In briefRobustness in Mechanism Design: A Worst-Case Review of Auction and Exchange Mechanisms
SummaryThis article reviews robust design and analysis of auctions, using the computer science paradigm that measures robustness by the worst-case ratio, over environments, between a robust auction's performance and an optimal auction's. It covers robust analysis of the first-price auction for welfare, prior-robust auction design for revenue, and prior-robust exchange mechanism design for gains from trade. It also shows that the revelation principle can lose performance in robust design, quantified by the revelation gap.
In briefFast demodulation for lattice-based coset codes (quantize-walk method)
SummaryThis paper addresses the high complexity of soft demodulation for lattice constellations by proposing fast quantize-walk demodulation (FQWD) for multilevel coset codes. The method uses radial projections onto the Voronoi cells of the D4 and E8 lattices to reduce complexity from O(d2^{sd}) to O(sd2^d). Experiments show only negligible block error rate loss relative to maximum likelihood demodulation under AWGN.
Chloe Makdad, Leopold Bertholet, Mairead Cummins et al.
In briefePACT: An energy-performance-aware controller that tracks hourly power commitments for LLM serving
SummaryThe paper argues that cutting LLM serving energy alone does not guarantee lower costs, because deviations from preset electricity commitments can be costly, especially when overuse and underuse are priced asymmetrically. The authors propose ePACT, a two-level controller that adjusts serving capacity and GPU clocks while meeting request-level service requirements. In full-day trace simulations on H20 and H200 GPU pools, ePACT reduces the asymmetric deviation cost by 73.8% and 75.7% relative to vLLM while retaining near-vLLM SLO attainment.
In briefStorage codes from binary BCH codes that approach unit rate
SummaryA storage code assigns symbols to graph vertices so each symbol can be recovered from its neighbors. The paper proves that binary full-parity storage codes on triangle-free coset graphs of primitive BCH codes have rate approaching one for any fixed error-correction capability of at least two. For the double-error-correcting family over F_{2^m}, the storage deficiency is shown to be Θ(((1+√5)/4)^m).
In briefToken-based economy for fairer, efficient highway congestion control using express lanes
SummaryThis paper studies a non-monetary token economy for allocating highway lanes, aiming to improve fairness without reducing traffic efficiency. High-occupancy vehicles get unrestricted express lane access, while other users earn and spend tokens to switch between regular and express lanes. In a microscopic traffic simulation using real-world demand data, the proposed prices produced nearly the same average travel time as a scenario with no reserved express lane, while substantially reducing urgency-weighted perceived travel time.
Leonardo Pedroso, Juan Pablo Bertucci, W.P.M.H. Heemels et al.
In briefGiving LLMs social value profiles makes their play in a social dilemma more human-like
SummaryThe authors test eight large language models on two versions of the Centipede Game and compare their default choices with human experimental data. The models' default strategies vary widely and are generally far from the human reference. When conditioned on Social Value Orientation (SVO) profiles drawn from human participants, alignment with human behaviour improves by up to 70%, and higher SVO values reduce the probability of stopping the game.
In briefSpeeding up MPI programs by letting compute-heavy and memory-heavy parts drift out of step
SummaryThe paper shows that in MPI applications with little communication that alternate between compute-bound and memory-bound phases, communication stalls can unintentionally ease memory-bandwidth contention. The authors demonstrate this with a parallel optical flow solver combining ray tracing and an optical flow kernel, achieving considerable speedup through desynchronization. A tunable microbenchmark and a bandwidth-aware simulator are used to study and validate the effect.
In briefOptimal Universal Coding of Integers: Numerically Determining the Minimum Expansion Factor
SummaryThis paper studies universal coding of integers (UCI), which assigns binary codewords to positive integers so that the average codeword length stays within K times max{1,H(P)} for every nonincreasing source distribution. The authors show that the worst-case ratio for any universal code is attained by a family of one-point-plus-uniform-tail distributions. They also introduce a UCI inequality, analogous to the Kraft inequality for prefix codes, which yields an equivalent definition of the optimal minimum expansion factor C*. Numerical computation gives C* = 2.000124757036101…, with the first fifteen decimal digits certified.
In briefAgentic Wireless Digital Twin Building and Calibration from Real Measurements
SummaryThis paper proposes AWDT, an agent-based framework that automatically builds and calibrates wireless digital twins. Three agents handle the propagation environment, base station and sector configuration, and radio material properties, refining the model using differences between smartphone-measured RSRP and ray-tracing predictions. In real-world experiments, the RSRP prediction MAE dropped from 11.46 dB to 5.25 dB. Evaluation with an independent measurement system also showed transfer across devices with lightweight device-specific bias adaptation.
In briefGPU-Initiated Communication: A Low-Level Performance Dissection
SummaryThis paper examines how GPU threads post RDMA operations directly to the network interface card, a mechanism used by NVSHMEM, NCCL GIN, and DeepEP. The authors build minimal GPU and CPU-proxy transports and benchmark them against several libraries on NVIDIA H100, H200, B200, and GB200 systems. They report that a minimal GPU path issues an operation in 0.7 µs and completes it in 4.0 µs, with library overheads adding up to 4.6 µs. The paper concludes that the submission path alone does not predict overall communication performance.
WhyDecomposes GPU-to-NIC communication via minimal transports and measures multiple libraries on several platforms, separating hardware from library costs.
Javid Baydamirli, Ismayil Ismayilov, Kaan Oktay et al.
In briefJoint communication and sensing coverage for UAV corridors with beam misalignment
SummaryThis paper studies joint communication and sensing (JCAS) coverage probability for a terrestrial base station using a realistic 3D 3GPP antenna pattern, with UAV users deployed in a finite cylindrical aerial corridor. The base station uses radar sensing to track a known UAV, under a 3D beam misalignment error modeled as multivariate normal. Exact analytical expressions for JCAS coverage probability are derived under clutter and uplink interference. The results show corridor-length saturation at high UAV heights and a beamwidth trade-off between interference suppression and misalignment sensitivity.
Harris K. Armeniakos, Petros S. Bithas, Konstantinos Maliatsos et al.
In briefJoint Communication and Sensing in Aerial Corridors: A Stochastic Geometry Model
SummaryThis paper studies how well a drone base station can both communicate and sense in an aerial corridor filled with drone users. The corridor is modeled as a finite cylinder, with users randomly placed in three dimensions and the base station tracking a known user with radar. The authors derive exact formulas for coverage probability under clutter and uplink interference. They find that more directional base station antenna beams improve performance noticeably, especially in shorter corridors.
Harris K. Armeniakos, Petros S. Bithas, Athanasios G. Kanatas et al.
In briefUAV attitude-aware trajectory and attitude co-design for covert communication
SummaryExisting UAV covert communication trajectory designs usually treat the drone as a point mass and ignore how maneuvers affect antenna orientation. This paper studies a rotary-wing UAV with a directional antenna sending confidential data to a legitimate receiver while a warden is present, explicitly modeling acceleration-induced attitude and its effect on channel gain. It derives a covertness constraint using KL divergence at the warden and jointly optimizes trajectory and attitude with an MM-based algorithm to maximize average covert rate. The framework is also extended to a robust design under imperfect attitude control.</en
In briefRapidMoE: Adaptive Residual Offloading for Faster Large-Scale MoE Inference
SummaryThis paper addresses the mismatch between large Mixture-of-Experts (MoE) models and heterogeneous CPU-GPU hardware, where PCIe bandwidth limits and heavy CPU computation reduce efficiency. It proposes RapidMoE, a residual offloading system that shifts offloading from the expert level to the bit level, using compact data representation, dual-path routing matched to hardware, and balanced scheduling. The authors also introduce a runtime arbitration method to adjust critical experts. Experiments report up to 3.5x decoding speedup and 2.1x prefill speedup over state-of-the-art offloading systems.
In briefPractical codes for computing Boolean functions over channels for decision tasks
SummaryThe paper proposes a practical Reed-Solomon-based code construction for Boolean function computation, where a receiver determines a function value unknown to the transmitter. For noiseless binary channels, it derives finite-blocklength worst-case error bounds and characterizes rate-reliability tradeoffs across function families. The construction is extended to noisy channels by concatenation with a conventional channel code, and numerical results report SNR coding gains over conventional transmission for specific tasks.
In briefServing Revisable Agent Tasks Through Versioned Execution
SummaryThe paper presents RETIRE, a serving control-plane design for LLM agents whose running tasks get revised. Each execution version holds the authority to publish output or install state, so obsolete work is revoked first and the completed prefix that a successor can inherit is certified. The design is implemented in vLLM. In controlled paired experiments, it reduces the median revision-to-successor time-to-first-token by 17.1%.
WhyProposes versioned execution unifying obsolete-work revocation and KV state inheritance, implemented in vLLM with paired experiments showing a 17.1% median TTFT reduction.
In briefFair, non-wasteful matching of families and daycare places, including siblings and already-enrolled children
SummaryThis paper studies matching between families and daycare centers, where a family may have several children with joint preferences and some children are already enrolled and seeking transfers. The authors introduce resettlement, which lets displaced families return to their initial assignments, and give a greedy algorithm showing that matchings satisfying non-wastefulness and resettlement-based fairness exist. They also analyze computational complexity, finding that some existence questions are NP-complete and computing resettlement-based outcomes is PLS-hard.
In briefSeamless reconfiguration for DAG-based Byzantine atomic broadcast in asynchronous networks
SummaryThis paper proposes a seamless reconfiguration protocol for asynchronous DAG-based Byzantine atomic broadcast, which adds or removes nodes. Existing DAG-based BFT protocols assume a static set of processes, and prior reconfiguration approaches either pause the system or wait for a predetermined epoch boundary. The proposed protocol lets the DAG keep growing while membership and quorum thresholds change, without halting the system. The authors state it is the first such protocol for fully asynchronous networks.
Michael Yiqing Hu, Leander Jehl, Paul Franke-Bergmann et al.
In briefFeature freshness budgets for real-time ML inference under stream lag
SummaryThe paper models an online feature store as a materialized view over an event stream and defines a per-feature freshness budget, the gap between a feature's decision window and the staleness imposed by serving. It proves a per-feature staleness bound and a closed-form threshold below which a feature cannot be served within budget. The results are validated in a discrete-event simulation and on a real Kafka, Redis, and PostgreSQL pipeline, where the predicted threshold reproduces. The authors find the main simulation-to-infrastructure gap stems from sampling-phase artifacts rather than latency, and randomizing the phase reduces the drop-rate error to 0.018.
WhyFormalizes feature freshness as a bounded budget with a closed-form threshold, validated on real Kafka, Redis, PostgreSQL pipelines, and identifies a phase-locking simulation bias.
In briefReading While Writing: Baseline Data Needs for Molecular Neural Interfaces
SummaryThis paper studies molecular neural interfaces that both deliver and sense a neurotransmitter, and how they can estimate endogenous release despite their own chemical input. The authors analyze two exchanging regions with shared saturable uptake and show that an unknown resting concentration difference leaves the source region's uptake uncertain. A dopamine-inspired example and a Gaussian observation model are used to show how delivery and baseline information affect release classification.
In briefMANTA: ML-Augmented Tiering Advisor for Smarter Memory Page Placement
SummaryMemory tiering combines fast DRAM with slower tiers such as CXL-attached memory to expand capacity. The paper introduces ChOMP, an offline optimizer used as a reference, and a trace-driven simulator that reveals opportunities for online policies. It then presents MANTA, which predicts page usefulness from runtime access features using a lightweight learned model integrated into ARMS. On eight emulated CXL workloads at 4 GB fast memory, MANTA achieves geometric-mean speedups over ARMS of 1.12× on Linux 6.2 and 1.08× on Linux 6.18, and 1.69× across six Optane workloads.
Johannes Freischuetz, Kiet Pham, Sujay Yadalam et al.
In briefDitto: Generalized Reconfigurable Linearizable Reads
SummaryThis paper studies how to make linearizable reads faster in distributed datastores without giving up consistency. It classifies existing read algorithms into a generalized space and analyzes their latencies mathematically, revealing two fundamental tradeoffs. The authors propose a new algorithm, Pairwise Quorums, and Ditto, a system that reconfigures across this space at runtime as network and workload conditions change. They claim that together these subsume existing algorithms, being either strictly better or matching them where a real tradeoff exists.
In briefMegaFlux: Skew-resilient MoE megakernels with pipelined expert replication
SummaryThis paper addresses slowdowns in mixture-of-experts (MoE) models, where uneven routing overloads some GPUs while others sit idle. It presents MegaFlux, which decides at runtime whether to replicate popular experts and pipelines the resulting weight transfers and backward gradient reductions with expert computation. Across 147 configurations per direction on eight NVIDIA B200 GPUs, it reports geometric-mean speedups of 1.45x for forward and 1.28x for backward over fixed expert placement.
Jianzhu Yao, Siva Kumar Sastry Hari, Vignesh Balaji et al.
In briefForming Alliances Under Uncertainty: A General Lotto Study
SummaryThis paper studies whether two players fighting a shared opponent can benefit from alliances when they only partly know each other's or the adversary's resource budgets. It uses the coalitional General Lotto game, where players allocate budgets across valued contests. The authors characterize which games admit a budget transfer that helps both players under multiplicative uncertainty. They show such robust transfers exist in a positive-measure set of games at every uncertainty level.
In briefMarket design for two-sided ad auctions with autobidders: which mechanisms still work when both buyers and sellers are strategic
SummaryThe paper studies two-sided markets in which autobidders, such as Return-on-Spend (RoS) value maximizers, participate on one or both sides. It shows that, without prior information, many incentive-compatible, budget-balanced mechanisms have unbounded Price of Anarchy for liquid gains from trade when value maximizers are present. It also finds that McAfee's Trade Reduction mechanism keeps a constant Price of Anarchy for liquid welfare under RoS bidding. With distributional information, the authors design a mechanism that achieves first-best liquid welfare when at least one side consists of RoS maximizers, contrasting with the Myerson-Satterthwaite impossibility.
WhyFirst study of two-sided autobidder markets, with negative and positive results, and a first-best mechanism under distributional information contrasting Myerson-Satterthwaite impossibility.
In briefL2R-EV: Learning Which Repairs to Make in Electric Ride-Pooling with Limited Charger Queues
SummaryThe paper studies electric ride-pooling, where passenger service, routes, batteries, and limited chargers must be managed together. The authors build a discrete-event simulator and add a learned repair layer that reassigns uncollected passengers, guided by a supervised score and a two-stage PPO policy. On 12 held-out Manhattan episodes, supervised repair cut operational cost by 6.57% versus no repair and won in all 12 episodes, while queue-aware charging reduced waiting by 41.0%.
In briefPricing Time, Not Just Tokens: Latency-Aware Pricing Mechanisms for LLM Inference
SummaryThe paper models LLM inference as a service market in which buyers have private willingness-to-pay, task volume, and time preference, and utility depends on both token counts and latency slack. Its main result is a separation theorem: discrete hardware tiers make buyers self-select on time preference, reducing three-dimensional screening to one-dimensional screening within each tier. The authors characterize optimal tiered mechanisms using virtual-value techniques and calibrate them to 8-GPU H100 and B200 clusters.
In briefShaping and modulation augmentation for LDPC-coded QPSK
SummaryThis paper adds a short shaping code to LDPC-coded QPSK so that some codeword bits control how often lower- and higher-energy constellation points are sent. The lower-energy points are chosen more often, producing a Gaussian-like symbol distribution. The receiver refines shaping-bit likelihoods using LDPC decoder feedback, and the shaping codebook, bit mapping and constellations are jointly optimized. The authors report gains of up to 0.6 dB over 5G NR LDPC-coded QPSK and 0.53 dB over QAM-16.
Heping Wan, Joonyoung Cho, Sandesh Rao Mattu et al.
In briefLatencyLab: a DPDK-based tool for measuring P4 pipeline latency on FPGA SmartNICs
SummaryThis paper presents LatencyLab, a DPDK-based framework that measures how much latency a P4 program adds on an FPGA SmartNIC, without needing PTP support or clock synchronization. A switch duplicates each probe packet to two ports, one through the VitisNetP4 pipeline and one through a matched bypass path, and the arrival-time difference isolates pipeline latency. The authors evaluated four VitisNetP4 programs on an AMD Alveo U280 and report tight, reproducible latency distributions, with two independent checks agreeing with the measurements.[18]
