In briefGeneralized Prophet Inequalities for Online Auctions: Agents with Budgets and Mixed Objectives
SummaryThis paper studies prophet inequalities for online auctions, a setting that classically assumes agents maximize quasi-linear utility. The authors model agents' budgets and varying objectives, such as value maximization in autobidding, as stochastic inputs alongside valuations. They analyze anonymous posted-price mechanisms and generalize the balanced prices framework to incorporate budgets and mixed objectives. The paper shows constant competitive ratios in several central settings, including agents with subadditive valuations, and identifies trade-offs between competitive ratio and price sensitivity. It also gives a general reduction from budgeted settings to budget-free value maximizers.
WhyExtends prophet inequalities to budgeted, mixed-objective agents, proving constant competitive ratios and a reduction to budget-free value maximizers.
In briefFrom Pilots to Production: Lessons on Multi-Site Federated AI for Science
SummaryMany valuable scientific datasets cannot be centralized due to proprietary, export-control, classification, or data-sovereignty limits, so the model must move to the data. The paper synthesizes lessons from U.S. Department of Energy national laboratories, industry deployments, and open-source efforts on moving federated AI from pilots to production. It organizes these lessons around a five-domain readiness frame covering data architecture, privacy and security, trust and verification, governance, and operations. It also identifies open problems such as asynchronous federation, leakage auditing, verification standards, and harmonized data contracts.
In briefDarpan: A Digital Twin Framework for Next-Generation Computing Continuum Decisions
SummaryThis paper proposes Darpan, a digital twin framework for applications spanning devices, edge, fog, and cloud. A physical side runs real applications and observes their state, while a digital side keeps an executable virtual copy updated to evaluate candidate decisions from the same starting point. The selected decision is validated against the latest physical state before execution. The authors report a mean absolute error of 0.485 s for predicted response time and 93.1% scale-out efficiency at 40 nodes on real DAG workloads.
In briefAdaptive power allocation to improve mission efficiency for coordinated low-altitude aircraft swarms
SummaryThis paper studies how multiple heterogeneous low-altitude aircraft cooperate on complex missions, with some handling sensing, communication, computing, and execution. The authors introduce operational-capability entropy to quantify each aircraft's effective work capability. They formulate a power allocation problem that minimizes a linear quadratic regulator cost as a mission-efficiency metric, considering channel conditions. The problem is split into two convex subproblems solved iteratively with closed-form solutions, and simulations show the scheme outperforms traditional approaches.
In briefComparing adaptive Byzantine quorum systems to reduce latency in wide-area networks
SummaryIn Byzantine fault-tolerant replication, a quorum of replicas must vote on a value before it is decided, and the quorum's size and makeup affect client latency over wide-area networks. Existing quorum designs are usually tied to different protocol implementations, which makes fair comparison difficult. The authors extend BFT-SMaRt with a modular quorum-system abstraction so that several quorum constructions can run in the same framework with the same optimizations. They then use network simulation to study how these constructions affect client-observed latency.
In briefAge of information distributions in an Erlang loss system with shared servers
SummaryThis paper derives exact distributions of age of information (AoI) and peak AoI (PAoI) in a bufferless M/M/c/c Erlang loss system, where Poisson-arriving updates from one or more sources share c servers. Three policies handle arrivals when all servers are busy: blocking, preempting a random same-source update, or preempting the stalest one. Using an absorbing Markov chain method, the authors obtain matrix-exponential distributions, which extend to multiple sources affected only by aggregate update rates.
In briefKirin: An orchestrator for cloud-native WebAssembly services
SummaryThis paper presents Kirin, an orchestrator for cloud-native WebAssembly services that handles resource scheduling, lifecycle management, scaling, and request routing. The authors evaluate Kirin against container-based deployments across several service workloads. Results are promising for service lifecycle management, while known limitations remain for CPU-intensive workloads. For typical service workloads, Kirin performs competitively, suggesting WebAssembly could be a viable cloud-native deployment option.
In briefTight liquid welfare bounds for budget-constrained auctions via LP duality
SummaryThis paper studies the efficiency of auctions with budget-constrained bidders and derives tight price of anarchy bounds for coarse correlated equilibria. It introduces a general framework that reduces these guarantees to finding feasible solutions to the dual of a linear program, along with a new smoothness notion for budget-constrained settings. Using this framework, the authors obtain tight liquid welfare guarantees for many auction formats and new lower bounds for the budget-free uniform price and generalized second-price auctions.
WhyProposes an LP-duality framework and new smoothness notion, yielding tight PoA bounds across many auctions and matching a known upper bound.
In briefVersion Age-of-Information-Based Semantic Secrecy Region Analysis: examines semantic secrecy in wireless status updates, balancing information freshness against eavesdropping.
SummaryThis paper studies semantic secrecy in wireless status-update systems using Version Age of Information (VAoI). It models the tradeoff between informative updates and confidentiality with a three-node wiretap channel. The authors introduce the semantic secrecy region (SSR), analyze it geometrically, and derive closed-form expressions for the destination's average VAoI and the eavesdropper's semantic asymmetry probability using a two-dimensional Markov chain. For Rayleigh fading channels, they characterize the optimal SSR boundary and power allocation, and numerical results show that optimized power allocation enlarges the secrecy region compared with fixed-power benchmarks.
In briefFederated Calibration of UAV Motion Uncertainty in Multi-Operator ISAC
SummaryThis paper studies UAV tracking in integrated sensing and communication (ISAC) systems with multiple operators, where raw sensing data, channel information, and transmit decisions stay local. The authors propose a two-timescale framework: slow-timescale calibration of motion uncertainty using local radar innovations, and online fusion of track estimates via covariance intersection with predictive beamforming. Simulations show improved tracking-uncertainty consistency under motion model mismatch, and about 31% lower RMSE than communication-only transmission design with less than 2% extra power.
In briefFair Outcome Prospects Encourage Prosocial Behaviour in the Centipede Game
SummaryThis behavioural experiment studies how people make sequential resource-exchange decisions in the Incremental Centipede Game, comparing a zero-payoff deadline treatment with a treatment ending in a fair split. Even though the fair outcome is uncertain, it moves participants beyond self-interest. An evolutionary model indicates that the fair split shifts play toward mutually beneficial strategies when players are sufficiently prosocial and use probabilistic strategies.
Marco Saponara, Ant\'onio M. Fernandes, Elias Fern\'andez Domingos et al.
In briefFissionReady: Joint Workload and Power Scheduling for Small Modular Reactor Data Centers
SummaryThis paper presents FissionReady, a scheduler for data centers powered by small modular reactors (SMRs). It tracks each module's fuel age, xenon state, and remaining flexibility to set per-module operating limits, then coordinates load-following, batch deferral, and grid purchases through a two-timescale optimization. Compared with a same-sized base configuration, the authors report zero involuntary shutdowns and zero batch deadline misses, with water use reduced by about one-third and grid cost by about half.
Raghavendra Kanakagiri, Rohan Basu Roy, Yankai Jiang et al.
In briefCoded Computing for Dynamic Systems Using Cartesian Products
SummaryCoded distributed computing reduces Shuffle communication bottlenecks through coded multicast, but existing schemes assume a fixed set of workers and do not adapt to arrivals and disconnections. This paper proposes a dynamic scheme that keeps surviving workers' caches and Reduce assignments while using storage from newly arriving clusters. File placement uses Cartesian products of cluster-wise placement delivery arrays, and a greedy algorithm reassigns abandoned Reduce functions. The authors derive the exact communication load for arbitrary feasible disconnections and show, under one-shot communication and uncoded placement, that the scheme is within a factor of two of the lower bound when new clusters arrive without disconnections, and within a factor of four otherwise.
In briefCommunication-efficient (1+ε)Δ edge coloring and the Lovász local lemma
SummaryThis paper studies edge coloring in the two-party edge-partition model, where Alice and Bob each hold part of the edges and must jointly produce a proper coloring with little communication. The authors give a public-coin Las Vegas protocol that finds a (1+ε)Δ edge coloring for sufficiently large Δ, using expected O(ne^{-γΔ}+1) bits and O(log n/Δ+1) rounds. They also develop a two-party constructive Lovász local lemma, which they apply to several other classical problems.
WhyIntroduces a new randomized coloring procedure and a two-party constructive LLL, reducing colors to (1+ε)Δ and communication substantially; solid analysis, relevant to a subfield.
In briefMulti-View LP Bounds for Locally Recoverable Codes with Availability
SummaryThe paper develops a multi-view linear-programming framework for locally recoverable codes with fixed availability a, recording Hamming distances across partitioned helper-set blocks. Constraints from counting identities, product-Krawtchouk positivity, and collision inequalities give a polynomial-size relaxation for arbitrary finite-field codes. For availability 2, exact rational certificates establish M_max values of 8, 27 and 64 for three parameter sets. The four-block bound is strictly stronger than both coarser models.
Ming-Hsuan Kang, Maosheng Xiong, Yu Hsuan Hsieh et al.
In briefMechanism design for choosing a bridge location when agents have optional facility preferences
SummaryThe paper studies a bridge location problem in which two separated regions each have one prelocated facility, and each agent selects a nonempty subset of facilities she cares about. Costs are defined as the maximum, sum, or minimum distance to those facilities. The authors design deterministic strategyproof mechanisms that choose the bridge location and approximately minimize either social cost or maximum cost. They report optimal mechanisms and approximation ratios for several variants, along with lower bounds.
In briefVosti: A Verified Engine for Bitwise-Deterministic LLM Inference
SummaryThe paper formally defines deterministic LLM inference as producing bitwise-identical logits for the same prompt and initial sampler state under a fixed model and deployment configuration. The authors find that production modes such as vLLM's batch-invariant mode and SGLang's deterministic mode yield different logits under some execution variations. They present Vosti, an inference engine built against this specification, with a Verus inductive proof for scheduling and the prefix-sharing KV cache and a Triton analyzer proving kernel output equality. Vosti reports identical logits across all tested variations and performance comparable to vLLM's batch-invariant mode on decode-heavy workloads.
WhyFormalizes deterministic LLM inference and verifies scheduling, KV-cache, and kernels with Verus and a Triton analyzer, beyond empirical batch-invariance guarantees.
In briefExact embeddings of complete bipartite graphs and non-embeddable lattices under information distance
SummaryThe paper encodes vertices of finite complete bipartite graphs as binary strings at each scale, so that information distances approximate scaled graph distances with a constant additive error. This resolves an open problem of Hutter on scale-embeddings of such graphs and shows that no positive power of information distance is exactly representable in a real Hilbert space. The authors also prove that integer lattices admit no scale-embedding and extend this conclusion to (R^m, ||·||_1).
WhyResolves two open problems of Hutter with rigorous proofs, showing information distance powers are non-Euclidean and lattices cannot be scale-embedded.
In briefJoint source-channel coding of Gaussian sources over block erasure channels, with nonasymptotic bounds and channel-uniform normal approximations
SummaryThe paper studies lossy transmission of a Gaussian memoryless source over memoryless, possibly nonstationary block erasure channels at finite blocklength, under an excess mean-squared distortion criterion. It derives computable nonasymptotic achievability and converse bounds and establishes matching third-order, channel-uniform normal approximations. Numerical evaluations compare these joint source-channel coding bounds and their common approximation with an optimized symmetrized separate source-channel coding benchmark.
In briefEfficient HPC Systems for AI Agents: Challenges and Opportunities
SummaryThis paper examines how coding agents have become real users of high-performance computing systems, while HPC abstractions, interfaces, and policies were designed for human-driven workflows. The authors report that agent users are 19.5% of observed users but account for 55.8% of job submissions, 29.1% of CPU core-hours, and 42.7% of GPU-hours. They argue that facilities should co-design systems that treat agents as first-class principals, and outline challenges and opportunities in compute, storage, agent memory, and safety.
Yunjia Zheng, Bintang Dwi Marthen, Zachary Pan et al.
In briefKeeping Provenance in Shared KV Caches to Prevent Unsafe Reuse in LLM Serving
SummaryThe paper finds that LLM serving systems keep provenance such as adapter and weight configuration in local prefix caches, but the shared KV-cache tier keys entries only by token content and coarse model metadata. This lets identical tokens from incompatible contexts collide; cross-adapter collisions cut accuracy from 0.94 to 0.64 in the authors' experiments on vLLM and SGLang. The authors propose a KV provenance contract that binds provenance into lookup and store keys through a canonical descriptor, implemented across three cache paths. The implementation eliminates unsafe reuse while keeping legitimate sharing, with hit-path latency changes within 0.34 ms.
WhyFirst systematic study of provenance loss in shared KV caches, with broad runtime experiments across engines and models; practically impactful for LLM serving.
In briefEfficiency of Generalized Proportional First-Price Auctions Under Auto-bidding
SummaryThis paper studies how efficiently auctions allocate ad slots when advertisers use auto-bidding with return-on-spend constraints. It measures efficiency loss via the Price of Anarchy across prior-free auction formats. For two bidders, the standard proportional first-price auction achieves a tight PoA of at most 1.5. For n bidders, setting the exponent to 2n yields a PoA of 2 minus 1/(4n+1), beating the deterministic barrier of 2.
WhyProves tight PoA of 1.5 for two bidders with matching lower bound, and breaks the barrier of 2 for all finite n.
In briefEfficient client selection for federated learning as a service in multi-application IoT systems
SummaryThis paper examines federated learning as a service (FLaaS) in IoT, where models are trained across devices without sharing raw data, enabling collaboration across applications. The authors propose selecting client nodes for model training based on delay, energy consumption, and link status. They evaluate the method on a Human Activity Recognition dataset under simulated IoT network conditions, comparing it with random and Q-learning based selection. The reported results show improved delay and energy efficiency while maintaining communication performance.
Prachi Nandi, Sonakshi Satpathy, Timam Ghosh et al.
In briefAuctions with Price Predictions: Designing Single-Item Auctions Using a Predicted Revenue-Maximizing Price
SummaryThe paper designs auctions for selling one item in unlimited supply, given only a single prediction of the revenue-maximizing uniform price rather than predictions of each bidder's value. The authors characterize the Pareto frontier of consistency and robustness achievable by any universally truthful auction. They show that randomizing between posting the predicted price and running an optimal prior-free fallback auction is Pareto-optimal. The paper also extends the results to graceful revenue degradation with prediction accuracy and studies learning such predictions from historical market data.
WhyDesigns auctions from one predicted revenue-maximizing price, characterizes the consistency-robustness Pareto frontier, and gives a learning-theoretic path from historical data.
In briefWhy does scale-invariant regret matching converge linearly at the last iterate?
SummaryThe paper studies IREG-PRM+, which normalizes cumulative regret by its own norm and reaches optimal regret without knowing the payoff scale. On zero-sum matrix games it converges linearly in the last iterate, and the authors attribute this to norm saturation, where the regret norm rises to a finite limit and freezes the step size. They prove this always happens and that the last-iterate Nash gap vanishes on every matrix game. Near a unique strictly complementary equilibrium they give a closed-form linear rate and an observable ratio certificate.
WhyExplains linear last-iterate convergence of scale-invariant regret matching via norm saturation, with explicit bounds and closed-form rates, backed by a sizable testbed.
In briefAn information identity showing the geometry behind rare deviation events
SummaryThe paper studies the probability of a deviation event, where the empirical measure of n independent draws falls in a set of distributions. At every n, the exponent of this probability splits into two nonnegative terms: a rate, equal to n times the relative entropy of a typical draw's distribution under the event relative to the population, and a dependence term among the draws, measured by total correlation. The relative sizes of the two terms depend on the geometry of the set, which the paper examines.
WhyGives an exact split of large-deviation exponents into rate and total correlation, linking it to set geometry with rigorous proofs, though mainly theoretical in scope.
In briefA transmitter design that adapts to the receiver for a 200 Gbps C-band optical access network.
SummaryThis paper proposes and numerically demonstrates a receiver-informed transmitter (RITX) waveform design for 200 Gb/s PAM-4 C-band direct-detection passive optical networks. Using 50 GHz optoelectronics, RITX reaches N1/E1 loss classes with an FFE-free ONU and supports a 40 km reach. At 1:64 and 1:128 splits, it yields network-wide complexity savings of 48.1% and 66.7%, respectively.
In briefThe Cost of Truthfulness in Fair Cloud Resource Sharing
SummaryThis paper studies how to fairly and truthfully divide several divisible resources among users with Leontief preferences, a setting motivated by cloud computing. It introduces a parametric mechanism family (ASF) and a Corrected Resource Balancing (CRB) mechanism, measuring efficiency with the fair-ratio benchmark. For two resources, CRB achieves fair-ratio at most 1+1/n, which is asymptotically optimal; for three or more resources, deterministic mechanisms satisfying sharing incentive and strategy-proofness have fair-ratio exactly m.
WhyIntroduces ASF and CRB mechanisms, improving the two-resource fair-ratio from 1.268 to 1+Θ(1/n) with matching lower bounds, rigorously proven.
In briefSimple Mechanisms in Matching Markets: A Tight e-Approximation for Gains from Trade
SummaryThis paper studies how well simple mechanisms approximate gains from trade (GFT) in two-sided matching markets with independent buyer values and seller costs. It shows that the Generalized Random-Offerer mechanism achieves an exact worst-case approximation ratio of e relative to first-best GFT, improving the previous 3.15 bound. The authors also construct markets with regular distributions and a single knapsack constraint showing this guarantee is tight.
WhyEstablishes GRO's exact worst-case ratio e with matching tightness, substantially improving the prior 3.15 bound and resolving the question for this mechanism.
In briefHow GPU server layout, parallelism and congestion control jointly affect MoE inference speed
SummaryThis paper uses ASTRA-sim with an NS-3 backend to simulate MoE model inference across 32 GPU ranks under controlled settings. It tests six server topologies, four TP/EP partitions, two TP collective algorithms and four network congestion-control modes, giving 768 deterministic simulations. The authors find that exposed communication accounts for 89.9% to 95.8% of mean completion time, and that topology, parallelism, collective implementation and congestion control jointly determine completion time.
In briefInnovation-Based Sampling: Conveying Information Through the Direction and Timing of New Information
SummaryThis paper introduces innovation-based sampling of continuous stochastic processes, generating a sample only when the process reveals sufficiently new information relative to earlier samples. For a Wiener process, it characterizes the direction, timing, and frequency of innovations, showing they become sparser over time so the sampling rate vanishes asymptotically. The paper also studies remote estimation from sparse innovations, finding that the absence of an innovation carries information and yields substantially lower mean-square error than a conventional zero-order hold estimator.
