In briefRethinking KV Cache Eviction for Agent LLM Serving
SummaryThis paper addresses memory pressure from key-value caches in large language model serving, which requires eviction when capacity runs short. It proposes keeping shared prefixes, evicting cache from the start of the remaining history, and retaining the cached suffix, and builds a system called Penelope around this idea. Implemented in SGLang and tested on three Qwen3 model sizes with real agent traces, it reduces mean time to first token by up to 60% and job completion time by up to 37% with a GPU-only cache. With a two-tier GPU-CPU cache, the reductions reach up to 76% and 60%.
In briefThe Pareto Secretary Problem: online selection of Pareto-optimal observations
SummaryThis paper studies an online selection problem in which observations arrive one at a time from an i.i.d. sample, and the goal is to pick one that is Pareto-maximal, meaning no other observation is at least as good in every coordinate and strictly better in one. It determines the optimal failure-probability order for fixed dimensions, covering independent coordinates and fixed positive definite Gaussian covariances. It also reports results under arbitrary dependence, where optimal success can tend to about 0.58.
In briefTransfer principles for gradient-based learning from simple scalar games to large network zero-sum games
SummaryThe paper studies a broad class of gradient-based learning rules in zero-sum games. It shows that if a rule converges on all scalar zero-sum games, it also converges on any network zero-sum game, under a step-size condition tied to the norm of the game operator. Finite-time rates also transfer mode by mode, so the worst-case network response equals the largest scalar rate. Thus, analyzing scalar games is enough to characterize convergence over the network class.
In briefEmbodied Movable Antennas: Reconfiguring Wireless Channels Using a Physical Agent's Mechanical Redundancy
SummaryThis paper proposes embodied movable antennas (MAs), which change antenna position and orientation using the mechanical degrees of freedom already present in physical agents rather than dedicated actuators. The authors distinguish task-coupled and task-preserving modes and use a redundant robotic manipulator as an example, deriving an exact Fourier representation of the multipath channel. Numerical results show substantial communication gains from task-preserving motion.
In briefCapacity and the Four Subspace Problem: Distributed Linear Computation and Cross-Ratio
SummaryThe paper studies a setting where several users each observe a fixed linear function of common data, and a server must recover another linear function with zero error from one message per user. It examines the smallest case, three users observing a, b, and a+b while the server wants a+λb, whose four functions form lines with cross-ratio λ. The capacity depends on λ and the prime subfield: it equals 2 for harmonic cross-ratios, 3/2 for λ outside the prime subfield, and lies in [3p/(2p-1), 2] for other prime-subfield values.
In briefExtremal character spectra of finite projective systems (a math paper on how many distinct values a point multiset's character function can take on flats)
SummaryThe paper studies finite spanning multisets of points in the projective space PG(k-1,q) and the character function of such multisets. Under a total-multiplicity budget, it examines the maximum number of distinct values the character function takes on codimension-r flats. The case r=k-1 is solved completely, including minimum-cost multiplicity profiles. For r=1, the known upper bounds on the minimum length of these systems are improved.
In briefMatching Workloads to Blackwell: Power-Resolved Benchmarks of NVIDIA B200 and B300 for HPC and AI
SummaryThis paper benchmarks NVIDIA's B200 and B300 (Blackwell Ultra) GPUs on two separate eight-GPU clusters, covering computer-vision inference and training, Llama-3 LLM inference, and six production HPC codes. It finds that the B300 is not a uniform successor to the B200, trailing it by up to several times on some workloads such as convolutional inference and double-precision Quantum ESPRESSO. The authors use per-GPU power draw to explain the convolutional gap and provide workload-specific procurement guidance.
In briefDecentralized Dynamic Stable Matching: Computing and Learting Optimal Stable Matchings Without Clocks or Direct Communication
SummaryThis paper studies how players can decentrally compute the player-optimal stable matching when players arrive and depart unannounced, without a common clock or direct communication. For known preferences, it gives a deterministic polynomial-time self-calibrating protocol with improved cost bound O(t_*+K^2 log T). For unknown preferences, it presents a decentralized learning protocol with expected cost O(t_*+K^4 Δ_min^{-2} log T), and a lower bound shows the dependence on T and Δ_min is tight.
WhyDecentralized dynamic stable matching without clocks or communication; the authors report improving cost from O(T^{1-1/K}) to O(t*+K^2 log T) with a matching lower bound.
In briefApproximate-support Donoho–Stark uncertainty principles for k-Hadamard operators
SummaryThis paper studies Donoho–Stark type uncertainty principles for k-Hadamard operators in finite dimensions, building on earlier work by Wigderson and Wigderson. It examines versions for approximate supports under ℓ^p norms. The authors show such results cannot extend to p>1 for matrices that are not special k-Hadamard. For special k-Hadamard operators, they prove the principle holds for 1≤p≤2, while showing it fails for p>2 even under the speciality condition.
In briefLayered reproducibility for HPC: a Feel++ and Ktirio Urban Buildings case study
SummaryThis paper addresses the difficulty of reproducing high-performance computing work, caused by complex software stacks, heterogeneous hardware and costly validation. It proposes a layered strategy in which Feel++ provides a traceable software supply chain and Ktirio Urban Buildings (KUB) builds on it to manage versioned urban-building data, executable simulations and checksum-verifiable results. The approach moves reproducibility from late manual packaging to an artifact chain linking source commits, packages, containers, benchmark configurations, dataset versions, provenance records and generated results.
Javier Cladellas, Vincent Chabannes (IRMA), Christophe Prud'Homme (IRMA)
In briefAchievable rates and nearly lossless state transmission over a multi-access channel with causal state information
SummaryThis paper studies transmitting both data and state over a memoryless state-dependent multi-access channel where encoders have causal state information. The state has a common component known to both encoders and two private components, each known to one encoder. Data must be recovered with arbitrarily small error probability, and the state sequence with arbitrarily small symbol-error rate. Assuming the capacity region without the state reconstruction constraint is nonempty, the authors derive the capacity region and show it is achievable using a Block-Markov scheme with binning and superposition coding.
In briefBatch universal prediction via conditional additive smoothing
SummaryThe paper studies batch regret, a measure of how well a predictor performs on a fresh test sequence after observing a batch of training sequences from the same unknown source. It analyzes conditional additive smoothing estimators for binary memoryless and first-order Markov sources, deriving their asymptotic batch regret. For memoryless sources, the optimal bias and resulting regret are shown to vary with the training batch size, extending Krichevsky's Poisson-based analysis. The authors also introduce a Rényi analogue of batch regret and identify the optimal predictors as conditional α-NML estimators.
