In briefPricing LLM services from a cold start under two-sided learning
SummaryThis paper studies how providers of online large language model services can set prices when they lack historical data, while users' willingness to pay changes with repeated use. It models the problem as two-sided learning, where the user learns from private experience and the provider adjusts prices using only purchase decisions. The authors derive matching minimax regret bounds across two serving-cost regimes and propose corresponding algorithms. Simulations confirm the predicted regret scaling and show the algorithms outperform standard online learning baselines.
WhyThe authors formulate LLM pricing cold start as two-sided learning, derive matching minimax regret bounds with a cost-dependent phase transition, and report simulations supporting the predicted scalin
In briefUsing Antenna Mutual Coupling to Boost Cell-Free Massive MIMO
SummaryThis paper studies a cell-free massive MIMO system in which several access points each use a small, fixed-size linear antenna array with possibly closely spaced antennas. Using multiport communication theory, the authors analyze how mutual coupling affects uplink and downlink spectral efficiency under different impedance-matching strategies. Numerical results show that with a full matching network, mutual coupling can itself be beneficial when an access point serves several user devices.
In briefAllocating Verification in LLM Agent Pipelines with Hidden Tool Degradation: A Mean-Field Analysis
SummaryThe paper studies a simplified model of verification allocation in role-structured LLM agent pipelines where errors are inherited and tools may degrade silently. It establishes fixed-point results for serial and multi-role pipelines and a finite-population deviation bound under independence. Under explicit conditions it characterizes a two-action threshold policy and a restricted four-action interval policy, which synthetic experiments are used to check. The analysis is conditional on the model and is not a general mean-field-game result or empirical evidence about deployed LLM systems.
In briefCovering radius of the third-order Reed–Muller code RM(3,8) is 56
SummaryThe paper proves that the covering radius of the third-order Reed–Muller code in eight variables is 56. The proof splits even- and odd-weight inputs, using moment-controlled slicing and a covariance-rank obstruction. Explicitly stated local covering and compatibility assertions are verified by complete finite certificates.
In briefMulti-GPU acceleration of interface problems in FEniCS for cardiac modeling.
SummaryThe paper extends cuDOLFINx so that mixed-dimensional interface problems can be assembled on GPUs within the FEniCS framework. Combined with GPU-resident Krylov solvers and algebraic multigrid preconditioning, it runs the EMI cardiac model end to end on GPUs. Strong and weak scaling experiments on regular and realistic 3D cardiac geometries, with up to 8000 cells and 22 million degrees of freedom, compare GPU and CPU performance.
Edoardo Centofanti, Benjamin Pachev, Joshua Steyer et al.
In briefSearching a plane using lines between robot pairs: competitive ratios for two and many robots
SummaryThis paper studies deterministic search by several unit-speed robots starting at the origin of the plane, where a target is found once it lies on a segment joining two robots. For two robots, the optimal competitive ratio is 1+√3+π/6 when the target distance is known, and lies between about 3.286 and 8.929 when it is unknown. For three or more robots, the competitive ratio is sec(π/n) whether or not the distance is known.
In briefReed-Muller codes are optimal linear 3-query locally correctable codes
SummaryThis paper proves that any linear 3-query locally correctable code with constant correction radius must have length at least 2 raised to Ω(√k). The previous best lower bound was 2 raised to Ω(√(k/log k)). The result shows that linear 3-LCCs built from Reed-Muller codes, which have length 2 raised to O(√k), are optimal up to constant factors in the exponent. The proof combines chain-derived equations with Kikuchi graphs and shows that a certain edge-coloured Kikuchi graph has a unique colour cycle.
WhyThe authors improve the length lower bound for linear 3-query LCCs to 2^Ω(√k), showing Reed-Muller constructions are optimal up to constant exponent factors.
In briefDynamic coalition formation under capacity limits: a regime-switching value-investing game
SummaryThis paper studies how a network operator and service providers coordinate investment and resource allocation for edge computing under uncertain demand. It proposes an audit-based framework for daily coalition formation, using a two-stage coalitional game that separates infrastructure investment from operational participation, with demand modeled as a geometric random walk. In a reference run over five service classes and three years, the approach gains 12.4% over static grand-coalition membership and recovers 94.6% of the full-information optimum.
In briefDeep reinforcement learning for multi-drone communication aided by BD-RIS and RSMA
SummaryThis paper studies a downlink in which multiple drones communicate through a hybrid setup using beyond-diagonal reconfigurable intelligent surfaces (BD-RIS) and rate-splitting multiple access (RSMA). Drones in good BD-RIS coverage use both direct and reflected links, while others use direct links only. The authors jointly optimize drone trajectories, RSMA rate allocation, beamforming, BD-RIS matrices and drone-to-surface association using a DDPG-based method. Simulations show a sum rate comparable to SAC and higher than TD3, PPO and DQN, at about 1/65 of the per-slot runtime of fixed-trajectory alternating optimization, with gains coming mainly from RSMA.
Aamer Mohamed Huroon, Ali A. Nasir, and Daniel B. da Costa
In briefHotplug Coded Caching: Uncoded vs. Row-Wise Coded Placement Comparison
SummaryThis paper studies the hotplug coded caching model, where K users fill caches of size M from N files, but only an unknown set of A users is active during delivery. Under uncoded placement, it shows that YMA delivery with MAN placement is optimal, and offline users incur a load penalty at intermediate memory sizes. Under row-wise placement, classical schemes are embedded in Reed-Solomon codes, yielding the classical tradeoff for A users with no load penalty, though the placement depends on K. The paper also derives converse bounds showing several of these schemes are optimal under specific conditions.
In briefA game-theory method for settling real-time energy imbalances in local communities
SummaryThis paper proposes a cooperative game-theoretic framework for settling real-time deviations in community-based local energy markets. The settlement process is modeled as a transferable-utility cooperative game, and the Shapley value and nucleolus are compared with a proportional imbalance responsibility rule. The framework is demonstrated using real-world data from a residential energy community in Amsterdam.
Angel M. Gonz\'alez-Rueda, Iago N\'u\~nez Lugilde, Jose L. Crespo-Vazquez et al.
In briefDisC: Game-Based Incentive Design for Fair, Robust, Energy-Efficient Client Participation in Over-the-Air Federated Learning
SummaryThis paper addresses over-the-air federated learning (OTA-FL), where client heterogeneity, fading channels, energy limits, and uneven participation can harm fairness and aggregation accuracy. It proposes DisC, a game-theoretic distributed client selection framework in which each client chooses its participation level by maximizing a utility that accounts for fairness, learning relevance, OTA distortion, energy use, and delay. The authors show that DisC has a Nash equilibrium and can be formulated as an exact potential game.
In briefGPU-parallel AC N-1 contingency analysis using low-rank updates to a shared Jacobian
SummaryThis paper presents a GPU-parallel Newton–Raphson solver for N-1 contingency analysis, which checks power system performance after single-element outages. It inverts one reference Jacobian matrix and reuses it across all contingencies, handling topology changes with low-rank Woodbury corrections and remaining state perturbations with an inner iterative solver. On transmission-scale benchmarks, per-topology solve times are competitive with an optimized 80-core CPU baseline, with the low-rank correction accounting for about 1–2% of the workload.
Marc Hunkem\"oller, Nico Westerbeck, Joost van Dijk et al.
In briefJoint Port Position and Beamforming Optimization for Fluid Antenna Near-Field Sensing
SummaryThis paper proposes a planar fluid antenna array-assisted near-field integrated sensing and communication framework for multiuser communication and target sensing. The authors derive the Fisher information matrix and Cramér-Rao bound for jointly estimating target range, zenith angle, azimuth angle, and radial velocity. They jointly optimize the array port positions and transmit beamforming under communication SINR, transmit power, and deployment constraints, using a two-stage alternating algorithm. Simulations indicate favorable sensing performance with reasonable computational cost.
In briefSame function, different CPU: how hardware differences affect serverless cost and speed
SummaryThis paper studies how the type of CPU a function runs on affects its performance across four commercial Function-as-a-Service platforms. Using variance decomposition, it finds CPU heterogeneity alone explains 4% to 97% of performance variance on AWS Lambda and Google Cloud Run Functions, while Azure Functions shows a platform mechanism that overshadows CPU effects. Because all platforms bill by duration, invocations on the slowest CPU type can cost up to three times as much as those on the fastest.
Jonas Heisterberg, Natalie Carl, Trever Schirmer et al.
In briefCluster Efficiency for LLM Serving via Fast Interconnects
SummaryThis paper studies cluster efficiency for long-context LLM serving, measured in tokens per dollar and per joule, using economical accelerators pooled over fast interconnects. The authors implement their serving stack on a Tenstorrent Galaxy cluster of Blackhole chips, combining integrated communication engines, pipelined ring transfers, communication-computation overlap, and device-side decode replay. Across three dense models in BF16 and 8-bit configurations with 8K to 256K-token contexts, the Galaxy achieves 1.07 to 2.24 times the tokens per dollar of an eight-GPU RTX PRO 6000 Blackwell server. Emulating TCP-class communication reduces throughput by 23.5% to 67.4% and removes the cost advantage in every evaluated configuration.
Youhe Jiang, Pavle Glu\v{s}ac, Binhang Yuan et al.
In briefDispatch control for multi-tenant workflows under multi-resource external rate limits
SummaryThis paper addresses workflow engines calling external APIs with shared rate limits, where retried throttled calls hold worker slots and starve other tenants. The authors propose Backpressure, a worker-side admission gate that atomically checks each call's cost against shared state, enforces fairness over both requests and tokens, and releases worker slots for denied calls to wait on durable timers. On a three-priority workload, it raises Jain's fairness index from 0.60 to 0.89 and cuts high-priority tail latency roughly eightfold.
Stanislav Kosorin, Trever Schirmer, Leonard Wossnig et al.
In briefCustomizing wireless channels with reflecting surfaces: feasibility and limits
SummaryThis paper studies whether a phase-only intelligent reflecting surface (IRS) can reshape the wireless channel between a base station and a user into a prescribed form, under three progressively relaxed criteria. It draws on a mathematical link to the signal error of a point-to-point MIMO system to characterize per-coordinate gain limits and derives array-size conditions for avoiding phase conflicts among streams. Numerical results show that jointly choosing singular vectors and IRS phase shifts reduces average customization error, but gains beyond the propagation spectral limits cannot be achieved exactly.
In briefPersuasion by Parts: Separate Sources Persuade Through Independent Coordinates
SummaryThe paper studies how an information designer can persuade a receiver to act when the payoff depends on independent coordinates observed by separate sources. The designer commits to a product experiment in which each source conditions its signal only on its own coordinate. When the receiver's decision rule is known, the achievable probability of action is at least a constant times the square of the joint-persuasion probability, and the exponent 2 is tight. When the rule is unknown, a product experiment built from the prior alone achieves an exponent of 2 plus epsilon.
In briefPredicting GPU latency on low-end devices without collecting traces from them
SummaryThis paper proposes a zero-shot method for predicting execution latency on low-end consumer GPUs without collecting execution traces on each target device. It configures a local host GPU by selectively modifying device parameters so that its measured latencies approximate those of the target hardware. Evaluated on 60 computer-vision models across 21 consumer GPU models, the method reaches a mean absolute percentage error of 14.21%, versus more than 25% for comparable baselines. Predicted and measured latencies show an average Spearman correlation of 0.961.
In briefAnalyzing Position Estimation Error in Cooperative OFDM Sensing-and-Communication Systems
SummaryThis paper develops a receiver-specific mean-squared-error framework for localization in cooperative integrated sensing and communication systems that use random OFDM payloads. It analyzes local peak perturbations and incorrect ambiguity-peak selection for matched filtering and reciprocal filtering. Matched filtering is examined for how symbol-energy fluctuations cause frequency spread and interference that affect delay errors, while reciprocal filtering removes symbol-dependent weighting but amplifies noise for low-energy symbols. Numerical results show receiver-dependent estimation errors and the benefits of frequency-dependent constellation allocation.
In briefGraph Conditions for Angle Rigidity and Angle-Based Network Localization
SummaryThis paper studies angle rigidity, a theory that characterizes the geometric shape of planar graphs under angle constraints and supports global convergence of angle-based network localization algorithms. It establishes the equivalence between angle constraints and bearing constraints and gives conditions under which cycle compatible frameworks are angle fixable. The authors show for the first time that Laman graphs characterize minimally globally angle rigid graphs. They also propose a distributed algorithm for angle-based sensor network localization that guarantees global convergence under a milder topology condition than prior work.
In briefRose: Graph Optimization and GPU Resource Control for Efficient Embodied RL
SummaryThis paper presents Rose, a runtime-aware system that automates fine-grained overlap and GPU resource management for embodied reinforcement learning while preserving synchronous training semantics. It first uses a window-guided graph optimizer to find low-activity GPU windows and overlap complementary work, then uses NVIDIA MPS with a slack-aware resource manager to tune per-process SM quotas. Across diverse policies, simulators, RL algorithms, and GPU platforms, Rose reportedly achieves 1.20 to 1.92 times higher training throughput and 23.8% to 39.3% higher average SM activity than state-of-the-art systems.
In briefX-PASS: Pinching-antenna systems using radially arranged waveguides, designed and analyzed
SummaryThis paper proposes X-PASS, in which a central hub feeds several radially arranged waveguide arms, each carrying one pinching antenna, and operates them by selecting, aggregating, or multiplexing the arms. The authors model the hub's insertion losses and derive a unified signal-to-noise ratio expression for the three operating modes. They also give a closed-form radius inside which aggregation outperforms selection. Numerical results validate the analysis and compare X-PASS coverage with other layouts at equal transmit power.
In briefInformation cost of inner product and matrix multiplication, via a Shannon-to-min-entropy conversion
SummaryThis paper determines the information complexity of matrix multiplication over finite fields with independent uniform inputs, up to a universal additive constant. For the binary inner product, it resolves an open question posed by Braverman, Garg, Pankratov, and Weinstein about the optimal dependence on error probability. The proof relies on a new theorem about convolution on discrete groups, stating that with enough distributions, the min-entropy of their ordered product is at least the average Shannon entropy minus a constant.
In briefOctEx: GPU vector search over billions of vectors using precomputed lattice neighbor lists
SummaryThis paper presents OctEx, a GPU system for approximate nearest-neighbor vector search that stores hundreds of randomly projected dataset copies in GPU memory. Each copy assigns vectors to nearly round cells of a regular grid and precomputes the nearest vectors to each cell center, so queries read candidates in parallel rather than through many dependent graph-walk rounds. On four H100 GPUs at recall 0.95 to 0.995, the authors report speedups of 2.0x to 46.4x over CAGRA on datasets up to one hundred million vectors, and 1.46x to 2.05x on one-billion-vector datasets.
In briefWhen cache pressure becomes a trap: metastable failures in LLM agent serving
SummaryThe paper studies how a brief traffic surge can break prefix reuse for LLM agents in prefill-decode disaggregated serving engines, leaving service degraded even after load returns to previously sustainable levels. The authors attribute this to a population-cache feedback loop explained by a fluid model. They present Basin, a scheduling approach that reserves predicted session contexts, extends cache with idle decode memory, and sheds excess demand, reporting higher goodput than Continuum and HiCache on real H200 deployments.
WhyThe authors identify cache-population feedback as a metastable failure mechanism in PD-disaggregated serving, propose a fluid model and the Basin policy, and report goodput gains on real H200 deployme
In briefTime-Aware Urban Flood Routing Using Forecast Water Depth and Sticky Road Closures
SummaryThis paper presents a routing engine that uses a two-hour water forecast, updated every 15 minutes, to plan trips around roads closed for a given vehicle class. Any road closing mid-trip stays closed for that query. In tests on a recorded storm in western Hanoi, a router checking only departure conditions sent 7.8% of routes onto roads flooded by arrival. The water depths come from an unverified hydrodynamic simulation, and the rainfall was replayed rather than forecast live.
In briefFinding the best measurement sequence when a measuring device is broken or limited
SummaryThe paper addresses how to choose the best sequence of measurements from the remaining available ones when a measuring device is worn, damaged, or malfunctioning. It frames this as a sequential decision process using information theory and dynamic programming. The authors show that greedy measurement sequences are suboptimal for a simple class of constraints, and propose an algorithm related to optimal binary decision trees with derived complexity bounds.
In briefDesigning finite-state channels in reverse using dynamic measure matching
SummaryThis paper studies an inverse problem for communication channels with memory: constructing a finite-state channel for which a given signaling behavior supports capacity-achieving codes. It introduces dynamic measure matching to turn a directed-information converse into conditions on channel dynamics, defining a notion of dynamic surprise. For a specific noisy refractory family, the paper establishes an exact capacity plateau, differences between Shannon and zero-error feedback capacities, and an exponential strong converse.
In briefA sharp upper bound on mutual information between two binary variables
SummaryThe paper proves a sharp upper bound on the mutual information between two binary random variables, expressed through their distribution-dependent KL contraction coefficient. It resolves the binary-pair conjecture of Anantharam, Gohari, Kamath, and Nair. Equality holds exactly when the variables are independent or form a doubly symmetric binary source. Combined with tensorization and data processing, the result also bounds mutual information between binary outputs from two correlated observation sequences, and this bound depends only on one pair's law when observation pairs are independent copies.
WhyThe authors resolve the binary-pair conjecture, giving a tight bound on mutual information via the KL contraction coefficient and characterizing equality cases with supporting-tangent arguments.
In briefA beam-switched sparse circular array that reduces angle-estimation errors in Wi-Fi sensing.
SummaryThis paper addresses the need for unambiguous angle-of-arrival estimation across the full azimuth plane in Wi-Fi sensing. It proposes an angularly staggered beam-switched uniform circular array that uses spatial complementarity to suppress grating-lobe ambiguities without adding RF chains. Numerical evaluations show significantly fewer AoA estimation outliers and better omnidirectional tracking performance.
In briefCoalitions and Stability Frameworks Unifying Majoritarian and Proportional Multiwinner Allocation
SummaryThis paper proposes two frameworks for multiwinner elections: coalitions, covering properties such as fixed majority and solid coalitions, and stability, covering generalizations of Condorcet's criterion such as Gehrlein stability and local stability. The authors develop theory and tools characterizing the mathematical and computational properties of both frameworks across preference structures including unrestricted candidate domains and restricted party-list domains. Building on this, the paper analyzes majority jackpot electoral systems, which have received limited mathematical treatment in social choice research.
In briefExact shortest paths and maximum flow in planar networks under the CONGEST model
SummaryThis paper studies exact shortest paths and maximum flow in planar networks within the distributed CONGEST model. Using approximate distance labels from a recent preprint, it reduces the round complexity of exact directed single-source shortest paths to Õ(D), where D is the hop diameter. The authors also apply a weighted face-disjoint graph to convert dual distance problems into planar problems, yielding exact maximum flow bounds.
In briefConcatenation locality for codes correcting successive adjacent transpositions
SummaryThis paper studies codes that correct up to t successive adjacent transpositions, using block concatenation so that encoding and decoding can be done blockwise. Concatenations of words from a fixed block set remain correcting at every length, and a single finite test over 2t+1 blocks certifies this property. For binary words, the authors give explicit constructions whose redundancy is close to optimal.
In briefWideband Microwave Analog Computer (MiLAC) for Flat and Frequency-Selective Processing
SummaryThe paper studies the microwave linear analog computer (MiLAC), an analog architecture built from reconfigurable impedance networks that performs computation in the microwave domain. It considers LC resonator implementations with tunable capacitance, maps optimal MISO precoding at a reference frequency to the capacitance settings, and identifies the remaining degrees of freedom. Two techniques are proposed to use these degrees of freedom: a closed-form method for frequency-flat processing and a numerical method for multi-frequency processing, with simulated results for millimeter-wave and sub-6GHz scenarios.
Juan Vidal Alegr\'{i}a, Matteo Nerini, Bruno Clerckx
In briefFairway: Load-Aware KV Cache Reuse for LLM Prefill Scheduling
SummaryThis paper presents Fairway, a scheduler for the prefill stage of LLM serving that balances KV cache reuse with fairness. It prioritizes cache reuse while queues are short and shifts toward fair allocation as requests wait longer relative to their prefill time. A dual-lane design, with a reuse-first Fast lane and a Fair lane, lowers tail time-to-first-token under transient overload without raising the median. The authors report p99 TTFT up to 8.6x lower than prior prefix-aware schedulers.
Hamidreza Alikhani, Nazmul Takbir, Matin Khajavi et al.
In briefPaying for Fractional Error: Tight Constant-Round Approximation of Covering Problems
SummaryThe paper presents a general framework for deterministic constant-round approximation of covering problems in sparse networks under the broadcast CONGEST model, where each vertex sends the same O(log n)-bit message to its neighbors. It combines a solver giving a (1+δ)-approximate solution to local covering linear programs and a rounding rule that turns an approximate fractional solution into an exact coefficient. For ordinary domination, it achieves a 2a+1 approximation against the fractional optimum under a minor edge-density assumption, which equals seven on planar graphs and matches a known lower bound.
WhyThe authors combine a local covering LP solver with a rounding principle, reporting tight constant-round approximation coefficients matching lower bounds on planar graphs.
Saeed Amiri, Sebastian Siebertz, Wojciech Wawrzyniak
In briefOptimal online learning in bilateral trade with interest
SummaryThis paper studies online learning in bilateral trade on platforms such as Uber, Airbnb, and eBay, where a platform posts prices to buyers and sellers and observes only whether trade occurs. It introduces global budget balance with interest, allowing trading profits and accrued interest to finance subsidies. The authors show that optimal regret against the best fixed common price drops from Θ̃(T^{3/4}) to Θ̃(T^{5/7}) as the interest rate grows, and that the optimal competitive ratio for the α-regret extension rises from 2 to 3.
WhyThe authors introduce global budget balance with interest and report tight regret bounds in adversarial and i.i.d. valuation models across interest-rate regimes.
