In briefEntanglement-swapping quantum key distribution over MIMO free-space optical links
SummaryThis paper analyzes a quantum communication system that uses polarization-encoded DV-QKD with entanglement swapping over MIMO free-space optical channels, aiming to extend secure range beyond a direct link. The channels are decomposed into parallel eigenchannels via singular value decomposition, and closed-form expressions for QBER and secret key rate are derived. Numerical results show that entanglement swapping keeps QBER low over longer distances than direct transmission, at the cost of lower key rate at short range, with the crossover distance growing as MIMO configurations increase.
Sushil Kumar, Soumya P. Dash, and George C. Alexandropoulos
In briefBenchmarking Federated Learning Frameworks on Medical Image Classification
SummaryThis paper compares three federated learning frameworks, Flower, NVIDIA FLARE, and FedBioMed, using ResNet-10 on the MedNIST image classification task in a single-host containerized setup. Across 81 executions, the authors measured execution time, CPU, RAM, GPU utilization, network traffic, and final accuracy. Significant differences were found in server memory, GPU utilization, network traffic, and execution time, while no significant accuracy difference appeared within the short training schedules tested.
In briefStudy of whether first-order learning methods converge in signaling games
SummaryThis paper studies how learning algorithms behave in Bayesian signaling games, a simple model where a Sender with private information sends a signal and a Receiver updates beliefs and responds. The authors find that standard online learning methods reliably converge to strict Perfect Bayesian Equilibria but not to non-strict ones. They show that finite signaling games have a PE that is variationally stable exactly when it is strict, which yields local convergence guarantees near such equilibria.
WhyThe authors characterize variational stability of PE in signaling games as equivalent to strictness, and report convergence results plus experiments across several first-order algorithms.
In briefOneReel: Efficient Multi-Scene Video Generation via Scene-Structured Attention
SummaryThis paper presents OneReel, a training-free system for generating videos with multiple scenes. It keeps full attention within each scene and limits information exchange between scenes to a bounded shared memory, reducing attention cost from quadratic to linear in the number of scenes. Across four video generation models, it reports 1.7-4.9x end-to-end speedup over dense attention and 34-37 point gains in text-following accuracy while maintaining cross-scene consistency.
In briefLearning-assisted beamforming optimization for hybrid XL-RIS communication systems
SummaryThis paper studies a downlink multi-user system in the near field that uses a hybrid active-passive extremely large-scale reconfigurable intelligent surface (XL-RIS). It jointly optimizes the base station precoder, discrete phase shifts, and active-element gains using a discrete coordinate ascent-based alternating optimization framework. A low-complexity deep unfolding alternative is also proposed, which achieves sum-rate performance close to the alternating optimization scheme with substantially lower and nearly constant runtime. Both schemes outperform manifold optimization under discrete hardware constraints.
In briefA game-theory model for small farmers joining carbon farming together
SummaryThis paper proposes a cooperative game-theoretic framework to study how smallholder farmers can join carbon farming through coalitions. It combines optimizing individual farmer portfolios, optimizing joint coalition portfolios, and allocating surplus using concepts such as the Shapley value and core. The authors release CarbonFARM-Bench, an expert-built dataset, and report computational experiments showing solo participation is largely infeasible while coalitions create substantial value.
Geetha Charan, Thandava Sunkara, Rohit P Suresh et al.
In briefMeasuring how much freedom a multi-robot team can move with and be observed
SummaryThis paper proposes a mobility information capacity framework that measures how much maneuver freedom a multi-robot system can use while its motion remains observable, defined as the mutual information between a maneuver sequence and its noisy trajectory observation. For linear Gaussian models with feedback, it derives finite-horizon and stationary long-horizon capacity limits and converts workspace, separation, and chance constraints into maneuver covariance restrictions. An aerial-robot simulation evaluates spectral convergence, safety-aware allocation, finite-codebook identification, and fleet-size scaling.
In briefRobust analysis of multi-unit first-price auctions with large supply, measuring efficiency loss
SummaryThis paper studies how efficient multi-unit first-price auctions are, where bidders have independent private values and want one unit each. It uses a smoothness framework to measure the price of anarchy when supply is large. It shows that worst-case bid distributions cannot occur in equilibrium under a mild density lower bound on values, yielding a new asymptotic lower bound of 1-Θ(ε log(1/ε)) for ε-approximate equilibria.
WhyThe authors show worst-case bid distributions cannot arise in equilibrium under a density lower bound, yielding asymptotic tight bounds beyond the classic 1-1/e.
In briefProportional Elections with Feasibility Limits and Ranked Ballots: A New Solid Coalition Rule
SummaryThis paper studies elections where voters submit ranked ballots and feasibility constraints specify which sets of candidates may be selected. The authors introduce the first proportionality axiom for this general setting and extend the Solid Coalition Refinement (SCR) rule to general constraints. They show the rule satisfies the axiom exactly under matroid constraints, discuss its computational implementation and limitations for approval ballots, and provide its first empirical evaluation.
Grzegorz Kwacz, Georgios Papasotiropoulos, Piotr Skowron
In briefAsymptotic Analysis of Data Deduplication Under a Constant Edit Model
SummaryThis paper studies data deduplication algorithms from an information-theoretic perspective, focusing on data streams where repeats are approximate rather than exact. It considers two probabilistic source models in which repeated blocks are modified by a constant number of substitution edits or burst-insertion edits per block. For both models, variable-length deduplication is shown to achieve asymptotically arbitrarily large compression ratios, and multi-chunk deduplication achieves compression within a constant factor of the source entropy under mild conditions.
In briefA cheap way to keep two bandit players in sync, explained as a communication problem
SummaryThis paper studies how to synchronize two players in a distributed multi-armed bandit setting. It treats synchronization as a communication problem and analyzes the probability of synchronization error. The authors derive error exponents for the maximum-likelihood decoder using a random coding ensemble, and introduce a universal decoder that achieves the same exponents. They also apply these results to bound the regret of their synchronization scheme.
Raymond Zhang, Henrique Miyamoto, Richard Combes et al.
In briefCompetition and convergence among LLMs on leaderboards
SummaryThe paper models LLM leaderboard competition as a game in which developers allocate resources across capabilities to maximize win rate. It shows that similar production functions can produce similar capability profiles, while comparative advantage can sustain distinct ones, and that diminishing returns make the competitive frontier increasingly crowded. Using LM Arena and OpenRouter data, the authors find developers improve most where they lag competitors while retaining category-specific advantages.
In briefRunning FP32 and FP64 units concurrently on GPUs, and mixed-precision placement in WENO reconstruction
SummaryThis paper investigates whether idle native FP64 units on NVIDIA GPUs can run alongside FP32 work, and compares precision placements for WENO-Z reconstruction. Overlap is substantial on A100 but small on GH200, GB200, and RTX 6000 Blackwell. For a one-dimensional entropy wave problem, the W32P64 placement fails to reach the design order for WENO7 and WENO9, while double-float emulation stays close to design order and gives a 2.51 to 3.94 times speedup over FP64 on RTX 6000 Blackwell.
Jun Hatayama, Daichi Mukunoki, Yusuke Onishi et al.
In briefFast, practical classification of HPC I/O time series for scheduling
SummaryThis paper introduces SAGE, a method that splits HPC job I/O traces into short windows, learns a symbol alphabet, and clusters the resulting symbol sequences to characterize temporal I/O behavior. Evaluated on 819,771 real traces from two HPC systems, SAGE-2 lowers median intra-cluster metrics by 9.1% versus K-shape while cutting clustering time from 3.8 hours to 2 minutes. Using cluster centroids for I/O scheduling reduces I/O stretch by up to 24% compared with round-robin placement.
Maxime Gonthier, Luan Teylo, Francesco Antici (RIKEN CCS) et al.
In briefCapacity of distributed linear computation: the geometry of relations
SummaryThis paper studies distributed linear computation, where several users observe linear functions of common finite-field data and send messages to a server that computes another linear function. It characterizes the generic capacity, the minimum total number of symbols sent, when coefficients are random and fixed during coding and the field size grows. The capacity curve's shape is determined by the geometry of linear relations among the observations and the demand.
In briefTPIC: Energy-conserving space plasma simulations on GPU supercomputers
SummaryThis paper presents TPIC, a massively parallel GPU implementation of the Energy Conserving Semi-Implicit Method for kinetic space plasma simulations. The authors extend its exact energy conservation to closed domains with conducting boundaries by deriving and canceling a spurious term in the discrete Poynting flux theorem. Across several electrostatic and electrodynamic test cases, TPIC matches analytical or reference results and conserves total energy to machine precision. On 4,096 AMD GPUs, it advances 1.9 trillion particles with 96.8% weak scaling efficiency.
