efficiently on GPUs for mixed-integer optimization problems. It combines linear-relaxation solves with a batched local-search procedure and generates multiple restricted subproblems in parallel. On 283 MIPcc26 and MIPLIB instances, the single-node GPU version reduces the average gap integral by 8-18% compared with its CPU counterpart, and parallel exploration of multiple nodes reduces it by a further 12-19%.
In briefCoda: Using Admission Flexibility to Speed Up Coding-Agent Serving
SummaryCoding agents alternate between LLM inference and tool calls, producing long sessions with reusable KV states. The paper finds two inefficiencies in shared serving systems: KV states sit across storage tiers with unequal preparation costs, and requests with different context lengths decode poorly together. It presents Coda, which uses Tiered-Aging state admission and Compatibility-Aware execution admission, plus a routing layer for multi-worker setups. Coda improves output-token and SLO-compliant throughput by 20.3% and 70.5% on average over prior systems and vLLM.
In briefCoverage Depth in Distributed DNA Storage: When Sampling Cuts Sequencing Cost
SummaryThis paper studies how many DNA reads are needed to recover a full message in distributed DNA storage when strands are randomly sampled from containers. It derives exact formulas for recovery time for arbitrary linear codes and ordered partitions. The authors show MDS codes are optimal for each fixed partition and identify when distributed sampling reduces total reads rather than only parallelizing them.
WhyDerives exact recovery-time formulas, proves MDS codes optimal, and resolves a conjecture on simplex codes, a strong theoretical contribution for DNA storage.
In briefPaxosLease in Relativistic Inertial Frames
SummaryThe paper examines whether the PaxosLease distributed lease protocol remains safe under special relativity, where simultaneity is frame-dependent. It restates the safety property using light cones, requiring a lease expiry to lie in the causal past of any later acquisition by a different holder. The authors modify classical PaxosLease by adding a relativistic Doppler factor to its containment and quarantine rules.
In briefByteSplat: Faster distributed 3D Gaussian Splatting training by cutting GPU data movement
Summary3D Gaussian Splatting (3DGS) reconstructs photorealistic scenes, but training large scenes demands much memory and computation. ByteSplat is a distributed training framework that targets two bottlenecks: repeated off-chip memory access during rasterization and inter-GPU communication of partial gradients. It fuses forward and backward rasterization into one kernel, adds hardware-aware pruning, and removes zero gradients from communication. On six datasets, it cuts intra-GPU off-chip traffic by 63.4% and backward inter-GPU communication by 65.8%, reaching up to 6.1x training speedup on eight GPUs while preserving reconstruction quality.
In briefWhether OFDM is optimal for aperiodic ranging signals
SummaryThis paper studies the expected aperiodic autocorrelation sidelobes of random communication symbols sent through a unitary modulation matrix, for constellations such as uniform QAM and PSK. For each lag it determines the exact minimum expected sidelobe level and a unitary modulation that attains it. The full OFDM modulation attains this minimum if and only if the lag divides the block length N, and simultaneous optimality over all positive lags is shown to be impossible for N ≥ 3.
WhyDerives exact minimum expected sidelobe levels per lag with rigorous proofs, and resolves the full EISL conjecture for a broad class of constellations.
In briefAI-assisted GPU optimization of a variational method for few-body boson systems
SummaryThe stochastic variational method is a precise approach for quantum few-body systems in fields such as nuclear and atomic physics. The authors built a parallel code for helium-4 few-body clusters, optimized for NVIDIA GH200 and AMD MI300A GPUs, with the whole code written by an AI coding agent. They introduce a secular-equation method with the Gu–Eisenstat prescription for the generalized eigenvalue problem, and tune performance by batching small matrices and optimizing memory layout. For a five-body system with 4800 basis states, the tuned code is reported to run 16.2 and 14.7 times faster than the tuned CPU version, and 168 and 153 times faster than the first CPU implementation.
Noritaka Shimizu, Shigeyoshi Aoyama, Yusuke Tsunoda et al.
In briefServeTwin: A Benchmark-Validated Simulator for Exploring Distributed LLM Serving Designs
SummaryThis paper presents ServeTwin, a closed-loop simulator for evaluating distributed LLM serving designs without costly physical clusters. It couples specification-driven analytical timing with a stateful serving loop and can run unmodified serving benchmarks. Compared with real deployments, it reproduces a steady-state throughput-interactivity frontier with 3.6% mean error and predicts multi-turn performance with 9.9% error.
In briefAnnotation-driven migration of CUDA programs to Tenstorrent Blackhole hardware
SummaryThe paper presents an MLIR-based compiler that migrates statically shaped affine CUDA programs to the Tenstorrent Blackhole platform, deriving data placement, inter-core communication, and tile computation. Declarative annotations let developers specify choices the source leaves open, such as where compute and data-movement operations are placed and streaming granularity. On Gaussian elimination, a five-point stencil, and symmetric rank-k update, the best configurations reach a 4.2x speedup for BF16 Gaussian and 2.1 to 2.2x for FP32 Gaussian and Jacobi, relative to the compiler's defaults. The authors note that a choice's performance effect can reverse depending on surrounding policies, precision, and loop schedule, so alternative per-core realizations should be compared rather than policies selected independently.
In briefBeaver: Elastic GPU Sharing for ML and Latency-Critical vRAN Workloads
SummaryThis paper presents Beaver, a system that lets a latency-critical radio network workload and a throughput-oriented machine learning workload share the same GPU. It adjusts streaming multiprocessor allocation per time slot and rewrites ML kernels to yield memory bandwidth during the radio workload's memory-intensive phases. On an H200 GPU, the authors report that the radio workload's p99.9 latency stays within its 1.5ms uplink deadline while retaining 74% of Llama-3.3-70B serving throughput.
In briefExistence of Two-Sided Product Expanding Codes Using Rademacher Matrices
SummaryThe paper proves that two-sided product expanding codes exist without relying on explicit constructions of asymptotically optimal locally testable codes (c³-LTCs), unlike an earlier result by Kalachev and Panteleev. It shows that independent random linear codes over a sufficiently large prime field have constant two-sided product expansion with probability tending to one, for fixed numbers of component codes with rates bounded away from zero and one. The trade-off is that the field characteristic must grow with the block length, whereas the earlier work used extensions of F₂.
WhyReplaces explicit c^3-LTC constructions with Rademacher matrices and trace power bounds to prove existence of two-sided product expanding codes; novel technique with rigorous proof, but existence only
Eshan Chattopadhyay, Noam Ringach, Nicholas Spooner
In briefBurning Signed Graphs: Two competing fires spread on signed graphs, aiming to burn as many vertices of one colour as possible.
SummaryThe paper introduces a new graph burning model in which yellow and green fires ignite vertices of a signed graph and spread along edges, with an edge's sign determining whether the fire changes colour. Each step, a player ignites a new vertex in a chosen colour. The optimal achievable count, called the plurality number, is expressed through Hamming distances to a binary code associated with the graph. The authors determine exact plurality numbers for certain signed paths and prove hardness results for computing or approximating it.
Shaun Fallat, Himanshu Gupta, Kamyar Khodamoradi et al.
In briefOptimizing simulation and data pipelines for experiments on exascale computers
SummaryThis paper addresses how large experiments in high-energy physics and cosmology, which generate growing volumes of data, can use exascale computing resources efficiently. The authors propose a multi-stage task scheduling approach and demonstrate it on a Monte Carlo simulation pipeline from the Short-Baseline Near Detector (SBND). They also present theoretical and numerical tools for modeling task scheduling, which can help identify optimal strategies and choose resources for a given problem size.
Thomas Wester, Harikrishna Tummalapalli, Christine M. Simpson
In briefConditional Correctness in List Decoding: When a Later Second Codeword Boosts Confidence in the First
SummaryThis paper studies how the likelihood ranks of noise sequences in list decoding relate to the posterior probability that a decoding is correct. Under a random codebook model, the authors show that soft-output expressions for list decoding exactly match finite-blocklength conditional correctness probabilities, and they characterize the asymptotic reliability of the first list entry under hard-decision maximum likelihood decoding. The reliability depends only on the first, second, and last ranks. For list size two, a sufficiently late second entry can make the first entry asymptotically correct even when its rank alone suggests otherwise.
In briefTopology-aware integrated sensing, communication, and charging for large-scale low-altitude wireless networks
SummaryThis paper addresses low-altitude wireless networks that must support sensing, communication, and charging at once, which creates tightly coupled optimization problems. The authors propose a unified topology-aware framework that represents terrestrial and airborne devices as nodes and their interactions as edges in a bipartite graph. Based on this representation, they develop a low-complexity topology reconfiguration algorithm using a unified resource-adjustment rule. Simulations show better multi-functional coordination efficiency than state-of-the-art methods, with robustness and scalability in scenarios with many ground and airborne users.
In briefFrom Alert Floods to Precedence Forests: LOGOS, a Zero-Prior-Knowledge Incident Triage System
SummaryThis paper asks whether failure propagation can be isolated using only raw, unannotated text logs. The authors present LOGOS, an unsupervised system that uses entity-event co-occurrence and temporal precedence to compress millions of log lines into a compact precedence forest. Tested on 25 production enterprise outages and 12 open-source issues without any prior knowledge, it reports a median runtime of 4.5 minutes, a median 99.8% noise reduction, and a mean recall of 0.76.
In briefMoE-CORE: A system that coordinates expert offloading and storage to run large Mixture-of-Experts models on limited device memory.
SummaryThis paper presents MoE-CORE, a system for running Mixture-of-Experts (MoE) model inference when device memory is limited. It loads complete expert layers in alternating buffers during prefill and, during decoding, uses layer-wise cache sizing, routing-history-aware replacement, and cross-layer prefetching to reduce host-to-device transfers. The authors report substantially lower mean time per output token than the evaluated vLLM Prefetch configuration on DeepSeek-V4-Flash-W4A8 and GLM-5.2-W4A8C8.
In briefRecovery sets and service rates of codes built by Plotkin-type construction
SummaryThis paper studies service rate regions for binary codes used in distributed storage, where data can be recovered from several disjoint server groups. It shows how recovery sets of a Plotkin-type code arise from those of its underlying code, and uses recovery hypergraphs to bound service rates. The results are applied to first-order binary Reed-Muller codes, showing that service rates for R(1,m) can be derived recursively from the generator matrix of R(1,2).
