In briefUniform-price mechanism design for automated market makers
SummaryThis paper studies batch double auctions in decentralized exchanges where an automated market maker (AMM) must keep post-trade reserve products at least as large as initial reserves. It proposes a randomized uniform-price mechanism based on binary quantity buckets, which randomly selects among three clearing modes and is ex-post feasible, individually rational, and weakly quantity incentive compatible. The mechanism achieves an expected volume of Ω(1/(1+log γ)²) times the optimal unconstrained volume minus O(q̄). The authors also show that full incentive compatibility forces a worst-case approximation ratio of O(1/γ).
In briefA Smaller Subspace Is Enough: Faster ISAC Beamforming Through Dimensionality Reduction
SummaryThis paper studies beamforming for integrated sensing and communication (ISAC) in monostatic sensing mode, a technology considered important for 6G networks. Because MIMO systems have many antennas, the optimization is computationally expensive. The authors show that the problem can be solved in an even lower-dimensional subspace than existing exact methods, with only a small loss in sensing performance. Numerical results compare sensing performance and computational complexity across the exact and smaller subspaces.
In briefConstructing new MRD constacyclic codes over inseparable field extensions
SummaryThe paper constructs MRD constacyclic codes of length p over an inseparable field extension L/K, where K has characteristic p. The authors state that these codes cannot be obtained from known constructions.
In briefBayesian learning of mean field game parameters under common noise, recovering unknown settings from partial random observations
SummaryThe paper studies how to infer unknown parameters of Mean Field Games from partial observations of population trajectories made random by common noise. The authors introduce the Bayesian-Extended Mean Field Game (BE-MFG), whose single Nash equilibrium encodes equilibrium policies for all parameters, enabling efficient learning of the conditional density of mean-field trajectories. On three synthetic environments, the method accurately recovers the ground-truth parameter from few partial stochastic observations, outperforming an adapted inverse reinforcement learning approach. An experiment on public electricity market data suggests it can also be applied to real-world data.
In briefGPU isotonic regression that merges convex hulls without copying vertices
SummaryThe paper presents a faster GPU method for isotonic regression, which finds the closest monotone sequence to data in weighted least squares. Instead of copying surviving hull vertices at each merge level, each chunk's hull is stored once and merged hulls are represented as segment lists. On dense hulls the merge phase is reported as 17 to 276 times faster than the copying approach, and the whole operator 3.9 to 19 times faster.
WhySegment-list hull merging avoids per-level vertex copying, with proofs and large GPU speedups over a copying baseline.
In briefFrom Border-Agnosticism to Causal Rewrite Systems: a theory paper reinterpreting borders in causal graph dynamics as missing information and recasting border-agnostic causal graph rewriting as a rewriting system to support reversibility.
SummaryThe paper builds on Causal Graph Rewriting, a recent extension of Causal Graph Dynamics that introduces controlled asynchronism into dynamical systems with dynamical graphs. It re-examines the notion of borders, which resemble dangling edges, and interprets them as missing information. This perspective adds structural rigor to the framework and allows border-agnostic causal graph rewriting to be reformulated as a rewriting system rather than a dynamical system. The authors present this shift as a step toward facilitating the treatment of reversibility.
In briefTrembling toward Duverger: how plurality voting converges or cycles
SummaryThis paper models a fixed, finite electorate in which voters see poll results and best respond while allowing for small chances that others deviate. It shows that a single update typically makes the dynamics settle on a stable profile where voters back one of the two poll leaders. It also shows that convergence can fail, with period-two and longer cycles possible, and that larger electorates allow longer cycles.
WhyProposes a finite-population trembling-hand dynamics model, proving one-update convergence and richer cycle structures; strong theory with impact mostly within theoretical game theory.
Maximilian Weigel, Damien Berriaud, Roger Wattenhofer
In briefUpper Limit on Code Rate for Insertion/Deletion List Decoding
SummaryThis paper derives an upper bound on the rate of linear codes that can be list decoded from insertion and deletion errors. It shows that, with a fixed list size L, a linear code can list decode from at most n − (L/(L+1))k + 1 insdel errors, and that this bound is strict when the code does not contain the all-ones vector. The authors also give a simple construction achieving the optimal rate for any list size while correcting a single insdel error. Results apply to linear codes over finite fields of size q ≥ 3.
In briefTiming gradient returns in over-the-air split learning on an aging single-antenna channel
SummaryThis paper studies over-the-air split learning over an aging single-antenna Rayleigh channel, where the mean returned gradient equals the antenna-domain gradient scaled by temporal correlation. Uncorrected training collapses in negative-correlation regions, and near zero correlation no scalar calibration improves the power ratio of a single transmission. Returning at the first correlation extremum with phase correction reaches 90.2% accuracy versus 91.8% with channel state information, at 3.5 GHz and 15 m/s in a tested line-of-sight setting.
In briefMoving Antennas Keep Gradients Intact for Learning on Mobile Devices
SummaryThe paper studies over-the-air training in which a moving terminal rotates the backpropagated gradient. A movable antenna that keeps pace with the terminal can rotate it back without channel estimation. The authors give a speed law and a bias-jitter law, and simulations on 3GPP CDL-A channels report 98.58% accuracy versus 98.59% with ideal CSI.
In briefCovert digital semantic communication using movable antennas for concealed, meaning-preserving wireless transmission
SummaryThis paper proposes a covert digital semantic communication system assisted by movable antennas, jointly designing the constellation, antenna positions, and semantic transceiver. The authors derive a closed-form maximum covert rate against an optimal detector with unknown noise level. A Gauss-Laguerre ring constellation and a covert symbol mapping are developed to carry semantics while staying hidden. Simulations report higher achievable covert rates and a 24% improvement in reconstruction quality over state-of-the-art semantic communication schemes.
In briefA unified geometry-based framework explaining how handover rates and sojourn times of aerial users in 3D cellular networks depend on network geometry and speed.
SummaryThis paper studies frequent handovers of aerial users in 3D cellular networks, modeling service regions as a stationary random tessellation and user motion with a cruise-turn model. Under stationarity, isotropy, and independence assumptions, it derives an exact handover rate law, H = (S_V/2)v̄, and an approximation for mean remaining sojourn time. The results are combined with link coverage to define a handover-aware effective spectral efficiency, applied to Poisson-Voronoi and Poisson-Delaunay models and checked numerically.
WhyLinks random tessellation boundary geometry with kinematics to derive an exact handover-rate law, extended to Poisson-Voronoi and Delaunay models; theory is strong, numerical validation is modest.
In briefScaleRAN: A Runtime for Elastic GPU Resource Management in AI-RAN
SummaryThis paper presents ScaleRAN, a node-local runtime that elastically manages GPU resources shared between radio access network (RAN) processing and AI workloads. It uses NVIDIA Green Contexts to repartition streaming multiprocessors at RAN slot granularity, with both reactive and predictive controllers driving the reallocations. The authors also introduce LDPCSim, a GPU-accelerated workload generator for evaluation. Results indicate that predictive fine-grained partitioning meets slot-level RAN deadlines while giving co-located AI workloads more GPU resources.
Arthur Hamard, J\'er\'emie Leguay, Gopalasingham Aravinthan et al.
In briefAchievable rate limits for interference networks using rotatable antennas
SummaryThis paper characterizes the achievable rate region of a rotatable antenna (RA)-assisted MISO interference channel. It exploits the extra spatial degrees of freedom from dynamically orienting RAs to boost desired signals and suppress interference, enlarging the rate region. The authors formulate a sum-rate maximization problem that jointly optimizes transmit beamforming and array rotation under mechanical constraints, and derive closed-form solutions for the two-receiver case. Numerical results show near-optimal performance and that RAs allow simple maximum ratio transmission beamforming with negligible loss.
In briefBaGuaLu-2: Handling Shifting Memory Demands to Train Mixture-of-Experts Models on Over 10,000 GPUs
SummaryThis paper presents BaGuaLu-2, a large-scale training system for Mixture-of-Experts language models. It replaces static memory allocation with fine-grained, on-demand provisioning of activation memory for each training batch, addressing memory spikes caused by uneven expert loads. The authors report training a 1040-billion-parameter model on 12,288 GPUs at over 2 EFlops and 35% MFU. They also state average throughput gains of 1.14x over Megatron-Kuai and 1.13x over AdaPipe.
In briefOn the Convergence of Iterated Round Robin
SummaryRound Robin is a simple, widely used method for allocating indivisible goods that guarantees envy-freeness up to one item under additive preferences. This paper studies an iterated version where agents observe outcomes and revise their reported preferences via best-response dynamics. With full information, cycles can occur even under strong restrictions, but restricting to item-preserving swaps guarantees convergence to a pure Nash equilibrium. With partial information, a more cautious agent behaviour makes iterated Round Robin strategyproof.
Maxime Lucet, Henri Surugue, Nawal Benabbou et al.
In briefIdentifying Signals in Multi-Antenna Gaussian Channels
SummaryThis paper studies how many messages can be reliably identified over discrete-time Gaussian MIMO channels with per-antenna, per-symbol peak-amplitude limits. It gives upper and lower bounds on identification capacity that grow linearly with the channel rank r. The number of identifiable messages can grow super-exponentially with blocklength n, as 2^((n log n)R), where R is an achievable rate. Higher channel rank yields a spatial-multiplexing gain, raising achievable rates and codebook growth.
In briefNew lower bounds for binary constant-weight codes via symmetry-based search, plus two quantum codes reaching upper bounds
SummaryThe authors give explicit binary constant-weight codes that raise lower bounds for A(n,6,w) in eight entries of Brouwer's table, released in two rounds in October 2026. Each code is a union of orbits under groups such as semilinear affine groups, and was verified by two separately written programs. They also report two qubit stabilizer codes, [[20,11,4]] and [[43,32,4]], which raise listed lower bounds to the listed upper bound of 4.
In briefMailRail: Reducing Incast Congestion with User-Space, Node-Level Flow Control
SummaryThis paper presents MailRail, an out-of-order, asynchronous messaging framework paired with a user-space, node-based credit flow control system. Instead of tracking credits per process, it uses shared per-node pools to reduce incast congestion caused by irregular workloads such as GNNs and Mixture-of-Experts models. According to the authors, the approach achieves up to a 1132x speedup over unconstrained execution at a 4096:1 incast ratio, with minimal overhead on uniform workloads.
Noel Gomillion, Lance Fletcher, Trevor Steil et al.
In briefUnderstanding KV Memory Retention in Large-Scale AI Agent Workloads
SummaryThis paper analyzes four production Microsoft M365 agent workload traces covering more than 121 million model calls. It finds that agent contexts grow quickly and that keeping their key-value state in GPU memory needs very large amounts of memory, though a one-hour idle cap cuts that ceiling about 3.6 times. Under tighter memory budgets, the authors' multi-resolution retention method, which demotes KV state to int4, outperforms plain eviction. They also report that routing a call to a different deployment sharply reduces context reuse.
WhyCharacterizes KV residency on large production agentic traces and proposes MRR, cutting recompute rates; substantial trace evidence and clear implications for inference system design.
In briefExplicit construction of optimal binary linear complementary pairs using Belov codes
SummaryThis paper studies linear complementary pairs (LCPs) of codes, a topic in coding theory. The authors construct additional optimal binary LCPs from Belov codes, a generalization of Solomon-Stiffler codes. The results extend Güneri's earlier construction. They also allow more flexible construction of optimal binary LCPs across a wide range of minimum distances.
In briefCausal information exchange as a necessary condition for adaptive regulation in life and agency
SummaryThe paper proposes a perturbational framework that separates causal transmission, directional imbalance, internal retention, response structure, and functional regulation. Randomised interventions define Shannon information channels, and Coding Theorem and Block Decomposition estimates describe response and trajectory structure. For a specified binary sensor–internal channel–actuator task, the authors derive a positive information lower bound from successful regulation and construct controllers with identical information profiles but opposite outcomes. The conclusion is that causal information is a necessary resource for this form of adaptive regulation, but neither causal coupling nor response complexity alone measures agency or life.
Brian LuValle, Abicumaran Uthamacumaran, Hector Zenil
In briefOptimal Event Curation: Mechanism Design for Balancing Participants Under Budget Balance
SummaryThis paper studies how to design a self-financing event that must weigh costly participants who add positive value against revenue-generating participants who may have a negative impact. The authors introduce a two-sided mechanism design framework that maximizes overall event quality under a budget-balanced constraint, so the designer does not run a deficit. For additive utility, they fully characterize the optimal incentive-compatible Bayesian mechanism under both ex-ante and ex-post budget balance. For submodular utility, they propose an ex-ante budget-balanced mechanism that achieves a constant-factor approximation to the optimal Bayesian mechanism.
MohammadTaghi Hajiaghayi, Sebastien Lahaie, Mohammad Mahdavi et al.
In briefSingle-letter characterization of rate-cost tradeoffs in static time-invariant control systems
SummaryThis paper studies the minimum communication rate needed to reach a given average control cost for static time-invariant control (STIC) systems, where the next state depends on history only through the latest action via a time-invariant kernel. Previously known formulations involved directed-information problems whose size grows with the time horizon. The authors show that for STIC systems these problems admit a single-letter form, giving a constant-size convex characterization that goes beyond linear quadratic Gaussian control.
WhyProves single-letter characterization of rate-cost tradeoffs for static time-invariant control, yielding the first constant-size convex characterization beyond LQG.
Eray Unsal Atay, Venkat Chandrasekaran, Victoria Kostina
In briefUAV rotatable antennas with rotation delay considered
SummaryThis paper studies how rotatable antennas on UAVs can improve wireless link performance by changing their boresight directions. Existing work usually assumes instantaneous reconfiguration, but mechanical rotation takes time in practice. The authors build a rotation-delay-aware model for finite time slots, revealing a tradeoff between gain from reorientation and lost transmission time. They propose a joint design of antenna orientation and beamforming to maximize the minimum effective rate, with a low-complexity algorithm for the multi-user case.
In briefAccelerated Dataflow: Scaling Efficient Inference in Streaming Data Analytics
SummaryThis paper presents a system on Google Cloud Dataflow for streaming pipelines that combine data processing with machine learning inference, such as large language model calls. It uses autoscaling co-designed for heterogeneous worker pools and prefix-aware batching. Compared with existing streaming data processing systems, the authors report 1.5x to 2x higher throughput at similar latency, up to 75% peak resource savings, and 84% lower peak tail latency.
RuiLong Jiang, Dustin Rhodes, Danny McCormick et al.
In briefFaster distributed edge coloring for bipartite graphs using (Δ+1) colors
SummaryThis paper studies distributed (Δ+1) edge coloring in the LOCAL model on bipartite graphs with n vertices. The authors give a deterministic algorithm running in O(f(Δ) log n) rounds, which yields O(log n) rounds for bipartite graphs of constant degree. These bounds improve on earlier results that hold for general graphs and also work for a local-palette version of Vizing's theorem.
WhyImproves deterministic distributed bipartite edge coloring from polylog to O(log n) rounds, with proofs for a local-palette variant; theoretical advance.
In briefFresh Information Is Not Enough: Age-of-Actuation-Optimal Scheduling for Energy-Harvesting Actuators
SummaryThis paper studies transmission scheduling for an energy-harvesting actuation system, where actuation needs both a received update and available energy. The authors formulate minimizing average age of actuation (AoA), the time since the latest completed actuation, as a Markov decision process and show that an optimal policy has a threshold property. They also derive the average AoA in closed form under the always-transmit AoI-optimal policy. Numerical results indicate AoA-aware scheduling can improve actuation freshness while reducing transmissions.
Ali Nikkhah, Anthony Ephremides, and Nikolaos Pappas
In briefThe Information Bottleneck Problem Is NP-hard
SummaryThis paper studies the information bottleneck method, which seeks a compressed representation of X that retains as much information as possible about Y, given a joint distribution on finite alphabets. The authors prove that computing the optimal value of this problem, including its Lagrangian relaxation, is NP-hard. Unless P equals NP, neither can be approximated in polynomial time within an additive error on the order of |X| to the power of negative 3/2.
WhyProves the information bottleneck and its Lagrangian relaxation are NP-hard, with a polynomial-time inapproximability bound, a notable complexity result.
In briefMinimizing the age of incorrect information in pull-based remote estimation of a Markov source
SummaryThis paper studies a discrete-time pull-based remote estimation system in which a monitor requests samples from a source modeled as a finite-state irreducible Markov chain. Samples travel over zero-delay or unit-delay channels, and the monitor uses a maximum a posteriori estimator. Performance is measured by the mean age of incorrect information (MAoII), under a long-run sampling rate constraint. The authors derive optimal policies with a semi-Markov decision process for the zero-delay case, and use belief-based MDP and constrained deep reinforcement learning for the unit-delay case.
In briefRate-adaptive LDPC convolutional codes with short constraint length
SummaryThis paper designs high-rate, rate-adaptive LDPC convolutional codes with short constraint length using interleaved parity-check polynomials. Building on known polynomial selection results that avoid short cycles in the Tanner graph, it proposes constructions based on Golomb rulers and relaxed variants that keep similar structural properties. Conditions for the absence of length-4 cycles are derived, along with free distance upper bounds that are met in all considered examples.
Massimo Battaglioni, Edoardo Carnevali, Franco Chiaraluce et al.
In briefEfficient KV cache migration for distributed LLM inference across edge servers
SummaryThe paper proposes EdgeFlow, a hybrid method that migrates KV caches by combining direct transmission with activation-based recomputation, running computation and communication in parallel. The authors formulate a multi-node KV migration optimization problem and design a 2-approximation algorithm using improved linear programming relaxation and bipartite graph techniques. Experiments on a real-world distributed LLM inference platform show migration time reduced by up to 45%.
In briefSmoothed-time analysis of two methods for solving deterministic turn-based games on any graph
SummaryThis paper studies algorithms for solving deterministic discounted and mean-payoff games played on any graph. It removes the ergodicity assumption from earlier work and proves that a policy iteration variant by Loff and Skomra runs in smoothed polynomial time on any game graph. It also proposes three new value iteration algorithms for deterministic mean-payoff games, shown to be correct and to run in smoothed polynomial time.
