In briefProbabilistic construction of maximum h-scattered subspaces (a combinatorial structure in linear spaces over finite fields)
SummaryThis paper studies maximum h-scattered subspaces, linear subspaces over finite fields that are connected to coding-theory notions such as MRD codes. The authors build such subspaces from random matrices over a suitable algebra. They prove that they exist when the dimension r satisfies a stated inequality and (h+1) divides rn. This yields the existence of F_{q^n}-linear MRD codes for all sufficiently large lengths N.
WhyProbabilistic construction improves existence bounds for maximum h-scattered subspaces and yields MRD code existence with rigorous proofs, impacting coding theory.
In briefMigration as a Steady-State Runtime for Virtual Stationarity in Low-Earth Orbit
SummaryThis paper addresses the problem that satellites in low-Earth orbit pass overhead within minutes, so applications must keep handing off to successive satellites to stay logically near a fixed geographic area. The authors reframe live migration as a continuous steady-state execution model and propose a Virtual Stationarity Runtime (VSR) framework. Tests on physical heterogeneous hosts report 95th-percentile downtime below 6 milliseconds over 300 recurrent handovers, and a 1,600-satellite simulation confirms migrations complete within orbital handover windows.
Daniel Kleebinder, Alireza Furutanpey, Stefan Nastic
In briefBeyond Hotness: High-Resolution Memory Profiling of Cloud and LLM Workloads for Tiered Memory Using ARM SPE
SummaryThis paper presents a high-resolution memory characterization method for studying where multi-tier memory can suit data analytics and LLM workloads. The authors implement it on ARM processors using the Statistical Profiling Extension (SPE) and analyze eight memory-intensive workloads from CloudSuite and LLM applications. They identify region-specific access patterns, assess cloud-oriented cache hierarchies, and propose metrics based on memory idle periods and latency profiles to guide tiered memory data placement.
In briefTile-based decompression inside compiled GPU query engines, avoiding decompressed data in GPU memory
SummaryThis paper addresses GPU memory pressure in analytical queries by adding a compression subsystem to compiled GPU query engines that treats decompression as a first-class physical operator. The authors implement the design in Proteus, fusing cooperative tile decompression with tuple-at-a-time execution in one generated GPU kernel so decompressed columns are not materialized in HBM. Compared with HeavyDB and Sirius on GPU-resident data, the approach reports competitive query times, up to 4x lower resident memory footprint, and up to 6x lower memory-bandwidth consumption.
Antonio Boffa, Ananya Gupta, Hamish Nicholson et al.
In briefComparing RISC-V, Arm, and x86 chips for supercomputing workloads
SummaryThis paper compares the Sophgo SG2044 RISC-V processor with NVIDIA Grace and Intel Xeon 8480+ using a common software stack and methodology. Tests cover four HPC benchmarks and four applications, including LLM decoding. The SG2044 is 8 to 15 times slower in absolute performance than the Arm and x86 systems, but shows competitive scaling and energy efficiency on several workloads. The authors conclude that RISC-V processors are closing the gap with established architectures.
Emanuele Venieri, Giacomo Madella, Gabriele Ceccolini et al.
In briefGLP Networking API Specification: a technical spec for the network layer under madGLP
SummaryThis paper specifies the networking layer beneath madGLP, a deterministic runtime for Grassroots Logic Programs whose agents run on users' smartphones and communicate only by asynchronous message passing. The layer provides only what GLP cannot supply: the network stack, location service, application attestation and cryptographic primitives. It defines three transport underlays (BLE/PAN, LAN and WAN via rendezvous servers) under one interface, using mutually authenticated Noise sessions and attestation. A single message protocol supplies the fair, unordered, atomic and authenticated delivery that madGLP assumes.
In briefStudy showing asynchronous session processes with nested session types are Turing complete
SummaryThis paper studies the expressiveness of Asynchronous Session Processes (ASP), the process model tied to Nested Session Types (NSTs), which can describe nested recursive protocols beyond traditional regular session types. The authors prove the model is Turing complete by encoding two-counter machines into typed processes. The encoding uses a state-aware representation of counters and instructions governed by NSTs.
WhyProves asynchronous nested session processes Turing complete via a correct two-counter machine encoding, substantially advancing known expressiveness limits for this model.
In briefDistortion-rate outage probability for non-ergodic source classes
SummaryThe paper studies the notion of outage in source coding, a dual of the outage concept used for non-ergodic channels. When source statistics are not ergodic, the authors define outage, discuss its properties, consider the special case of compact source classes, and provide examples.
Dongmin Park, Ibrahim Abou-Faycal, Razane Tajeddine
In briefPareto-optimal Scheduling with Predictions: a paper on single-machine job scheduling that balances trust in predictions against worst-case performance.
SummaryThe paper studies preemptive single-machine scheduling where a prediction of job order or sizes is available, with the goal of minimizing total completion time. The authors exactly determine the best possible trade-off between consistency and robustness for deterministic algorithms. They give a greedy algorithm achieving this trade-off and prove a matching lower bound.
WhyExactly characterizes the optimal consistency-robustness tradeoff for deterministic algorithms with matching bounds, substantially improving prior Θ(1/ε) robustness with complete proofs.
In briefMetaDB: Bringing File-System Structure Back into Distributed Databases
SummaryThis paper presents MetaDB, a metadata-aware distributed transactional database that serves as the shared metadata layer for Baidu's large-scale storage systems. It addresses mismatches between general-purpose databases and file-system metadata workloads using tree-topology sharding, off-critical-path semantic processing, and a lifecycle-aware storage engine. The authors report 2.2x to 4.1x the throughput of their previous database with half as many metadata machines, and say it has run in production for over five years, scaling to over 1,000 metadata nodes in one cluster.
In briefA counterexample to Tokareva's conjecture on sums of two bent functions
SummaryThe paper studies whether Boolean functions of degree at most n/2 can be written as sums of two bent functions in even dimensions. Tokareva had conjectured this converse, which the authors disprove by showing that x1x2(x3x4+x5x6+x7x8) is not such a sum on F^8. They extend the construction to counterexamples in every even dimension n≥8. Using Shaporenko's reduction, they also give explicit counterexamples to Tokareva's derivative conjecture in every even dimension n≥10.
In briefCertificates for Ruling Out Syndrome Patterns in LDPC Codes
SummaryThis paper studies how to prove that an LDPC code's parity-check matrix has no set of at most t variable nodes leaving at most b checks unsatisfied, a question relevant to minimum-distance and trapping-set analysis. It expresses this impossibility as a polynomial system and gives certificates whose degree depends on how the cardinality constraints are encoded. A local encoding yields certificates of degree at most t + b + K + 1, independent of block length, while a standard counting encoding yields lower bounds linear in block length for several code families. All computed results are verified in exact rational arithmetic.
Malihe Aliasgari, Hassan Khodaiemehr, and Khadijeh Bagheri
In briefHierarchical scheduling that caps queue waits for bursty HPC jobs to keep access fair
SummaryShared HPC clusters face unfair queueing when a few heavy users submit bursts of jobs and occasional users wait with no predictable delay. The paper proposes HyDRA, a scheduler that assigns each job a deadline on arrival based on size, priorities, backlog and fairness position, then ranks jobs to bound waits. Evaluated on a year of ALCF Polaris job traces, it reports twice the project-level Jain's fairness of the best baseline, at a moderate cost in utilization.
Karaen Senthilkumar, Suman Raj, Maxime Gonthier et al.
In briefBest achievable error for locating closely spaced point sources
SummaryThe paper studies how accurately closely spaced point sources can be located from noisy superpositions of a known smooth response with unknown complex coefficients. Because relative phases can offset changes in location, local Fisher information alone does not determine the attainable error. For each fixed source count K≥2 the authors prove matching Bayes risk bounds across small separations and noise levels, and show that a projected least-squares estimator attains these orders without knowing the phases or prior.
WhyProves matching Bayes risk bounds for closely spaced source localization across separation and noise regimes, distinguishing odd and even source counts; impact largely confined to signal processing.
In briefDifferentiable combinatorial auction design using graph attention networks to balance revenue, welfare, and fairness
SummaryThis paper proposes a differentiable, constraint-aware framework for designing combinatorial auctions, aimed at allocating public infrastructure such as airport slots and energy resources. It learns approximately strategy-proof mechanisms end to end and enforces fractional feasibility by construction. The authors introduce the graph-based CAGraph model, with CANet and CAFormer as baselines, and a decomposition that maps fractional allocations to feasible discrete ones. In synthetic benchmarks and an airport slot-divestiture case study, the learned mechanisms achieve up to 73.5% and 180% higher expected revenue than VCG, with small empirical regret.
In briefRandomization for Balancing Performance Under Normal and Compromised Communication
SummaryThe paper studies multiagent systems where agents can share information, asking how much agents should randomize their action choices so performance stays good both when communication works and when it is compromised. The authors show that for any scenario, some level of randomization exists where efficiency under nominal communication is near-optimal while efficiency under compromised communication is strictly positive. They also find this guarantee weakens as agents must choose among more resources.
In briefPermutation and automorphism groups of univariate multiplicity codes
SummaryThe paper gives explicit characterizations of permutations and linear isometries that leave univariate multiplicity codes invariant, for any evaluation subset of a finite field, multiplicity, and dimension. Automorphisms are defined with respect to the NRT metric. For multiplicity one, the results recover those of classical Reed-Solomon codes under the Hamming metric. The authors also give an algebraic characterization of automorphisms for the folded Hamming metric.
Eduardo Camps Moreno, Hiram H. L\'opez, Umberto Mart\'inez-Pe\~{n}as et al.
In briefKVStream: Adaptive Parallel Storage for Key-Value Caches in Large-Scale LLM Serving
SummaryLong-context LLM serving reuses prefixes whose key-value cache exceeds GPU and host memory, so retrieving it from storage slows inference. The paper treats prefix retrieval as aggregate-bandwidth scheduling across storage tiers and presents KVStream, which batches reads, defers writes, replicates chunks across tiers, and lets free tiers adaptively claim unread chunks. Integrated into LMCache and tested on four models, six datasets, and up to 128 GPUs, it reaches within 5% of the bandwidth bound, improves offline throughput up to 1.49 times, and lowers mean time-to-first-token up to 6.1 times under bursty arrivals.
In briefStochastic geometry and optimization for data-driven base station planning
SummaryThis paper proposes a data-driven stochastic geometry approach that optimizes base station deployment intensity to maximize a network performance metric. It uses samples from an unknown user-location distribution and applies an inhomogeneous Poisson point process for coverage analysis. The authors characterize discretization and sampling errors and solve the empirical problem with stochastic mirror ascent. Numerical experiments with synthetic and real-world spatial demand data demonstrate the approach.
Abhishek K. Shukla, Abhishek K. Gupta, and Ketan Rajawat
In briefCheng-Murray Conjecture on Deep Holes of Full-Length Reed-Solomon Codes
SummaryThis paper studies deep holes of full-length Reed-Solomon codes, where a received word is a deep hole when its interpolation polynomial has degree equal to the code dimension. The authors prove the Cheng-Murray conjecture for all finite fields in the range 3≤k≤⌊q/2⌋, completing the classification in all nonexceptional dimensions. The proof combines tangent envelopes of marked arcs with local syndrome compatibility and combinatorial counting.
WhyResolves the remaining Cheng-Murray conjecture cases using tangent envelopes and syndrome compatibility, giving complete proofs with notable theoretical advance in coding theory.
In briefSharper bounds for the entropic central limit theorem under log-concavity (how close sums of random vectors get to Gaussian, measured by entropy)
SummaryThe paper studies the relative entropy between the normalized sum of independent random vectors drawn from a log-concave isotropic probability measure and the Gaussian measure. It proves a new upper bound in which the dependence on dimension is polylogarithmic rather than polynomial, improving the previous best known result. The proof combines the De Bruijn identity with Eldan's stochastic localization process.
In briefDeep holes of full-length Reed–Solomon codes: uniform bounds and exact classifications
SummaryThis paper studies deep holes of the full-length Reed–Solomon code RS_q(k), meaning received words whose interpolation polynomial has degree k+m. For m from 3 to 7 it classifies all such deep holes over every finite field. Odd m yield none in this range, while m=4 and m=6 have only the Segre sextic and Frobenius-power exceptions, occurring in dimension two in characteristic two. It also gives uniform bounds for arbitrary m≥3.
WhyGives complete deep-hole classifications for full-length RS codes with m from 3 to 7, using rigorous proofs and verified computations; impact confined to coding theory.
In briefProcrustes-Bering-Sheaft: A toolchain for automated model discovery and virtual crash tests of microservices
SummaryThis paper presents a toolchain that automates lightweight model discovery and virtual crash tests for microservices. Each test estimates operation availability by Monte Carlo sampling of service and communication failures. The toolchain has three components: Procrustes checks deployment and tracing evidence, Bering discovers dependencies, and Sheaft evaluates declared operation-success conditions. The workflow is demonstrated on one HTTP operation and the implementation is verified with analytical oracles and controlled tests.
In briefA Resource Allocation Tournament: a contest model where players split limited resources across head-to-head matches
SummaryThe paper introduces a round-robin simultaneous resource allocation contest in which players allocate limited resources in advance across matches against all opponents, with outcomes set by relative allocations. It compares a total points payoff with a winner-takes-all payoff. For three players, complete equilibrium characterizations are provided under both schemes. For four or more players, total points admits Blotto-type equilibria, while winner-takes-all is studied numerically, using a best-response dynamics algorithm with stochastic inertia to approximate equilibria for four and five players.
