In briefTruthful 2-approximation auction for ad placement in AI-generated answers
SummaryThis paper studies auctions for sponsored content inserted at multiple points in AI-generated answers, where click-through rates depend on each advertiser-position pair. Under the cascade model, users read ads in order and stop after their first click, so both assignment and display order matter for welfare. The authors give a polynomial-time mechanism for arbitrary pair-specific click-through rates that is truthful and individually rational in expectation and achieves a 2-approximation to optimal cascade welfare.
WhyGives a truthful-in-expectation 2-approximation for general pair-specific CTRs, improving prior bounds, though validation is mainly synthetic.
In briefFlexible beamforming with a fluid antenna array, from discrete to continuous arrays
SummaryThe paper studies far-field beam synthesis in arbitrary directions using a planar fluid antenna array (FAA), which activates radiation ports from dense candidate positions within a finite region to gain geometry diversity. It proposes a coarse-to-fine scheme: an FFT-based phase retrieval and greedy port selection with least-squares weight initialization, followed by projected Adam-style joint refinement of port positions and weights. For the evaluated two-beam configuration, compared with a fixed array of the same aperture and active-port count, the mainlobe energy fraction rises from 35.73% to 48.92% and the peak sidelobe level drops from -2.30 dB to -3.83 dB.
In briefEarly selection for polar ensemble decoding with SC-style decoders
SummaryThis paper proposes early ensemble selection (EES), which ranks members of a parallel successive-cancellation-style decoder ensemble using intermediate decoding states and keeps a set number at predefined stages to reduce computation. Two selection rules are introduced: path-metric-based (PM-EES) and neural-network-based (NN-EES), and computational cost is characterized via stage-dependent LLR update counts. Simulations on the (128,64) extended BCH and RM(4,8) codes indicate that suitable schedules approach full-ensemble error rates with fewer LLR updates.
In briefNon-vanishing determinant space-time codes from abelian crossed product algebras
SummaryThe paper observes that all previously explicit non-cyclic crossed product space-time codes for more than four antennas rely on transcendental structure constants, so their coding gain is not bounded away from zero. The authors introduce abelian crossed products with a commutator twist, give a parametric form of the codewords as a 2×2 block matrix of cyclic-algebra codewords, and prove the division property. This yields algebraic codes over Q(i) with non-vanishing determinants and closed-form minimum determinants, including a 4×4 (Z/2)^2 code and a family of 8×8 Z/4×Z/2 codes.
WhyProves division property and closed-form minimum determinants for new abelian crossed product codes, with an infinite DMT-optimal family; advances theory, impact confined to space-time coding.
In briefDo bio-inspired algorithms really save energy in data centres? This paper separates server packing from migration search to find out.
SummaryThe paper reproduces an existing virtual machine consolidation method in CloudSim Plus and adds a matched control that keeps the same initial packing and empty-host shutdown but drops migration search. Ant colony optimisation, particle swarm optimisation and a modified genetic algorithm are tested across two fleet configurations and four workload levels. The authors find that most measured energy savings come from initial packing and host deactivation, and that metaheuristic search more often increases energy use than reduces it.
In briefRadio signal point processes with a reflecting wall: how reflections change the signal pattern
SummaryThe paper studies how a long reflecting wall gives each wireless transmitter a mirror image, so direct and reflected paths come in pairs. It compares two models: one that keeps the powers separate and superposes the point processes, and one that adds the powers per transmitter. For large thresholds it derives the limit of the ratio of their mean measures. It also covers count fluctuations for Poisson transmitters and the probability that the strongest path is direct.
In briefSecure integrated sensing and communication using movable beyond-diagonal RIS
SummaryThis paper studies integrated sensing and communication where the sensed target is also an eavesdropper, so the signal used for sensing overlaps with the leakage that secrecy must suppress. The authors show that a movable-element beyond-diagonal RIS (BD-RIS) can separate the two tasks in a monostatic setting, and they give a closed-form zero-forcing design. In simulations, the dual-stream design loses only 1.1 bit/s/Hz when the sensing requirement grows by 30 dB, compared with 8.3 bit/s/Hz for a single stream.
In briefMDS codes that contain a fixed smaller code over field extensions
SummaryThe paper studies [n,k] maximum distance separable (MDS) codes over finite extensions of F_q that contain the scalar extension of a fixed (k-1)-dimensional code U. An auxiliary linear code built from U has full-weight words that characterize the MDS completions. U admits such a completion over some finite extension exactly when d(U) ≥ n-k+1. A formula counts these MDS codes over each extension field.
In briefA class of Niho-type four-valued cross-correlation functions for sequences of different periods
SummaryThis paper studies the cross-correlation between a binary m-sequence of period 2^n−1 and its short-period decimated sequences, of which there are 2^gcd(s,k)+1. Under the conditions n=2k, 1<s<n and v_2(s)>v_2(k), the authors prove the cross-correlation function takes only four values and completely determine their distribution. The result includes several previously known binary cross-correlation functions as special cases and is a theoretical contribution to sequence design.
In briefInferScale: Kubernetes-native predictive autoscaling that keeps AI inference services within their latency targets
SummaryThis paper presents InferScale, an autoscaler built into Kubernetes that uses inference-aware signals, short-term demand prediction, and tenant-priority headroom to scale AI inference services. On Azure Kubernetes Service, it raised SLO compliance for a ResNet-50 workload with a 21-second cold start from 47.8% to 65.3%. In multi-tenant contention tests, it reduced variability in Gold-tier SLO results. The authors conclude that in startup-sensitive serving, prediction matters mainly because it lets the controller act early enough to offset cold-start delay.
In briefSGL: A Bridging Model for Heterogeneous Hierarchical Parallel Platforms
SummaryThis paper proposes SGL, a simple imperative Scatter-Gather language and parallel execution model. It builds on experience with bulk-synchronous parallel (BSP) programming and targets multi-level, heterogeneous architectures such as multi-core processors, graphics accelerators and hierarchical routing networks. SGL is consistent with Valiant's Multi-BSP model and offers a simpler interface than BSML, covering a large subset of BSP algorithms without complex message passing. It supports predictable, portable and scalable performance, and its explicit data distribution is intended to enable automatic or programmable load balancing.
In briefProof of the Golomb–Welch conjecture for lattice tilings
SummaryThis paper proves the Golomb–Welch conjecture for lattice tilings. It shows that for every n≥3 and r≥2, the integer lattice Z^n has no lattice tiling by Lee balls of radius r. Combined with classical constructions for small n and r=1, this determines all parameters admitting a lattice Lee tiling.
WhyResolves the long-open Golomb-Welch conjecture with a complete proof, fully settling existence of lattice Lee tilings, a significant advance for the subfield.
In briefBalancing throughput and fairness in cell-free massive MIMO via percentile optimization
SummaryThis paper studies downlink power control in user-centric cell-free massive MIMO networks, maximizing the sum rate of the lowest q-percentile users, a utility that includes max-min and sum-rate maximization as special cases. The problem is non-convex, non-smooth and NP-hard, and the authors use a change of variables to restore block-concavity, yielding an algorithm that converges to a directional stationary point. Simulations show higher lower-percentile rates and improved fairness compared with the weighted MMSE benchmark, massive MIMO and small cells across tested access point densities.
WhyA change of variables restores block-concavity, yielding a provably convergent SLqP power control algorithm, compared against other architectures.
In briefHow much is costly information worth when a group must cooperate to avoid collective loss?
SummaryThe paper studies an evolutionary game model of the collective-risk dilemma, where a group must reach an uncertain threshold to avoid loss. Individuals can pay a fixed cost to learn the threshold and condition their cooperation on it. Low-cost information raises group success, while high costs reverse this effect, and information is worth buying only when it is much cheaper than contributing. The authors also examine information spreading through networks and find that buyers can limit the spread of disinformation.
Elias Fern\'andez Domingos, Genki Ichinose, Wakaba Tateishi et al.
In briefPartial information decomposition when some sources are known, with a source chain rule and an impossibility result
SummaryThe paper studies how partial information decomposition, which splits the information several sources carry about a target into atoms, should behave when some sources are already known. It shows that two natural conditioning approaches agree atom by atom exactly when redundant information satisfies a source chain rule analogous to the chain rule for mutual information. For three or more sources, the authors prove that this source chain rule and local positivity cannot hold together in any partial information decomposition.
WhyIntroduces a source chain rule and proves a no-go theorem for three or more sources, a substantive theoretical advance with likely broad influence on PID research.
In briefZipFusion: Fusing lossless KV cache decompression with attention to speed up LLM decoding
SummaryLarge language models with long contexts and many concurrent requests put heavy demands on GPU memory and bandwidth from the key-value (KV) cache. ZipFusion lets attention read the losslessly compressed KV cache directly, reconstructing values in registers with simple bit operations to reduce GPU memory reads. On an NVIDIA H100, it reaches about 1.25x the decode attention speed of FlashInfer. With 8K context under the same GPU memory budget, it improves decode throughput by up to 26.3% over uncompressed SGLang.
