In briefPipelining vs. Combining: Two Strategies for Aggregation in Networks
SummaryThis paper studies aggregation (convergecast) in networks when the values to be combined are too large for a single message and must be sent as pipelined message trains. The authors give general lower bounds and optimal algorithms for architectures such as hypercubes and complete bipartite graphs, under splittable and unsplittable flow models. They also present an O(log n) approximation for general graphs in the unsplittable flows model.
Gabriel Marques Domingues, Claire Mathieu, Boaz Patt-Shamir
In briefSum-rank versions of Reed–Muller codes over Abelian Galois extensions
SummaryThis paper studies sum-rank metric codes over finite Abelian Galois extensions. The authors show that a quotient of a multivariate skew polynomial ring is isomorphic to a product of copies of End_K(L), so one skew polynomial can represent all blocks of a sum-rank space. They define the sum-rank weight via the rank of Dickson and generalized Moore matrices. They then define sum-rank Reed–Muller-type codes and give formulas for their dimension and minimum distance, which include linearized Reed–Solomon, Gabidulin, and rank-metric Reed–Muller codes as special cases.
In briefConstructing Non-GRS MDS Codes Using Imaginary Secant Lines
SummaryThe paper constructs non-GRS MDS codes from projective systems made of n-2 points on a rational normal curve plus a pair of Galois conjugate points on an imaginary secant. For odd q, the code lengths satisfy k+4 ≤ n ≤ (q+5)/2, improving the attainable lengths and dimension ranges of earlier TRS and CTRS constructions. A geometric argument establishes the non-GRS property for 3 ≤ k ≤ n-4, and the structure of the dual of any CTRS code is determined.
In briefNearly tight distributed approximation for dominating sets in sparse graphs
SummaryThis paper studies distributed approximation of the minimum dominating set problem on sparse graphs, extending near-optimal results from minor-free graphs to bounded expansion classes. It gives a deterministic constant-round LOCAL algorithm with approximation ratio d+1+ε, which is optimal up to ε for even d≥2. This yields 7+ε approximations for planar and bounded Euler genus graphs, and 5+ε for their triangle-free subclasses.
In briefLLMs and Bayesian Optimization for Self-Evolving Wireless Digital Twins
SummaryThis paper explores combining large language models with Bayesian optimization to let wireless digital twins evolve their optimization algorithms, not just adapt online. Bayesian optimization serves as a stable, sample-efficient backbone, while the LLM decides whether to keep or adapt its components based on the task and past experience. In simulations on a multi-user wireless beam-selection problem using Gaussian process kernel discovery, the LLM-discovered kernel improved final performance by 19.7% on training tasks over the strongest standard kernel baseline.
In briefA local-to-global principle for f-divergence homogenization
SummaryThe paper studies what happens when a family of independent statistical experiments is replaced by independent repetitions of their average. It gives a simple, complete characterization of the f-divergences for which this increases divergence by at most a constant depending only on the generator, uniformly in the number of observations and the measurable space. The criterion follows a local-to-global principle: a comparison of two observations on a binary alphabet determines whether the general inequality holds.
In briefBalanced fixed prices for a two-sided market with limited storage
SummaryThe paper studies an intermediary that buys storable items from Poisson-arriving sellers and sells them to Poisson-arriving buyers under limited storage, using the same fixed prices for both sides. For buyers with XOS valuations, the authors compute static common prices in polynomial time that guarantee half of the long-run gains-from-trade benchmark, even with one storage slot per item. For a single item with capacity C, the guarantee is C/(C+1), which is shown to be tight.
WhyIntroduces balanced posted prices with a tight C/(C+1) guarantee for single items and half-benchmark guarantees for XOS buyers, extending prophet-inequality ideas to recurring two-sided markets.
In briefRotating the array in near-field ISAC to balance communication, detection and estimation
SummaryThis paper studies near-field integrated sensing and communication with a rotatable base-station array under an exact spherical-wave model. It derives a Cramér–Rao bound decomposition and analyzes two low-complexity transmit architectures, a single-stream superposition design and a dual-waveform split design. Numerical results show clear gains over a fixed array orientation, especially for range estimation.
In briefExtending MDS Codes over Finite Fields to Longer Non-Reed–Solomon Codes
SummaryThis paper studies how to extend non-Reed–Solomon (non-RS) [n,k,n-k+1] MDS codes into longer [n+1,k,n-k+2] non-RS MDS codes. It proposes an iterative algorithm, inspired by meet-in-the-middle methods in cryptanalysis, that preserves the non-RS structure. Applied to several known code families over fields of order at most 32, it yields new MDS codes inequivalent to Reed–Solomon codes, including [9,4,6] codes over F9 and F13 and two [10,4,7] codes over F17.
In briefAssigning Tasks to Agents: Mechanism Design for Multi-Agent Contracts
SummaryThis paper studies task assignment when agents have private types and may take hidden effort, under adverse selection and moral hazard. For a single task, the authors show that incentive constraints depend only on the ratio of effort cost to success probability, yielding a simple Myerson-style characterization and a polynomial-time optimal mechanism. They establish a tight worst-case bound relative to the full-information benchmark. For multiple tasks, an optimal mechanism may not exist, but for a constant number of agents a polynomial-time approximation scheme for the supremum profit is provided.
Matteo Bollini, Matteo Castiglioni, Alberto Marchesi
In briefProduct assortment optimization under a perception-adjusted choice model
SummaryThis paper studies assortment optimization under the Perception-Adjusted Luce Model (PALM), a sequential choice model that allows violations of regularity and choice overload. Flores et al. (2019) conjectured that revenue-ordered assortments by level remain optimal for any number of levels. The authors confirm this conjecture, showing every PALM instance has an optimal assortment including all products at or above a level-specific revenue threshold. They also show the problem is polynomial-time solvable when the number of levels is fixed.
WhyResolves a conjecture on revenue-ordered optimality under PALM for any number of levels, and proves polynomial solvability for fixed levels.
In briefCan a reciprocal RIS match a non-reciprocal RIS's performance?
SummaryThis paper studies whether reciprocal reconfigurable intelligent surfaces (RIS) can achieve the performance of non-reciprocal ones. It finds that this holds if and only if some optimal non-reciprocal RIS makes the round-trip transmitter-RIS-transmitter channel symmetric. The condition is satisfied when the transmitter-RIS channel has rank one or when single-stream transmission is optimal.
Matteo Nerini, Juan Vidal Alegr\'ia, Bruno Clerckx
In briefuGDS: A User-Space Toolkit for Faster GPU Direct Storage
SummaryThis paper presents uGDS, an open-source user-space library for direct data transfers between NVMe SSDs and GPUs, aiming to remove kernel overhead from each I/O request. It separates synchronous and batch traffic across hardware queues and offers cuFile-compatible APIs on NVIDIA and AMD platforms. The authors report higher batch and synchronous bandwidth and lower write latency than NVIDIA GDS on an A100 with a Samsung 990 PRO, along with reduced time-to-first-token in LMCache.
WhyRedesigns GDS's control path in user space and separates sync and batch queues, a substantial improvement over prior work, with thorough evaluation on NVIDIA, AMD, and LMCache.
In briefAn integer determinant method for the length of Type II codes and the weights of shadows
SummaryThe paper gives elementary proofs that the length of a binary Type II code is divisible by 8, avoiding invariant theory. The main tool is a congruence modulo 8 for the determinant of the integer Gram matrix of a binary orthogonal matrix. The same argument yields the Conway–Sloane result that shadow vectors of a self-dual binary code have weight congruent to n/2 modulo 4.
In briefA Mechanism for Matching with Ties on Both Sides that Resists Group Manipulation by Workers
SummaryThis paper studies many-to-one matching markets where both workers and firms may be indifferent between options, and firms have publicly known preferences over sets of workers. For markets where firm preferences admit a gross substitutes representation, the authors propose a deterministic generalized deferred acceptance mechanism. The mechanism is weakly stable, Pareto efficient for both sides, and weakly group strategy-proof for workers, and can be computed in strongly polynomial time given value oracles.
WhyBridges indifferent two-sided matching with GS competitive equilibria, yielding a deterministic mechanism with stability, efficiency, and incentive guarantees; strong theory but applicability is confi
In briefTwo-timescale antenna positioning for pinching-antenna systems using statistical channel information
SummaryThis paper proposes a two-timescale framework for positioning antennas in pinching-antenna systems. Large-timescale positions over the waveguide are chosen using statistical channel information, and small-timescale refinements within a local region use instantaneous channel information. Simulations show the scheme closely approaches derived bounds and outperforms conventional approaches.
In briefProof of the ternary and quaternary cases of the {2,4}-distance code conjecture, which concerns the largest code size when pairwise Hamming distances are only 2 or 4.
SummaryThe paper studies q-ary codes of length n in which any two distinct words have Hamming distance 2 or 4. The authors prove that for q in {3,4} and n at least 6, the maximum size is 1+C(n,2), settling the remaining alphabet sizes of a conjecture by Boyvalenkov, Delchev, Zinoviev, and Zinoviev; the binary case was already known. The quaternary upper bound is proved combinatorially for large n, while finite lengths are verified by exact computation.
Yen-Chi Roger Lin, Valjakas Djaljapayan, and Wei-Hsuan Yu
In briefKnowing When to Hand Off: Critic-Guided Adaptive Execution for Cloud-Local Embodied Agents
SummaryThis paper presents CARE, a framework in which a lightweight critic model monitors a local executor and hands control to a stronger cloud model when local execution becomes unreliable. A shared handoff memory keeps execution context so the cloud model can intervene without restarting. Across diverse embodied benchmarks, CARE raises task success by 6.3 to 18.3 percentage points over the strongest collaboration baseline while approaching or exceeding cloud-only success.
In briefClassification of binary, ternary and quaternary Hermitian LCD codes
SummaryThis paper extends the complete classification of linear complementary dual (LCD) codes to binary codes of length 15 to 19, ternary codes of length 10 to 15, and quaternary Hermitian codes of length 12 to 13. The method reduces monomial equivalence of LCD codes to graph isomorphism on the coordinates and combines it with canonical augmentation.
In briefRepair changes which layers to prune: exact optima for depth pruning of LLMs
SummaryThis paper studies depth pruning of large language models, which removes whole transformer layers and replaces them with a training-free linear map. The authors exhaustively evaluated about 53,000 pruned models and found that the best layers to remove depend on whether repair is applied, since linear repair turns the optimum into a contiguous block. Methods that ignore the repair can miss the repaired optimum by up to 189%. Selecting layers under the repair that will be applied lowers perplexity by 20% to 66% relative to ReplaceMe on five models.
In briefConway's Game of Life dynamics mapped on an entropy-complexity plane
SummaryThis paper studies Conway's Game of Life using ordinal pattern analysis. Each generation is converted into a one-dimensional time series, and the resulting ordinal patterns are plotted on an entropy-statistical complexity plane. The authors compare encoding choices such as patch size and linearization method and examine how they affect the trajectories. They characterize the emergent dynamics through convergence behavior, trajectory patterns, and statistical complexity.
Lucas Malacarne Astore, Alejandro C. Frery, Petr\^onio C\^andido de Lima e Silva et al.
In briefOutage analysis and two-level resource optimization for RIS-assisted clustered UAV-NOMA networks
SummaryThis paper studies downlink UAV communications assisted by reconfigurable intelligent surfaces (RIS) and based on NOMA. It proposes a cluster-based framework and a greedy RIS-aware clustering method, and analyzes received SNR and outage under Nakagami-m and double Nakagami-m fading with imperfect SIC. A bilevel efficiency-reliability optimization problem is formulated and solved with a two-level iterative algorithm. Numerical results, baseline comparisons, and ablation studies are reported.
In briefHow Far Price Discrimination Can Go When the Market Is Unknown
SummaryThe paper considers a seller who knows only the set of possible consumer values, not their distribution. It characterizes the consumer and producer surplus pairs that some market and some segmentation can produce, which form a convex polygon. With a continuum of values up to v̄, consumer surplus is at most π ln(v̄/π) when producer surplus is π, a bound reached without segmentation.
In briefCutting a Cake for Human Consumption: a simple protocol for fairly dividing cake among n people
SummaryThe paper presents a bounded envy-free cake-cutting protocol that works for any number of agents n in the Robertson–Webb query model. It uses 2^{O(n log n)} queries and comes with a short, self-contained correctness proof. The authors note that most of the ingredients are already known in fair division research, and their main contribution is organizing these ideas into a protocol that can be fully described and proved in a few pages.
In briefReal-time VAT compliance for split payments using Open Banking and a DLT architecture
SummaryThis paper presents MTS-OB, a DLT-based architecture built on R3 Corda that integrates real-time VAT compliance into payments through split payments and Open Banking APIs. As a proof-of-concept, it enables automated real-time VAT split payments without disrupting existing payment infrastructure. Scalability tests show the system remained operationally stable as transaction loads increased. The authors conclude that real-time tax compliance requires redefining money to include its informational value.
In briefOptimal (v,5,3,1) optical orthogonal codes: a paper on the best possible size of certain codes
SummaryThe paper establishes a tight upper bound on the size of (v,5,3,1) optical orthogonal codes and calls codes that reach it optimal. An optimal code is called perfect if its codeword-set differences cover all non-identity elements of Z_v. The authors present direct and recursive constructions for infinite classes of optimal and perfect (v,5,3,1)-OOCs.
Laura Johnson, Lorenzo Mella, Anita Pasotti et al.
In briefWhen Majority Voting Fails: Recovering Binary Sequences from Noisy Copies with Fixed Error Counts
SummaryThis paper studies how to recover a binary sequence when it is sent several times over a channel, each time with exactly t substitution errors. It determines the minimum number of distinct outputs needed for an optimal decoder to reconstruct the sequence uniquely. The authors show that bit-wise majority voting, known to be optimal for errors of up to t bits, is suboptimal in this setting, and they propose an alternative algorithm that needs fewer observations.
Vivian Papadopoulou, Yonatan Yehezkeally, V. Arvind Rameshwar et al.
In briefEnumeration-based triangle counting on GPUs
SummaryThis paper presents a distributed-memory implementation for exact triangle counting, which counts closed length-3 cycles in a graph. The authors evaluate it on billion-edge real-world graphs using scale-out and scale-up NVIDIA GPU platforms based on Ampere and Hopper architectures. Compared with existing approaches, the reported end-to-end execution times improve by up to 10 times over a distributed CPU baseline.
In briefCapacity limits of secure summation when parties can exchange messages over several rounds
SummaryThe paper studies interactive secure summation, where K parties with private inputs compute their sum at one party while resisting up to T colluding parties. It characterizes the capacity region for communication and randomness rates. For 0≤T≤K−2, the optimal communication rate is K(K−1)/(K−T), and the optimal total randomness rate is T(K−1)/(K−T). Linear schemes achieve these optimal rates.
WhyExtends secure summation to multi-round interaction and fully characterizes the capacity region with rigorous information-theoretic proofs, impacting mainly information theory and secure computation.
In briefNashtopes and Their A-Discriminants: Studying how game equilibrium equations can degenerate
SummaryThe paper studies the mixed discriminant of the indifference system that determines Nash equilibria in normal-form games, viewing it as an A-discriminant whose associated polytope is called the Nashtope. For every balanced game format with at least three players, the authors compute the codimension and degree of this discriminant. Combining these results with earlier work, they obtain the codimensions and degrees of the irreducible components of the Nash discriminant variety.
In briefHAWK: A Transfer Mechanism for Moving KV Cache Between GPUs and NPUs in Split LLM Inference
SummaryThis paper proposes splitting large language model inference so GPUs handle the compute-heavy prefill phase and NPUs handle the memory-bound decode phase. It introduces HAWK, a mechanism that moves KV cache between these different devices over RDMA, using NPU-side buffers matched to the target tensor layout so data can be used without extra rearrangement. On commercially available GPUs and NPUs, the authors report performance within 7.88% of a GPU-only system under the same SLO constraints, with up to 1.26× higher power efficiency and 1.72× higher cost efficiency.
