In briefTruthful-in-expectation fair-share allocation of chores with a constant guarantee
SummaryThis paper studies how to allocate indivisible chores fairly while keeping truthful reporting incentives in expectation. It presents the first such mechanism that guarantees a constant ex-post maximin share approximation for any number of agents, achieving 4/3 for two agents and 3/2 for three. For two agents, the authors raise the known lower bound to 4/3, showing their mechanism is optimal there. For any n of at least 3, they prove a lower bound of 13/12 on the achievable ratio.
In briefWeaver: A System Using Idle GPU Capacity at AI-RAN Sites to Train Foundation Models
SummaryThe paper studies spare GPU capacity in AI-RAN cell sites, finding that 40-85% of GPU capacity goes unused, and that this spare capacity is complementary across sites. The authors present Weaver, which trains foundation models alongside radio access network workloads without degrading RAN performance, using a compute-aware MAC scheduler and elastic training. According to the paper, the prototype creates up to 4.9x more usable spare compute, and a multi-site testbed shows 2.1-3.7x higher training throughput than baselines.
In briefFinite-Resolution MiLAC-Aided Multiuser Beamforming
SummaryThis paper studies multiuser beamforming with a microwave linear analog computer (MiLAC), whose tunable elements can take only finitely many states instead of continuous values. Under lossless and reciprocal constraints, it jointly optimizes discrete susceptances and power allocation, and designs codebooks from channel statistics. The authors propose exact continuous penalty models, proven globally equivalent to the discrete problems, and ADMM-based algorithms. Numerical results show a 3-bit stem-connected MiLAC reaches over 95% of the sum-rate of an unquantized fully-connected MiLAC using only 22% of its tunable components.
In briefEnvy-free splits of random assignments, with four agents solved
SummaryThe paper studies whether a random assignment of objects to agents can be decomposed into deterministic assignments so that every agent envies any other with probability at most 1/2. It proves this is always possible for four agents using a computer-aided check over all vertices, and shows 1/2 is the best general threshold. It also gives proofs for special cases and shows some natural decomposition methods fail. Deciding whether an arbitrary random assignment admits such a decomposition is strongly NP-complete.
WhySettles the open four-agent case with a complete computer-aided proof and proves strong NP-completeness, a high-novelty result with limited practical impact.
In briefAReaL-TIK: An agent that optimizes unified RL GPU kernels using a search-tracking representation
SummaryReinforcement learning post-training often uses separate GPU kernels for rollout and policy updates, and numerical differences can distort probability ratios. The paper introduces KernelBraid, an agentic framework that starts from a hand-tuned, bitwise-consistent unified kernel and organizes code search with an optimization intermediate representation. Changes are kept only if they pass correctness checks and improve aggregate latency. Across 12 end-to-end training setups on H20, it reaches 1.10x average throughput relative to AReaL with log-probability recomputation.
WhyProposes an agentic optimization framework with an IR for bitwise-consistent RL kernels, validated across 12 end-to-end configs and multiple operators.
In briefDoes AI Advice Help or Harm Decisions? Study of Information Gathering with AI Input
SummaryThis paper studies how AI advice given before a costly information-gathering step and final decision affects decision quality, measured by the gross utility of the final action. It characterizes gross effects through the value function generated by optimal acquisition under uniformly posterior-separable costs. The authors show that convexity of this value function is necessary and sufficient for every AI signal to weakly improve gross utility at every prior, while nonconvexity can allow harm.
WhyCharacterizes when AI advice weakly improves decision quality via convexity of continuation value, with tight state-space results and counterexamples; theoretical, with limited empirical grounding.
In briefTighter limits for constant-power and low-power error-correcting cooling codes
SummaryThis paper studies binary constant-power and low-power error-correcting cooling codes for on-chip buses, which correct errors while controlling power and thermal effects, with e=w-3. The authors extend the earlier Zhao and Zhang upper bound for CPECC codes from w≥2t(t+1)+2 to w≥w0(t), where w0(t) is about √2·t^(3/2), and use Steiner systems to show this bound is attainable and asymptotically tight. For LPECC codes they give a new upper bound, ⌊C(n+2,3)/C(w+t,3)⌋, valid for w≥μ(t) with μ(t) about √2·t^(3/2), which is strictly smaller than the previous bound where both apply and is asymptotically tight.
In briefAutonomous AI system finds new kissing number constructions in 19 dimensions
SummaryThe kissing-number problem asks how many non-overlapping spheres can touch a central sphere, and exact values are known only in a few dimensions. The authors used the autonomous multi-agent system Qiushi Engine to obtain new lower bounds in nineteen dimensions. Each result is supported by explicit mathematical arguments and independently checkable finite certificates. Examples include K(25)≥197580 and K(55)≥53301140.
WhyGives new kissing-number lower bounds in nineteen dimensions via varied constructions with checkable finite certificates, though improvement magnitude over prior work needs full-text confirmation.
In briefA proof of the generalized packing-covering conjecture (a coding-theory conjecture now proven)
SummaryThis paper proves the generalized packing-covering conjecture for linear codes. The conjecture states that the t-th generalized Hamming weight is at most twice the t-th generalized covering radius plus two. The proof is computer-assisted, combining a parity-check reformulation, length bounds on possible counterexamples, and puncturing arguments. All orders t ≥ 32 are settled, and the remaining finitely many cases are excluded by exact computer verification.
WhyProves a long-standing conjecture via parity-check reformulation, length bounds, puncturing, and exact computer verification of remaining cases.
Gianira N. Alfarano, Giuseppe Marino, Alessandro Neri et al.
In briefMonitoring Multitenant Kubernetes Clusters with TLA+ Trace Checking
SummaryThis paper presents a monitoring tool that uses TLA+ trace-checking specifications at runtime, rather than only for design-time model checking. The tool infers system behavior from Kubernetes audit logs and detects violations in it. Its primary use case is multitenancy violations, though the pipeline is not limited to that setting. The work shows how formal reasoning can be incorporated into live Kubernetes environments to improve monitoring and correctness checking.
Ioana Sila\c{s} (West University of Timi\c{s}oara), Adrian Cr\u{a}ciun (West University of Timi\c{s}oara)
In briefA method for synthesizing safe update schedules for safety-critical systems
SummaryThis paper addresses how to apply software updates to safety-critical systems without interrupting operation or using cold spare hardware. The authors model the system and the update as a two-player timed game and use a bounded SMT encoding to synthesize a fixed global-time update schedule. The schedule is claimed to guarantee safe and complete deployment under all admissible system behaviors, demonstrated on an autonomous driving trajectory planner.
Janis Kr\"oger (Carl von Ossietzky Universit\"at Oldenburg), Paul Kr\"oger (Carl von Ossietzky Universit\"at Oldenburg), Martin Fr\"anzle (Carl von Ossietzky Universit\"at Oldenburg)
In briefWaveAlign: Cache-Aware Query-Row Scheduling for Faster Sparse Attention in Long-Video Generation
SummaryLong-video generation with diffusion transformers produces very long token sequences, so attention becomes the main inference bottleneck. The paper proposes WaveAlign, a lightweight framework that reorders query rows in dynamic sparse attention, grouping rows by similar key/value access patterns and sorting them within GPU waves to improve L2 cache hit ratio and reduce HBM traffic. Because only query and mask rows are permuted, sparse-attention semantics and existing kernels are unchanged. Reported results across two GPUs, two video DiTs, and four sparse-attention methods include up to 1.25x kernel and 1.17x end-to-end speedup without quality loss.
In briefDivergence-minimizing distribution matching for parallel and memory channels
SummaryThe paper extends distribution matcher theory to distributions with memory and shows that the divergence-minimizing matcher selects the highest-probability target sequences, as in the memoryless case. A divergence scaling law is generalized to distributions driven by innovation processes. The theory is applied to parallel additive white Gaussian noise channels using a modified enumerative sphere-shaping method with a weighted energy constraint. A three-channel example shows that joint shaping across channels substantially reduces rate loss at short blocklengths compared with product matchers, and simulations with probabilistic amplitude shaping and a 5G-NR LDPC code confirm the gains.
In briefRobust beamforming for OTFS massive MIMO under uncertain channel information, with GPU acceleration
SummaryThis paper studies robust beamforming for OTFS-based massive MIMO systems when channel state information is uncertain. The authors minimize total transmit power under a chance constraint that guarantees a set outage probability, and model the uncertain channel with support vector clustering as an asymmetric uncertainty set. They convert the problem into a semidefinite program and design a GPU-accelerated parallel ADMM algorithm. Simulations reportedly show lower transmit power than conventional schemes and a speedup from the GPU-accelerated ADMM.
In briefAutomatic self-reduction gives stronger randomized distributed lower bounds
SummaryThis paper studies randomized lower bounds for graph problems in the distributed setting. The authors show that self-reduction is a special case of classical round elimination and introduce a new error measure under which round elimination yields bounds beyond ω(log log n). They identify a class of problems where this improvement applies black-box. The paper proves Ω(√log n) randomized lower bounds for maximal matching on regular 2-colored graphs, 1/k-integral matching, and maximal H-packing.
WhyUnifies self-reduction as a special case of classical round elimination, yielding black-box randomized lower bounds beyond ω(log log n) with clear novelty.
In briefRobust Resource Management for Satellite-Air-Ground Networks Using Neural Network Channel Uncertainty Learning
SummaryThis paper studies joint robust beamforming and resource allocation in space-air-ground integrated networks when channel state information is uncertain. It formulates a chance-constrained problem that minimizes total transmit power while meeting user quality-of-service requirements under a set outage probability. A deep neural network learns the uncertain channel as an asymmetric uncertainty set, and an adaptive iterative algorithm jointly optimizes resource allocation and beamforming. Simulations show the method achieves better energy efficiency and robust reliability than traditional robust and non-robust approaches.
In briefScattering Measures and Information Geometry: Describing Scattering Experiments with Finite Measures and Their Information Structure via Fisher Metric
SummaryThe authors formulate scattering experiments as parameterized finite measures on outcome spaces and develop the information geometry defined by the Fisher metric. A scattering measure splits into total mass and normalized shape, separating Fisher information into changes in overall scattering strength and redistribution among outcomes. Measurement processes are modeled as probability or sub-probability kernels that contract Fisher information, linking the framework to optimal experimental design. The same finite-measure structure also arises in quantum measurements with incomplete detection.
In briefCovert communication over large fluid antenna arrays: how channel hardening affects detection
SummaryThis paper studies covert communication over enormous fluid antenna systems (E-FAS), where guided surface wave propagation creates a channel with a random covariance scale. With a passive warden present, the optimal likelihood ratio test reduces to an energy detector, and the authors derive its false alarm and missed detection probabilities. Numerical results show that finite-mode E-FAS channels improve covertness over the fully hardened Rayleigh limit for finite observation intervals. The effect of channel hardening is asymmetric across the legitimate and warden links.
Farshad Rostami Ghadi, Damoon Shahbaztabar, Kai-Kit Wong et al.
In briefComputational complexity of multi-agent binary-action contract design, mapping approximation limits within gross-substitutes rewards
SummaryThis paper studies how hard it is to compute optimal contracts when several agents each choose between two actions, focusing on gross-substitutes reward functions and related classes. The authors show the problem is APX-complete even for OXS rewards, and they give an EPTAS for weighted matroid rank functions, with no general randomized FPTAS. For partition matroid rank functions they give an FPTAS. For ultra reward functions, they prove that no polynomial-time randomized value-query algorithm achieves a 2^{o(n)} approximation in expectation.
WhyMaps a fine-grained approximability landscape within gross substitutes, with rigorous APX-completeness, EPTAS, and FPTAS results.
In briefTorch-PIM: Profile-guided automatic decisions on offloading PyTorch code to memory-near compute
SummaryThis paper presents Torch-PIM, a compiler framework that uses profile-guided optimization to decide whether each parallel loop in PyTorch-generated code runs on the host CPU or on processing-in-memory (PIM) hardware. Each loop is assessed by its amount of work and its memory boundedness, using host profiling or information from MLIR. The authors report speedups over CPU-only execution of up to 8.6x on tensor operators, 2.9x on MLP, 4.4x on Attention, 5.1x on GPT-J-6B, and 3.6x on LLaMA-7B across PIM configurations of 32 to 128 cores.
Heeeon Lee (1), Hyunwoo Nam (1), Junyong Heo (1) et al.
In briefContraction of Rényi divergences for discrete channels (how Rényi divergences shrink through noisy channels)
SummaryThe paper studies strong data-processing inequality (SDPI) constants for Rényi divergences on finite spaces, examining how they depend on the Rényi order. It proves monotonicity and convexity properties, identifies support restrictions on the relevant distributions, and gives bounds relating the constants to χ² and Hellinger divergences. The results are applied to local differential privacy and to finite-time convergence bounds for Markov chains.
WhySystematically characterizes Rényi-SDPI constants on finite spaces, with monotonicity, convexity, support restrictions, and tensorisation bounds, applied to LDP and Markov chains; strong theory.
Adrien Vandenbroucque, Amedeo Roberto Esposito, Michael Gastpar
In briefNash Equilibria in Pacing-Strategy Auctions: Complexity and Inefficiency
SummaryThe paper studies Auctions with Pacing Strategies (APS) games, where utility-maximizing bidders across many simultaneous first-price auctions each choose one pacing multiplier that scales their bids. It shows some instances have no approximate pure Nash equilibrium and that deciding whether an equilibrium exists is NP-complete in general but polynomial when the number of bidders or items is fixed. When equilibria exist, both the Price of Anarchy and the Price of Stability are exactly e/(e-1).
WhyIntroduces APS games, proves non-existence of approximate pure equilibria, NP-completeness, and exact tight inefficiency bounds; strong theory, narrow subfield impact.
In briefSemantics, Workflows, and Infrastructure: How Production AI Agents Are Served at Scale
SummaryThis paper analyzes a two-week trace of 11.7 million requests from a large-scale production platform for general-purpose AI agents, supported by inference infrastructure with over 10,000 GPUs. It examines the platform at three connected levels: task initiation semantics, workflow execution patterns, and infrastructure serving demands. The measurements reveal highly skewed request volumes across sessions, rare execution overlap among sibling requests, and context reuse across task boundaries. The authors then discuss deployment implications and open problems for future agent serving systems.
In briefSatellite semantic communication with vector quantization and index compression for high-resolution remote sensing images over AFDM
SummaryThis paper proposes DVQ-SDSC, a dual-branch vector-quantization-aided satellite semantic communication framework for transmitting high-resolution remote sensing images over affine frequency division multiplexing (AFDM). It processes high-frequency residuals and low-frequency structural semantics separately, and reduces index bitrate using PCA-based codebook reordering and group DPCM (G-DPCM) coding. Experiments on FAIR1M over 3GPP NTN-TDL-D channels show it outperforms the JPEG-LDPC scheme at 0.0625 bits per pixel.
In briefHow partition shape and rank density affect the speed of distributed quantum circuit simulations
SummaryThe paper evaluates 764 validated configurations of 39-qubit distributed quantum circuit simulations on 1,024 nodes of the Fugaku supercomputer. It finds that torus partition geometry dominates runtime, with flat partitions running 1.73 to 2.31 times slower than near-cubic ones despite identical data transfer. Rank density is a secondary factor, reducing runtime by about 11% at 16 ranks per node only on compact geometries.
In briefVarioPath: Workload-Aware All-to-All Communication for PCIe GPU Clusters
SummaryThis paper presents VarioPath, a scheduling framework for AlltoAllv communication on PCIe GPU systems, which is important for distributed mixture-of-experts model inference. It pairs an offline topology-aware analyzer with an online demand-aware scheduler to handle link contention and skewed, changing traffic. On four platforms with up to 256 GPUs, the authors report average AlltoAllv speedups of 5.88x over FAST and 1.72x over DeepEP. End-to-end, Qwen3 inference latency falls by up to 27.2% and Wan2.1 generation latency by 6.1%.
In briefHyDra: a scalable system that fixes load imbalance in dynamic context parallelism for long-context training
SummaryThe paper studies dynamic context parallelism (DCP) for long-context model training, where Megatron-Core DCP sizes parallelism to fit memory, causing uneven computation across ranks. In a production 256K-context job on over 11K GPUs, per-rank microbatch times differed by up to 5x. The authors present HyDra, a load-driven scheduler combined with a nested Ulysses-in-ring CP engine. On a 2,048-GPU production job, HyDra reduced the pipeline bubble from 36% to 14% and improved throughput by 1.10 to 1.43x over Mcore DCP.
In briefHeddle: Learning Structural Templates for Parallelism Planning on Heterogeneous GPU Clusters
SummaryHeddle is a learning-based planner for training large machine learning models on heterogeneous GPU clusters. It reduces the planning problem to searching over pipeline structures, learning offline to build plans from structural templates across diverse cluster configurations. Evaluation reports that Heddle matches or exceeds the best of five existing planners, with throughput gains of up to 84.5% on dense models and 4.6x on MoE models.
In briefRIS-assisted multiuser NOMA networks robust to hardware impairments and imperfect channel information
SummaryThis paper studies RIS-assisted single-beam and multi-beam NOMA networks affected by transceiver hardware impairments and imperfect channel state information. The authors propose a robust scheme that minimizes base station transmit power while meeting users' target rates, solved with Bernstein-type inequality and semi-definite relaxation. Numerical results show the proposed scheme outperforms a non-robust RIS-assisted NOMA scheme that ignores these impairments, and multi-beam NOMA achieves higher beam gain than single-beam NOMA.
In briefOffline-designed variances for parallel perturbed successive cancellation decoding of polar codes
SummaryThis paper proposes OVD-PSCP, a parallel successive cancellation perturbation-enhanced decoder for short and medium polar codes, in which the perturbation variances are chosen offline before decoding. Because all branches, including the original SC branch, can start simultaneously, the method avoids the extra latency of schemes that generate perturbations from earlier decoding outcomes. Simulations for rate-1/2 polar codes of lengths 64 to 512 show lower block error rates than conventional SCP with the same number of perturbation branches, with larger gains for shorter codes and more branches.
In briefRepairing DAG workflow schedules after link disruptions
SummaryThis paper studies how to repair schedules for DAG workflows in networked IoT systems when link disruptions invalidate parts of them. It compares four repair scopes: waiting out the disruption, rerouting, rescheduling only affected tasks, and rescheduling all pending tasks globally, each measured against an oracle and charged a latency cost. Across 100 workload instances and different disruption processes, no single scope wins, and the authors conclude the repair scope should adapt to the disruption process and repair cost.
Mohammadali Khodabandehlou, Jared Coleman, Bhaskar Krishnamachari et al.
In briefProvably efficient, manipulation-resistant linear regression: tight bounds
SummaryThe paper studies the trade-off between prediction accuracy and incentive compatibility in linear regression, where agents privately report labels and prefer predictions close to their true values. The authors give a deterministic group-strategyproof mechanism achieving a (d+1)-approximation to the least-squares optimum and show this is optimal even among universally strategyproof randomized mechanisms, answering an open question from Chen et al. (EC 2018). Under strategyproofness in expectation, squared individual loss keeps the factor d+1, while absolute individual loss achieves the tight ratio 2-1/(⌈d/2⌉+1).
WhyResolves an open question of Chen et al. with tight approximation ratios and optimality proofs for strategyproof linear regression mechanisms.
Yichen Huang, Yuqi Pan, Michael Mitzenmacher et al.
In briefGeographic Incentives of Multiple Concurrent Proposers
SummaryThis paper studies whether systems with multiple concurrent proposers (MCP) incentivize proposers to decentralize geographically. Proposers strategically choose locations to capture user market share, since users pick where to send transactions based on time to inclusion. The authors formally prove that proposers' quorum radii, their distance to the nearest quorum of attesters, determine their incentives, so proposers minimize this radius. This yields co-location when proposers also act as attesters, but geographic diversity when those roles are separate. Pre-confirmations and faster intervalidator connections are shown to weaken the quorum radius effect and further destabilize co-location incentives.
In briefPerformance vs Portability in HPC: Why Benchmark Before Running
SummaryThis paper compares running quantum chemistry software in Podman containers against running it directly on HPC systems, using precompiled executables with CPU-optimized builds across multiple compute nodes and no extra network tuning. The results show that performance depends on how the software was compiled, startup overhead, and execution configuration. Container preparation can dominate short tasks, while long calculations need progress monitoring. The authors conclude that workload-specific benchmarking and user-level monitoring are important when deploying scientific applications in containers on HPC systems.
In briefSoft-Output GRAND-Aided Macrosymbol: A Joint Detection and Decoding Method for NOMA
SummaryThis paper extends GRAND-AM, a coding-based joint multiuser detection and error correction method for NOMA systems, into a soft-output version called SOGRAND-AM. It produces calibrated bitwise a posteriori error probabilities and works with soft-decision outer forward error correction decoders. According to the paper, SOGRAND-AM achieves about a 2 dB bit error rate improvement at the outer soft-input decoder's output compared with hard-input GRAND-AM.
Sarah Khalifeh, Alexander Mariona, Ken R. Duffy et al.
In briefStatistical Dark Matter: Synergy Is Everywhere but Hard to Detect
SummaryThis paper studies synergy, an information-theoretic property found only in three or more variables jointly and absent from any subset of them. Using recent advances in multivariate information theory, the authors show synergy is far more common in complex systems than previously thought, and that it tends to dominate as the number of components grows. They also find that commonly used statistical modelling tools substantially underestimate these synergistic structures.
Alberto Liardi, Fernando E. Rosas, Daniele Marinazzo et al.
In briefProof of the Braun-Etzion-Vardy bound for binary subspace codes
SummaryLinear subspace codes are codes defined within projective spaces. Braun, Etzion and Vardy (2013) conjectured that a linear subspace code in projective space has at most 2^n codewords. This paper resolves that conjecture. It uses a novel character-theoretic argument and the proof is self-contained.
WhyResolves an open conjecture via a novel character-theoretic argument with a complete, self-contained proof.
