In briefBenchmarking Content-Defined Chunking Algorithms for Data Deduplication
SummaryThis paper isolates and compares eight base content-defined chunking algorithms used in data deduplication across four realistic workloads, measuring throughput, deduplication ratio, mean chunk size, and chunk-size variance. It derives formulas linking parameters of AE, RAM, and MII to expected chunk size. The authors find a long-standing error in the widely used AE parameterization that yields chunks 13-39% smaller than intended and inflates AE's reported deduplication.
Marcel Gregoriadis, Leonhard Balduf, Bj\"orn Scheuermann et al.
In briefPrivacy mechanism design with χ² utility under a bounded α-lift leakage constraint
SummaryThis paper studies designing a statistical privacy mechanism under a bounded α-lift leakage constraint. An agent releases U based on useful data Y, which is correlated with private data X, aiming to maximize Pearson χ² information while satisfying a point-wise privacy constraint. The authors derive a closed-form quadratic surrogate, an explicit feasible binary mechanism, and utility-gap bounds, and extend the results to point-wise Rényi-divergence privacy with numerical evaluation.
Amirreza Zamani, Mohammad Amin Zarrabian, Parastoo Sadeghi et al.
In briefHow far Kolmogorov-Loveland randomness is from Martin-Löf randomness
SummaryThe paper studies maps called encodings that send basic sets to clopen sets while preserving uniform Lebesgue measure and disjointness. It asks whether a computable encoding can send every non-Martin-Löf-random sequence to sequences that fail a given randomness notion. The authors show no such encoding exists for partial computable randomness, but one exists for Kolmogorov-Loveland randomness, together with two non-monotonic betting strategies, at least one of which wins on every image sequence.
In briefNeural polar codes for source compression learned directly from data
SummaryThis paper proposes a data-driven approach to lossless and lossy source compression based on polar codes. Neural polar decoders are trained directly from data to approximate conditional distributions of polar coordinates, without requiring an explicit source model. These learned distributions are used to estimate entropy rates, mutual information, and rate-distortion functions, and to build explicit polar compression schemes. Experiments show that the learned decoders closely match analytical successive-cancellation decoders when those are available, and remain usable when conventional methods are computationally infeasible or the source model is unknown.
In briefCompileMask: A method balancing expressiveness and efficiency in attention masking
SummaryThe paper notes that modern attention workloads need diverse masks depending on runtime metadata, but efficient kernels support only a few patterns. It presents CompileMask, which uses an offline compiler to analyze visibility rules and generate mask-specific preprocessing kernels, with rules written in a region-based DSL called MaskSpec. At runtime, the generated kernel produces a compact Execution Descriptor so the attention kernel can check visibility by comparing element indices against precomputed endpoints. Experiments show up to 3.8× speedup over state-of-the-art methods.
In briefGoal-oriented ordered autoencoder for progressive, successively refined communication
SummaryThe paper proposes GO-OAE, a progressively refinable quantized latent sequence optimized for downstream task utility rather than source reconstruction. Inference can be made from any received prefix of the code index sequence, with accuracy improving as more indices arrive. On MiniImageNet it reaches over 96% of the lossless upper-bound classification accuracy at 324 bits per image, and over 94% of the bound using only 70% of the indices.
In briefEmber: a framework making GPU unified virtual memory practical for oversubscribed workloads
SummaryGPU working sets now exceed GPU memory, so applications must oversubscribe it and keep only hot data on the GPU. This paper presents Ember, a UVM management framework that uses a hybrid address space, counter-driven region promotion, and decay-based eager reclamation to improve performance without hints, profiling, or application changes. At 1.5x oversubscription, it is reported to be 6.8 to 10.6 times faster than UVM on six applications across three PCIe GPUs.
WhyAdds hybrid address space, counter-driven promotion and decay-based reclamation to UVM; six apps on three PCIe GPUs show 6.8-10.6x speedups, fairly solid evidence.
Senad-Leandro Lemes Galera, Michail Georgoulakis Misegiannis, Ferdinand Gruber et al.
In briefThree-layer precoding for wireless power and data transfer using reconfigurable parasitic antennas
SummaryThe paper proposes a coupling-aware tri-hybrid precoding architecture for simultaneous wireless information and power transfer using reconfigurable parasitic antennas. It adds a third precoding layer that accounts for mutual coupling and reactive energy storage in the antenna domain. A low-complexity optimization solution is presented, and numerical results show better spectral efficiency and harvested energy than hybrid precoding at a lower hardware cost.
Thien Duc Hua, Mohammadali Mohammadi, Hien Quoc Ngo et al.
In briefShoreKeeper: Faster cloud data analysis via distributed, recovery-cost-aware caching
SummaryThis paper addresses the heavy memory demands of large iterative graph and machine learning workloads, which continuously generate intermediate data. It introduces ShoreKeeper, a framework that treats cluster-wide memory as a unified, elastic pool and uses a global optimization model over a single graph abstraction to minimize the recovery cost of cache-evicted data. The system is designed to run on both highly available VMs and preemptible cloud instances. The authors report that experiments show ShoreKeeper consistently outperforms state-of-the-art systems.
In briefQuantum Annealing for Wireless Resource Allocation: A Guide to QUBO Modeling
SummaryThis paper provides guidelines for formulating wireless resource allocation problems as QUBO models for quantum annealing. It uses a small terrestrial-aerial example to show how a fractional SINR maximization objective is converted into a two-bit QUBO and decoded into a scheduling decision. The authors discuss encoding, constraint handling, hardware embedding and feasibility verification, and introduce a QUBO Compatibility Index as a diagnostic for formulation readiness. A case study on max-min uplink power control in cell-free massive MIMO is included.
Nguyen Xuan Tung, Le Tung Giang, Mai Dinh Cong et al.
In briefStacked metasurface-based multimodal feature fusion for collaborative edge inference
SummaryThis paper studies using a Stacked Intelligent Metasurface (SIM) as a physical computing unit for collaborative inference (SIM-CI) in 6G edge networks. The SIM receives over-the-air aggregated features from multiple users on the same spectrum and performs spatial-frequency combining and inference in a single wave propagation. The design adds an adapter for network heterogeneity and channel state information, along with a hybrid beamforming network. Experiments report 98.6% of the accuracy of conventional digital-domain collaborative inference, with 25% lower computational energy and 75% fewer transmission time slots.
In briefSnapShare: A deduplication system for serverless snapshots that keeps tenant data isolated
SummaryServerless functions start quickly by restoring virtual machine snapshots, but content deduplication can enable timing attacks that leak tenant data, so platforms like AWS Lambda enforce strict isolation. SnapShare shares a snapshot page only when its full content matches a provider-approved source, selectively deduplicating common infrastructure across cluster workers while keeping private data isolated. Implemented as an extension to Knative, it reduces tail request slowdowns by 10-29%, cuts remote storage by 39%, and lowers local cache footprints by 38% compared with traditional approaches.
Leonid Kondrashov, Likun Zhang, Andr\'e Filipe do Pilar de Jesus et al.
In briefAdaptive optimization of information freshness (Age of Information) when system statistics are unknown
SummaryThis paper studies how to tune the update generation rate to minimize the Age of Information, a metric of how fresh information is at a receiver, when system statistics are unknown. The authors use analytical expressions of the AoI gradient, reducing the learning problem to estimating a finite set of system quantities. They propose an online stochastic approximation algorithm that jointly estimates these quantities and optimizes the rate, and prove almost sure convergence to the optimal rate under mild conditions. The framework is demonstrated in four incomplete-information settings.
In briefA ray-traced digital twin for testing 5G positioning with OpenAirInterface
SummaryThis paper presents a ray-traced digital twin of the EURECOM GEO-5G testbed that runs an unmodified OpenAirInterface 5G stack, with per-antenna channels computed by Sionna-RT. Treating distributed base-station antennas as independent devices preserves the absolute delays needed for time-difference-of-arrival positioning, and the twin reaches metre-level accuracy consistent with the physical deployment. Using the twin, the authors find that widening the carrier reduces the tail of horizontal error without improving typical positions, while vertical error is limited by array geometry. A height-diverse antenna layout reduces the RMS height error by a factor of four.
Nima Chaharbaghi, Florian Kaltenberger, Bartosz Podrygajlo et al.
In briefSizing Fluid Antenna Receivers for Target Outage: SNR, Aperture, and Port Count
SummaryThis paper studies how much SNR, aperture, and port count a single-user downlink with a fluid antenna receiver needs to meet a target outage probability. For known spatially correlated Rayleigh channels and ideal port selection, it introduces a sampling-free residual certificate that controls reduced-rank outage error over a continuous SNR range. At 1 bit/s/Hz and outage 10^-5, a four-wavelength, ten-port receiver requires 4.40 to 4.55 dB, and reducing ports from 256 to ten costs 0.70 to 1.25 dB under a joint 95% probability statement.
In briefUsing Reconfigurable Optical Networks to Capture Electricity Price and Renewable Windows Across AI Data Centers
SummaryThis paper studies how AI data centers can migrate training workloads across regions in response to electricity prices and renewable-energy availability. It proposes a transfer-aware framework that jointly coordinates training placement, checkpoint transmission, and optical circuit switching. In a five-data-center benchmark, compared with static optical operation, the method reduces real-time grid purchase cost by 12.50% and increases local renewable-energy utilization by 2.724 percentage points.
In briefScheduling LLM pre-training loads in renewable-heavy AI data centers
SummaryThis paper proposes a dispatch framework for large language model pre-training in AI data centers powered mainly by renewable energy. It converts LLM workloads into daily checkpoint-based tasks and maps them to feasible 3D-parallel GPU modes under throughput, GPU, memory, and IT-power limits. In a 2048-GPU case study, checkpoint-aware 3D dispatch cut operating cost by 20.02% and reduced renewable curtailment from 1.077 to 0.014 MWh compared with fixed continuous training.
In briefBeyond Hypercontractivity: Stronger Communication Converses via Rényi Contraction
SummaryThis paper studies how strong data-processing inequalities for Rényi divergences relate to hypercontractivity. For Rényi divergences of order greater than one, it characterizes their contraction through a property of the dual Markov operator called power contraction. Using the contraction coefficient, the authors derive an upper bound on rectangle probabilities in terms of marginals that is at least as strong as the hypercontractivity-based bound. Combined with inversion of the binary Rényi divergence, this yields communication lower bounds for one-way communication-assisted agreement distillation, illustrated with binary symmetric and rare-event Z-channels.
In briefSlackSwap: Bandwidth-aware swapping and recomputation to cut GPU memory use in neural network training
SummaryThis paper presents SlackSwap, a method that treats the time available for moving data between GPU and host memory as a limited resource. It uses one profiled training iteration to decide which activations to swap to host memory and which to recompute. In trace-driven simulations on an A100 with BERT-Large, GPT-2 XL, and a Llama-7B-style model, it keeps 96-100% of unlimited-memory throughput while using half the activation memory.
In briefDynamic matching with abandonment: a simple policy reaches the optimal 0.5 guarantee
SummaryThis paper studies stochastic dynamic matching, where agents of different types arrive over time, may leave while waiting, and are paired for type-dependent rewards. The authors show that a simple static priority policy, computable in polynomial time, earns at least half of the omniscient benchmark on every instance. This establishes that the optimal competitive ratio is exactly 0.5, improving the previous best guarantee of 0.192 for the general model.
WhyResolves a long-open problem by establishing optimal competitive ratio 0.5 via a polynomial-time static priority policy with a new stochastic lower bound and complete proofs.
In briefHow network structure affects whether AI agents can curb deception: truth-telling versus investigative agents
SummaryThe paper extends an evolutionary game-theoretic Deceiver's Game to structured networks, where agents can deceive, tell the truth, or investigate. It introduces committed AI agents fixed to truth-telling or investigative strategies and runs agent-based simulations across three network topologies. Passive truth-telling agents can amplify adaptive deception, while investigative agents increase adaptive truth-telling in the oscillatory regime. Intervention effects are non-monotonic in deployment density and depend on network topology.
In briefCoexponential Exploration in the Clef Programming Language
SummaryThis paper describes the Clef concurrent language, whose compiler tracks state ownership, waiting computations and request dependencies. It models stateful services and client pools with a coexponential account and uses delimited continuations to retain what each interaction needs. The authors prove state-ownership and reply preservation and derive circular-wait exclusion under stated assumptions. These results aim to keep service composition correct when lowered to multiple hardware targets.
In briefA simpler explicit way to build smaller strong blocking sets
SummaryEarlier work by Alon, Bishnoi, Das and Neri built sequences of strong blocking sets whose sizes grow linearly in k, equivalent to minimal codes with positive rate. This paper uses concatenation of (q-1)-power intersecting algebraic geometric codes to give an explicit construction with size about 2(q²-1)k. The construction is better than the earlier result when q ≤ 11, but worse for large q. Its approach is simpler and does not use expander graphs.
In briefWeight enumerators of irreducible cyclic codes of length 2^ℓ
SummaryThe paper uses a recently found spectral relation between irreducible cyclic codes and generalized Paley graphs to study the weight distribution of irreducible cyclic codes of length 2^ℓ. These codes were earlier studied by Sharma, Bakshi and Raka, who split them into two families with two cases each. The authors identify the associated generalized Paley graphs as complete, semiprimitive, Hamming, or Cartesian products of semiprimitive graphs, and compute their spectra. This yields explicit expressions for the weight enumerators of all such codes.
In briefLow-complexity SVD precoding for zero-padded ISI channels, a fast transceiver design
SummaryThis paper proposes a low-complexity SVD-based transceiver for zero-padded memory-1 ISI channels. Adding one zero-padding symbol creates a channel Gram matrix with an exact symmetric tridiagonal Toeplitz structure, enabling FFT-based SVD precoding and equalization at O(n log n) complexity with a 1/n rate penalty. Simulations at code rates near 0.945 show about 5 dB gain over a TX-FFE/RX-DFE system with hard-decision Reed-Solomon decoding at FER 10^-5, and near-SOVA performance with LDPC decoding at high SNR.
In briefGame Petri Nets: Modeling Large-Scale Strategic Systems Compositionally
SummaryThis paper introduces game Petri nets, formalized as outcome-driven compositional game Petri nets (OCGPNs), for modeling and analyzing large-scale strategic systems such as maintenance workflows and cyber campaigns. Local games attached to places of a colored Petri net produce typed tokens that enable, synchronize, or block later games. The authors give operational semantics and invariant-based structural results for concurrent nets, and prove that global equilibria equal collections of local equilibria with continuation values for sequential nets. Two running examples illustrate the theory.
In briefEfficient constructions of spherical 5- and 7-designs in odd dimensions using balanced halves
SummaryThis paper constructs spherical 5- and 7-designs in odd dimensions by lifting balanced halves of even-dimensional designs onto several suitable latitude layers. For every odd dimension it gives an explicit spherical 5-design with at most 60(d²−1) points, achieving the optimal quadratic order in the dimension. Assuming Zauner's conjecture, this bound improves to 7.5(d−1)². For odd d ≥ 9, it also builds a spherical 7-design with 110M points from a complex projective 3-design with M points in CP^((d−3)/2).
In briefDetecting random geometric graphs with general kernels
SummaryThis paper studies how to distinguish a random geometric graph with a fixed kernel from an Erdős–Rényi graph with the same edge density. It characterizes the critical dimension for kernels that need not be differentiable at zero, and for smooth kernels whose first nonzero derivative at zero has order greater than one. Across these classes, signed triangle or quadrilateral counts reach the optimal detection threshold.
In briefCommitment Protocols over Gaussian Channels Built from Deterministic Identification Codes
SummaryThis paper links commitment (BC), a basic cryptographic primitive, to deterministic identification (DI), a post-Shannon communication setting. For additive white Gaussian noise channels, the authors show that reliable commitment protocols can be built from DI codes in both honest-but-curious and fully dishonest security models. They establish linearithmic-scale capacity lower bounds of 1/2 for the honest-but-curious case and 1/4 for the fully dishonest case.
In briefFeatureZ: A General Framework for Feature-Preserving Compression via Pointwise Bounds and Star Classification
SummaryThis paper introduces FeatureZ, a lossy compression framework for structured volumetric scalar fields that aims to preserve geometric and topological features such as isosurfaces, merge trees, and Morse-Smale complexes. It combines pointwise upper and lower bounds with consistent classification of each point's star in the underlying mesh. FeatureZ works as an augmentation layer on top of an existing lossy compressor, first applying quantization and then an iterative procedure to ensure consistent star classification. The authors report that it preserves diverse features with minimal overhead and achieves compression ratios comparable to or better than feature-specific methods.
In briefEvent-native immersive communication for 6G: a framework that transmits physical event information instead of raw sensor streams, with a new quality metric.
SummaryThe paper proposes an event-native immersive communication framework in which a service-relevant physical event is the communication object and event beliefs describe what transceivers know before and after wireless delivery. It introduces event-belief deficiency (EBD) as a quality-of-experience measure and derives log-det EBD, SNR threshold, outage probability and required transmit SNR for Gaussian beliefs over Rayleigh fading. Simulations show that stream NMSE can stay nearly unchanged across physically distinct events while log-det EBD tracks them, and event-native delivery reduces normalized log-det EBD by up to 63.6% and lowers required average received SNR by 3.47 dB.
In briefCodes That Protect Function Values Under the Lee Metric
SummaryThis paper introduces function-correcting Lee metric codes over Z_m that protect both the data and the values of a prescribed function against errors. It presents a general construction and derives lower and upper bounds on optimal redundancy, including a Plotkin-type lower bound. Explicit redundancy upper bounds are given for locally binary Lee functions, the Lee weight function, and the modular sum function. Since the Lee metric equals the Hamming metric over Z_2, the results also hold there for the Hamming metric.
In briefUnified Localization for Multi-Panel Massive MIMO with Clock Asynchronism: combines geometry models and channel charting to locate users.
SummaryThe paper proposes UNILocMP, a unified localization framework that combines model-based geometry with channel charting for multi-panel massive MIMO under clock asynchronism. Users with line-of-sight links to at least two panels get joint position and clock bias estimation, while single or non-LoS users are handled by an unsupervised channel charting model with two-stage data augmentation. Numerical results show it outperforms model-based and channel-charting baselines and achieves acceptable performance relative to fully supervised fingerprinting. For a fixed number of antennas, the multi-panel architecture improves accuracy and robustness over a single-panel base station, especially under clock asynchronism.
Yuhao Zhang, Guangjin Pan, Musa Furkan Keskin et al.
