# characterization Published articles for characterization. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning DevFeed: [How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning]() Original publisher: [Read original article]() Author: Max W. Shen, Mark Goldstein, Zichu Wang, Aahlad Puli, Rajesh Ranganath Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Reinforcement learning]() Tags: [case-study](), [characterization](), [convergence](), [cs-lg](), [diffusion](), [expression](), [flow](), [functional](), [gradient](), [grounding](), [guarantees](), [learning](), [machine](), [machine-learning](), [map](), [maps](), [memory](), [models](), [objectives](), [optimizing](), [point]() ### AI overview The paper proposes a stopgrad regression principle that characterizes stationary points and their uniqueness across objectives for flow maps, reinforcement learning, and diffusion samplers. For flow map learning, it reports that the unique stationary point is the true flow map and establishes positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Under functional semi-gradient flow, it derives a closed-form learned flow map by composing the initial and true flow maps. It also proposes changed stopgrad placements for flow map objectives that the authors report reduce training memory by 2x. ### Source excerpt arXiv:2609.16222v2 Announce Type: replace Abstract: Stopgrads are widely used in training machine learning models, but stopgrads can alter the gradient, stationary points and convergence guarantees of the original objective, which can make stopgrad training theoretically ungrounded. We introduce a stopgrad regression principle, which identifies a general template for stopgrad objectives with a closed-form characterization of stationary points and their uniqueness, unifying stopgrad objectives for flow maps, reinforcement learning, and diffusion samplers. We provide theoretical grounding for optimizing stopgrad flow map objectives by showing their unique stationary point is the true flow map, and showing positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Remarkably, we show that under functional semi-gradient flow, the learned flow map has a closed-form expression composing the initial flow map and the true flow map. We additionally use our stopgrad regression principle to propose modified stopgrad placements for flow map objectives which reduce training memory by 2x. ## Privacy Filters are Captured by Residues: A Characterization of Free Natural Filters and the Cost of Adaptivity DevFeed: [Privacy Filters are Captured by Residues: A Characterization of Free Natural Filters and the Cost of Adaptivity]() Original publisher: [Read original article]() Author: Matthew Regehr, Bingshan Hu, Ethan Leeman, Pasin Manurangsi, Pierre Tholoniat, Mathias L\\'ecuyer Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [Cryptography](), [Security & Privacy]() Tags: [adaptive](), [adversary](), [budget](), [capture](), [characterization](), [cost](), [cs-cr](), [cs-ds](), [develop](), [factor](), [fail](), [family](), [filter](), [free](), [gaussian](), [leveraging](), [privacy](), [privacy-filter](), [private](), [profile](), [queries](), [techniques]() ### AI overview The paper characterizes worst-case privacy loss under adaptive interaction with privacy filters using a Bellman equation, then develops residue filters that unify existing approaches. Its Gaussian DP residue filter strictly improves on the naïve GDP filter; in the average case, it asymptotically permits as many pure-DP queries as the non-adaptive setting and π/2 times as many as the naïve GDP filter. The authors also find that natural filters, which account for each query's full privacy profile, are not generally free: free filters for arbitrary budgets require query families totally ordered under composition. Under an adaptive adversary, the natural approximate-DP filter can exceed its intended parameters, but the excerpt bounds the deterioration as at most polylogarithmic. ### Source excerpt arXiv:2602.15815v3 Announce Type: replace Abstract: We study privacy filters, which enable privacy accounting for differentially private (DP) mechanisms with adaptively chosen privacy characteristics. We show that the worst-case privacy loss of interacting with any privacy filter is characterized by a Bellman equation. We apply this theory to develop residue filters, which unify existing privacy filters. We develop the Gaussian DP (GDP) residue filter, which strictly improves upon the na\"ive GDP filter. We show that, in the average case, this new filter asymptotically permits as many pure DP queries as in the non-adaptive case and a factor $\pi/2$ more pure DP queries than the na\"ive GDP filter. We also show that residue filters capture the natural filter, which promises greater utility by leveraging exact privacy accounting techniques. Earlier privacy filters consider only simple privacy parameters such as R\'enyi-DP or GDP parameters. Natural filters account for the entire privacy profile of every query, promising more efficient use of a given privacy budget. We show that, contrary to other forms of DP, natural privacy filters are not free in general. We present a characterization of when a family of private queries admits free natural filters for a given budget. In particular, only families of privacy mechanisms that are totally-ordered when composed admit free natural privacy filters with respect to an arbitrary privacy budget. Finally, we show that, while the natural approximate-DP filter can fail in the presence of an adaptive adversary, it cannot fail too badly: the output remains approximate-DP with parameters at most poly-logarithmically worse than the intended privacy parameters. ## Generalized Priority-Aware Shapley Value DevFeed: [Generalized Priority-Aware Shapley Value]() Original publisher: [Read original article]() Author: Kiljae Lee, Ziqi Liu, Weijing Tang, Yuan Zhang Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Large Language Model]() Tags: [apply](), [arena](), [boundary](), [characterization](), [chatbot](), [comparisons](), [covers](), [cs-ai](), [cs-lg](), [data](), [develop](), [examples](), [graph](), [graphs](), [learning](), [llm](), [machine-learning](), [models](), [order](), [priority](), [random](), [value]() ### AI overview The generalized priority-aware Shapley value (GPASV) handles arbitrary directed, weighted priority graphs by penalizing order violations instead of forbidding them, addressing cyclic or nonbinary priorities that existing methods cannot represent. The authors give an axiomatic characterization, computational methods, and a priority-sweeping diagnostic, and show that classical models arise as boundary cases. In an application to LLM ensemble valuation on the cyclic Chatbot Arena preference graph, different balances between graph priority and individual soft priority produce substantively different valuations of the same data; priority-aware valuation therefore requires choosing that balance. This overview is based on the supplied abstract. ### Source excerpt arXiv:2605.15018v2 Announce Type: replace Abstract: Shapley value and its priority-aware extensions are widely used for valuation in machine learning, but existing methods require pairwise priority to be binary and acyclic, a restriction spectacularly violated in real-data examples such as aggregated human preferences and multi-criterion comparisons. We introduce the generalized priority-aware Shapley value (GPASV), a random order value defined on arbitrary directed weighted priority graphs, in which pairwise edges penalize rather than forbid order violations. GPASV covers a range of classical models as boundary cases. We establish GPASV through an axiomatic characterization, develop the associated computational methods, and introduce a priority sweeping diagnostic extending PASV's. We apply GPASV to LLM ensemble valuation on the cyclic Chatbot Arena preference graph, illustrating that priority-aware valuation is not a one-button operation: different balances of pairwise graph priority versus individual soft priority produce substantively different valuations of the same data. ## Gain-Dependent Entropy Characterization of a Zener Avalanche Noise TRNG via Dual NIST Standards: Toward Low-Cost Hardware Entropy for ICS Deployments DevFeed: [Gain-Dependent Entropy Characterization of a Zener Avalanche Noise TRNG via Dual NIST Standards: Toward Low-Cost Hardware Entropy for ICS Deployments]() Original publisher: [Read original article]() Author: Rishabh Das Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [federal information processing standards]() Tags: [22](), [abstract](), [all](), [amp](), [amplifier](), [amplifier-gain](), [arxiv](), [assessment](), [avalanche](), [based](), [bias](), [characterization](), [cost](), [cs-ar](), [cs-cr](), [deployments](), [entropy](), [hardware](), [ics](), [nist](), [noise](), [standards]() ### AI overview The study reports that a low-cost Zener avalanche noise TRNG's entropy quality depends on both amplifier gain and coupling capacitance. Across five gain settings and two capacitor values, it evaluates 1.44 billion raw bits per configuration using NIST SP 800-22 and SP 800-90B. A three-phase autocorrelation pattern appears at every gain, attributed to comparator saturation and recovery. Replacing the Zener diode with a stable voltage source cuts switching activity by about 97%, supporting avalanche noise as the necessary entropy source. The smaller coupling capacitor improves entropy, throughput, and bias. The authors report one gain configuration with a perfect statistical test score, NIST-compliant min-entropy, and peak throughput, at a bill-of-materials cost below five dollars. ### Source excerpt arXiv:2611.08774v1 Announce Type: new Abstract: True random number generators based on Zener avalanche noise offer a low-cost entropy source for industrial control system and IoT edge devices, yet the relationship between amplifier gain, coupling capacitance, and entropy quality remains uncharacterized under current NIST standards. This paper presents a comprehensive entropy assessment of an op-amp-based Zener avalanche noise generator across five gain configurations and two coupling capacitor values, evaluated against the NIST SP 800-22 statistical test suite and the SP~800-90B entropy source assessment using 1.44 billion raw bits per configuration. We identify a three-phase autocorrelation structure, attributed to comparator saturation-recovery dynamics, consistent across all gain settings. A direct-current substitution experiment, replacing the Zener diode with a stable voltage source at the same bias point, reduces switching activity by approximately 97%, confirming that Zener avalanche noise, rather than amplifier self-excitation, is the necessary entropy source. The coupling capacitor is further shown to govern comparator overdrive and entropy quality independent of this source mechanism, with the smaller capacitance improving entropy, throughput, and bias. The combined results identify an optimal gain configuration achieving a perfect statistical test score, NIST-compliant min-entropy, and peak entropy throughput at a bill-of-materials cost under five dollars, offering practical design guidance for low-cost entropy sources in industrial and embedded deployments. ## Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral Characterization DevFeed: [Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral Characterization]() Original publisher: [Read original article]() Author: Na\\"il B. Khelifa, Richard E. Turner, Ramji Venkataramanan Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [AI Models](), [Neural Network](), [stable-diffusion]() Tags: [characterization](), [cifar-10](), [cs-lg](), [data](), [diffusion](), [diffusion-models](), [distribution](), [error](), [errors](), [faster](), [filter](), [fine](), [gaussian](), [generative](), [generative-models](), [infinite](), [learning](), [limit](), [mechanism](), [model](), [models](), [stat-ml]() ### AI overview Recursive diffusion-model training can drift from the data distribution even with perfect score estimation and exact sampling: early stopping the reverse process for numerical stability progressively smooths the distribution. The authors prove geometric convergence to a unique limit, characterized as an infinite mixture of increasingly Gaussian-smoothed data distributions. A Hermite spectral analysis shows stronger attenuation of higher-order modes, which encode fine non-Gaussian structure, than of coarse modes. They prove that truncation schedules whose stopping times shrink toward zero across rounds asymptotically eliminate compounding. With discretization and score-estimation errors, the learned distribution stays within a Wasserstein-2 ball of the ideal limit, with higher-order errors contracting faster. The theory is validated on synthetic Gaussian mixtures and CIFAR-10. ### Source excerpt arXiv:2606.13796v2 Announce Type: replace-cross Abstract: Recursive training of generative models on their own outputs can lead to model collapse, a compounding drift away from the true data distribution. Existing theoretical works bound finite-round error accumulation in the context of diffusion models, but two questions remain open:~what distribution does the recursion converge to, and how fast? We answer both, isolating a mechanism distinct from imperfect learning: even with perfect score estimation and exact sampling, the early stopping of the reverse diffusion (required for numerical stability) drives a progressive drift away from the data distribution. We prove that this recursion converges geometrically to a unique limiting distribution, which admits a closed-form characterization as an infinite mixture of increasingly Gaussian-smoothed versions of the data distribution. A Hermite spectral decomposition of this limit reveals that recursive training acts as a low-pass filter: higher-order modes, which encode fine non-Gaussian structure, are attenuated much more strongly than coarse modes. This spectral picture motivates annealed truncation schedules that progressively shrink truncation times across retraining rounds; we prove that any schedule converging to $0$ asymptotically eliminates recursive compounding. Finally, we show our idealized characterization is robust: in the presence of discretization and score estimation errors, the learned distribution remains in a Wasserstein-2 ball around the ideal limit, with mode-dependent contraction rates that contract high-order errors faster than low-order ones. We validate the theory on synthetic Gaussian mixtures and CIFAR-10. ## The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention DevFeed: [The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention]() Original publisher: [Read original article]() Author: Ye Yuan Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [dpo](), [ai alignment](), [Fine-tuning](), [post-training](), [lora]() Tags: [alignment](), [characterization](), [cs-lg](), [dpo](), [financial](), [format](), [json](), [lora](), [margin](), [measurement](), [model](), [objectives](), [qwen3-5](), [retention](), [sft](), [synthetic](), [tasks](), [tax](), [testing]() ### AI overview In a study of Qwen3.5-4B on three structured financial-filing tasks, increasing the implicit reward margin under length-normalized DPO with an SFT retention term produced task-selective capability loss. Across 8 seeds, extraction was preserved within a ±5-point equivalence margin (−0.55 points); JSON performance fell 3.52 points, with equivalence not established; and router performance fell 30 points, mainly in a margin 1-3 damage window that included a transient format-break spike. At least 96% of margin growth came from suppressing rejected completions while chosen reward stayed approximately flat; the study does not assign the behavioral damage to a single gradient channel. Measured reward margins differed by up to 3.4x between nf4 training and fp16 evaluation on the unquantized base. These findings are limited to one 4B model, sparse LoRA, easily separable synthetic preferences, and the tested objective; the authors leave a no-retention control and measured format-alignment metric for future work. ### Source excerpt arXiv:2611.06476v2 Announce Type: replace Abstract: Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing tasks, and track downstream capability as the implicit reward margin grows to $\sim 7$ $\beta$-scaled units (nf4 training; $\sim 3$ under fp16 measurement on the unquantized base). Question. Does margin growth degrade capability uniformly across tasks, or selectively? Findings. In a pre-registered study (8 seeds, 6 unseen at freeze) with artifact-aware scoring and TOST equivalence testing, the alignment tax is graded and task-selective: extraction is preserved ($\Delta = -0.55$ pts, equivalent at $\pm5$ pts), json loses little ($\Delta = -3.52$ pts, equivalence fails on a clustered item-level interval), and router degrades severely ($\Delta = -30$ pts). Router's loss concentrates in a margin 1-3 damage window containing a transient format-break spike. The margin is accounted for almost entirely ($\geq 96\%$) by suppression of rejected completions, with chosen reward approximately flat, consistent with the retention term anchoring it; we do not attribute the behavioral damage to any single gradient channel. Gold-logp deflation is task-selective (extraction $-0.15$, json $-1.04$, router $-1.47$ nats/token under fp16), with gold selectively protected relative to incorrect controls on extraction (8/8 seeds). QLoRA reward margins differ up to $3.4\times$ between nf4 training and unquantized fp16 measurement, with an axis-dependent gap. Scope. Length-normalized rewards with SFT retention, sparse LoRA on 8/32 layers, one 4B model, easily-separable synthetic preferences, three structured tasks. A no-retention control and a measured format-alignment metric are left to future work. ## Invariant Features in Language Models: Geometric Characterization and Model Attribution DevFeed: [Invariant Features in Language Models: Geometric Characterization and Model Attribution]() Original publisher: [Read original article]() Author: Agnibh Dasgupta, Abdullah Tanvir, Xin Zhong Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Representation learning](), [Embeddings](), [AI Chat]() Tags: [characterization](), [cs-cl](), [cs-lg](), [framework](), [invariant](), [language-models](), [machine-learning](), [zero-shot]() ### AI overview The paper models paraphrase invariance as a local geometry: equivalent inputs vary along nuisance directions while preserving semantic identity in invariant latent subspaces. It presents a geometric characterization of these features, a contrastive method to discover subspaces separating semantic-changing from semantic-preserving variation, and zero-shot model attribution using invariant representations. The abstract reports that invariant structure appears at specific depths, semantic displacement lies largely outside nuisance subspaces, and representation interventions indicate invariant components play a causal role in outputs. Model-specific geometric patterns also enable accurate attribution; details and limits of the experiments are not provided in this excerpt. ### Source excerpt arXiv:2605.06458v2 Announce Type: replace Abstract: Language models exhibit strong robustness to paraphrasing, suggesting that semantic information may be encoded through stable internal representations, yet the structure and origin of such invariance remain unclear. We propose a local geometric framework in which semantically equivalent inputs produce representations that occupy structured regions in latent space, with paraphrastic variation along nuisance directions and semantic identity preserved in invariant subspaces. Building on this view, we make three contributions: (1) a geometric characterization of invariant latent features, (2) a contrastive subspace discovery method that separates semantic-changing from semantic-preserving variation, and (3) an application of invariant representations to zero-shot model attribution. Across models and layers, empirical results support these contributions. Invariant structure emerges in specific depth regions, semantic displacement lies largely outside the nuisance subspace, and representation-level interventions indicate a causal role of invariant components in model outputs. Invariant representations also capture model-specific geometric patterns, enabling accurate attribution. Together, these findings support a local geometric characterization of semantic invariance in language model representations. ## Trust Responsive Runtime Authorization for Agentic AI with Empirical Characterization of Policy Behavior and Parameter Sensitivity DevFeed: [Trust Responsive Runtime Authorization for Agentic AI with Empirical Characterization of Policy Behavior and Parameter Sensitivity]() Original publisher: [Read original article]() Author: Ahsan Habib Shihab, Imre Felde, Ramis Akhmedov, Stefan Gubo, Berik Sabdenaliyev, Madina Mansurova, Amir Mosavi Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Dynamic authorization rules](), [Access Control](), [Agentic AI](), [trust](), [test]() Tags: [about](), [abstract](), [adaptive](), [agent](), [agentic](), [agentic-ai](), [ai](), [all](), [arxiv](), [authorization](), [autonomous](), [behavior](), [characterization](), [cs-ai](), [parameter](), [policy](), [responsive](), [runtime](), [runtime-authorization](), [trust](), [with]() ### AI overview TRAP maps a weighted Authorization Readiness Score--combining runtime trust, operational stability, and contextual risk--to discrete authority levels. Escalation requires consecutive observations above a band edge plus a margin; a sharp readiness drop restricts authority immediately. Across 100 matched Monte Carlo traces, it restricted authority at the onset of abrupt degradation and re-escalated sooner after recovery than a symmetric policy. It had the highest and descriptively least variable stability among adaptive policies, but more post-recovery escalations fell below the selected trust threshold. Recovery time increased with escalation margin and run length, shifted with score weights, and was unchanged by the restriction threshold. With degradation spread across 5-20 intervals, TRAP restricted only on four traces retaining a sufficiently sharp single-interval drop, and often held more authority than readiness supported. Its restriction advantage therefore depended on abrupt degradation. ### Source excerpt arXiv:2611.06996v1 Announce Type: new Abstract: The policy that converts runtime estimates of an autonomous agent's condition into operational authority is rarely specified or measured. We empirically characterize the Trust Responsive Authorization Policy (TRAP), which maps an Authorization Readiness Score (ARS), a weighted combination of runtime trust, operational stability, and contextual risk, onto discrete authorization levels through an asymmetric rule: escalation requires consecutive observations above a band edge plus a margin, while a sharp drop in readiness restricts authority at once. On one matched set of 100 Monte Carlo traces, we compare TRAP with static, trust-only, risk-only, and symmetric policies. Under abrupt degradation, TRAP restricts at onset and re-escalates sooner after recovery than the symmetric policy, with the highest and, descriptively, least variable stability among the adaptive policies, although more of its post-recovery escalations occur below the selected trust threshold. Paired bootstrap intervals exclude zero for the remaining differences. In one-factor sweeps, recovery time rises with the escalation margin and run length, shifts with the score weights, and is unchanged by the restriction threshold. In an exploratory post-freeze experiment, when the same degradation is spread over 5-20 intervals, TRAP restricts only on the 4 traces that retain a single-interval drop above that threshold; averaged over all 100 traces, it holds authority above what readiness supports for 15.9-18.9 of the 20 degradation intervals, against 0.8 under abrupt degradation and about 10 for the symmetric policy. Its restriction advantage therefore depends on degradation being abrupt. All reported values are traceable to the publicly available, versioned research release. ## AtlasULP: Domain-aware Universal Link Prediction via Relation Atlas DevFeed: [AtlasULP: Domain-aware Universal Link Prediction via Relation Atlas]() Original publisher: [Read original article]() Author: Yujing Liu, Yixin Liu, Yu Zheng, Lianhua Chi, Alan Wee-Chung Liew, Heng Tao Shen, Shirui Pan Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Representation learning](), [generalization in machine learning]() Tags: [adaptive](), [applications](), [arxiv](), [atlas](), [based](), [capture](), [challenges](), [characterization](), [connectivity](), [context](), [costs](), [cs-lg](), [dataset](), [develop](), [domain](), [domains](), [essential](), [existing](), [experiments](), [link](), [prediction]() ### AI overview AtlasULP targets universal link prediction on unseen graphs, where existing subgraph-sampling methods may miss higher-order signals and domain-agnostic predictors may generalize poorly. It uses a relation atlas to represent node-pair structural associations from global and local perspectives, then generates domain-aware prompt tokens from contextual structures for prompt-guided in-context prediction. The abstract reports that AtlasULP consistently outperforms state-of-the-art methods across diverse graph domains; it does not provide experiment details or quantify the gains. ### Source excerpt arXiv:2611.07347v1 Announce Type: new Abstract: Link prediction (LP) is a widely applied task in graph learning. Conventional LP approaches typically follow a dataset-specific paradigm, requiring independent training for each graph and incurring high computational and maintenance costs in large-scale applications. Motivated by these limitations, recent work explores Universal Link Prediction (ULP), aiming to enable training-free inference on arbitrary unseen graphs. However, existing methods primarily rely on subgraph sampling strategies to construct universal representations, which are unable to capture higher-order information due to the exponential growth of higher-order neighborhoods. As a result, they often under-perform on graphs where long-range signals are essential, such as biological networks. Moreover, their prediction models are typically domain-agnostic and lack the ability to adapt to different graph domains, leading to limited generalization. To address these challenges, we propose an Atlas-based Universal Link Prediction framework (AtlasULP), which introduces the relation atlas to capture structural associations between node pairs from both global and local perspectives, enabling a comprehensive characterization of connectivity signals across diverse graph domains. Building upon this representation, we develop a domain-guided prompting ULP model, which generates domain-aware prompt tokens from contextual structures and performs prompt-guided in-context prediction for adaptive link inference. Extensive experiments demonstrate that AtlasULP consistently outperforms state-of-the-art methods across diverse graph domains. ## When Does a Cross-Device Run Justify PASS? Obligation-Level Evidence Qualification under Partial Observability DevFeed: [When Does a Cross-Device Run Justify PASS? Obligation-Level Evidence Qualification under Partial Observability]() Original publisher: [Read original article]() Author: Zimu Liu, Xiaoxi Guo, Xiaolu Tang, Liang Yu Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Software Engineering]() Topics: [test](), [Test automation](), [audit]() Tags: [association](), [atomic](), [business](), [characterization](), [closure](), [cross](), [cs-se](), [device](), [evaluation](), [field](), [flat](), [graph](), [input](), [isolates](), [matching](), [observability](), [over](), [pass](), [policy](), [post](), [run]() ### AI overview The evidence contract qualifies support before evaluating business predicates by binding each obligation to admissible subjects, episodes, objects, and observation scopes, and recording unresolved support. In a 12-seed controlled study of 2,304 claims with identical records and policy, Atomic-Facts Flat produced 160 false PASS judgments and 76.30% PASS recall. Adding computed associations reduced false PASS to five and raised recall to 99.09%; Deterministic Closure achieved zero false PASS and 100% recall, outperforming Atomic Flat in 11 seeds with one tie. The authors attribute most benefit to explicit association rather than graph notation. A separate characterization across four workflows, 40 physical episodes, and 80 claims found 95% accuracy and 100% PASS recall (27/27), with three false PASS judgments. These field results are descriptive: a policy-input mismatch blocked the planned matched-policy comparison, and post-hoc sensitivity did not establish superiority over a policy-informed textual baseline. ### Source excerpt arXiv:2611.06013v1 Announce Type: new Abstract: A cross-device test can complete every command yet judge a business claim from evidence belonging to another device or attempt. We introduce an executable evidence contract that makes support qualification an explicit stage before predicate evaluation. It binds each obligation to admissible subjects, episodes, objects and observation scopes, and records unresolved support rather than treating an apparently successful observation as proof. We evaluate two implemented parts of the contract. A prospectively specified, 12-seed controlled study isolates association under reused identifiers across 2,304 claims. Given identical records and business policy, Atomic-Facts Flat produces 160 false PASS and 76.30% PASS recall; supplying computed associations reduces false PASS to five and increases PASS recall to 99.09%. Deterministic Closure produces zero false PASS and 100.00% PASS recall, improving on Atomic Flat in 11 seeds with one tie (two-sided sign p=0.00098). Most of the benefit comes from explicit association rather than graph notation. A separate prospective field characterization spans four workflows, 40 physical episodes and 80 claims. Against frozen reference labels, the physical-evidence verifier attains 95.00% accuracy and 100.00% PASS recall (27/27), with three false PASS judgments. The field results are descriptive: a policy-input mismatch prevents the planned matched-policy comparison, and post-hoc sensitivity does not establish superiority over a policy-informed textual baseline. Together, the studies characterize qualification on physical evidence and isolate the value of explicit association under a controlled matching contract. ## The Price of Partial Information with Switching Costs: A Complete Characterization for Feedback Graphs DevFeed: [The Price of Partial Information with Switching Costs: A Complete Characterization for Feedback Graphs]() Original publisher: [Read original article]() Author: Nicol\\\`o Cesa-Bianchi, Tommaso Cesari, Roberto Colomboni, Emmanuel Esposito, Federico Fusco, Tomer Koren Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [graph theory]() Tags: [2](), [3](), [abstract](), [actions](), [algorithm](), [algorithms](), [all](), [argument](), [arxiv](), [bandits](), [but](), [characterization](), [classes](), [complete](), [costs](), [cs-lg](), [feedback](), [graphs](), [information](), [switching](), [with]() ### AI overview The paper characterizes minimax regret with switching costs across feedback graphs. Strongly revealing graphs--where every pair of actions has an endpoint that observes both losses--achieve regret \(\widetilde\Theta(\sqrt T)\). Other observable graphs have regret \(\widetilde\Theta(\gamma^{1/3}T^{2/3})\), with \(\gamma\) the domination number; graphs outside these cases have linear regret. Tournaments lie at the transition: they achieve the square-root rate, while deleting any pair's edge forces the \(T^{2/3}\) rate. Thus, an action observing all losses is neither necessary nor sufficient for the square-root rate. For strongly revealing graphs, TournamentSteering uses each observed comparison both to estimate losses and steer play; the steering loss reduction pays for estimate variance. Distributing comparisons among observing actions keeps that variance, and the regret, polylogarithmic in the number of actions. ### Source excerpt arXiv:2606.03851v2 Announce Type: replace Abstract: We study how switching costs affect minimax regret under partial feedback. Which features of the feedback preserve the $\sqrt T$ rate, and which force the $T^{2/3}$ rate? We address this question in the framework of feedback graphs, which interpolate between full information and bandits: playing an action reveals the losses of its out-neighbors in a known directed graph. Prior work on feedback graphs with switching costs gave algorithms with regret of order $T^{2/3}$, but matching lower bounds only for restricted classes of graphs, leaving the question open. We establish a complete characterization of the minimax regret with switching costs for all feedback graphs: $\widetilde\Theta(\sqrt T)$ if the graph is strongly revealing, meaning that every pair of distinct actions has an endpoint that observes the losses of both; $\widetilde\Theta(\gamma^{1/3}T^{2/3})$ for every other observable graph, where $\gamma$ is the domination number; and linear otherwise. In particular, tournaments, in which every action observes its own loss and each pair of actions is joined by one directed edge, sit at the phase transition: they behave like full information, but deleting the edge of any pair forces the rate $T^{2/3}$. Hence an action that observes all losses is neither necessary nor sufficient for the rate $\sqrt T$. Our algorithm TournamentSteering for the strongly revealing case rests on a new principle: each observed comparison of two actions serves both to estimate losses and to steer play, and the loss reduction from steering pays for the variance of the estimates. For every strongly revealing graph, distributing the comparisons among observing actions keeps this variance, and hence the regret, polylogarithmic in the number of actions, by an argument similar to the greedy construction of logarithmic dominating sets in tournaments. ## Towards Automatic Heart Murmur Characterization from Multi-spot Phonocardiograms DevFeed: [Towards Automatic Heart Murmur Characterization from Multi-spot Phonocardiograms]() Original publisher: [Read original article]() Author: Marcelo Nogueira, Jorge H. Oliveira, Carlos G. Ferreira, Miguel T. Coimbra, Al\\'ipio M. Jorge Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Model Development](), [Neural Network](), [Representation learning](), [clustering]() Tags: [architectures](), [characterization](), [classification](), [cnn](), [cs-lg](), [cs-sd](), [dataset](), [performance](), [structured]() ### AI overview The study frames heart murmur characterization as five patient-level classification tasks covering timing, shape, pitch, quality, and grade, using recordings from four auscultation locations. On 735 CirCor DigiScope patients with complete four-spot recordings, the best independent per-location classifiers combined by maximum-severity late fusion achieved 85.56%-93.81% unweighted mean sensitivity across tasks, with relevant class recalls of 80%-100%. A synchronous multi-input CNN was 7-16 points lower by task; a unified multi-input, multi-output CNN estimated all five attributes from a shared representation, reaching 72.59%-79.40%. Adding all minority-class segments for training balance reduced patient-level performance under the late-fusion rule, which the authors attribute to non-representative noisy segments not screened by the clustering-based evaluation selection. The abstract presents multi-spot characterization as feasible, while reporting higher performance for late fusion than for the jointly learned models. ### Source excerpt arXiv:2611.03654v1 Announce Type: new Abstract: Cardiac auscultation traditionally detects the presence of a heart murmur, but clinicians further characterize it by timing, shape, pitch, quality, and grade. Most automatic phonocardiogram analysis systems stop at binary detection and do not estimate this richer description. In this work, we formulate automatic murmur characterization as five patient-level classification tasks and compare three multi-spot strategies on recordings from the aortic, pulmonary, mitral, and tricuspid auscultation locations. First, classical and deep-learning classifiers are trained independently at each location, and predictions are combined via a maximum-severity late-fusion rule. Second, a synchronous multi-input CNN jointly processes one representative segment per location for a single task. Third, a unified multi-input, multi-output CNN predicts all five attributes from one shared multi-spot representation. Using 735 patients from the CirCor DigiScope dataset with complete four-spot recordings, the best late-fusion configurations reach an unweighted mean sensitivity between 85.56% and 93.81% across the five tasks, with class-specific recalls between 80.00% and 100.00% for every relevant category. The multi-input CNN trails late fusion by 7-16 points depending on the task, while the multi-output model reaches values between 72.59% and 79.40%, estimating all five attributes from one shared representation. We further show that naively balancing training data by adding all minority-class segments degrades patient-level performance under the maximum-severity rule, since it introduces non-representative, noisy segments not filtered by the clustering-based segment selection used for evaluation. These results indicate multi-spot murmur characterization is feasible with both fusion-based and jointly learned architectures, outlining a path toward automatic, structured murmur reports complementing binary screening. ## A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing DevFeed: [A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing]() Original publisher: [Read original article]() Author: William Nixon, Jon Durbin, Florian Standhartinger, Haryadi S. Gunawi, Juncheng Yang Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing](), [arXiv: Machine Learning]() Topics: [Large Language Model](), [LLM observability](), [Load Balancing](), [Performance Testing]() Tags: [abstract](), [aggregate](), [analyze](), [arxiv](), [balancing](), [behavior](), [benchmarking](), [caching](), [capture](), [characterization](), [cloud](), [critical](), [cs-ai](), [cs-dc](), [cs-lg](), [essential](), [evolution](), [existing](), [llm](), [llm-serving](), [load](), [load-balancing](), [serving]() ### AI overview The study analyzes a one-year production trace from CompanyX to characterize LLM serving across aggregate, temporal, model-level, and user-level views. The trace covers many users and models, including popular and long-tail models, and is publicly released to support research using production workloads. The supplied excerpt does not report specific caching or load-balancing findings. ### Source excerpt arXiv:2608.13573v2 Announce Type: cross Abstract: Large Language Model (LLM) serving has become a critical cloud workload, and realistic traces are essential for motivating and benchmarking serving systems. However, existing LLM serving workload studies remain limited in scale and scope. They often observe short time periods and provide limited visibility into how users interact with models in production. As a result, they do not fully capture how LLM serving workloads evolve over time or how user-model interactions shape production traffic. In this work, we further the understanding of real-world LLM serving workloads through both a global characterization and a longitudinal study of a one-year production trace from CompanyX. Unlike prior studies, our trace captures full production behavior across many models and users, including both popular and long-tail models. We analyze the workload from aggregate, temporal, model-level, and user-level perspectives, revealing workload evolution and user-model structure that are typically hidden behind aggregate views. To support future research, we publicly release the full one-year trace, enabling downstream studies of production behavior without relying on sampled or synthetically generated workloads. The trace is available at https://github.com/HarvardMadSys/chutes_workload. ## Deep Learning Compilers: Architecture, Optimization, and Empirical Characterization DevFeed: [Deep Learning Compilers: Architecture, Optimization, and Empirical Characterization]() Original publisher: [Read original article]() Author: Jinghao Wen, Ruixuan Wang, Xun Jiao Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Programming Languages / Compilers](), [Deep learning](), [PyTorch](), [High-Performance Computing]() Tags: [analysis](), [architecture](), [backend](), [characterization](), [compilation](), [compilers](), [cs-ar](), [cs-lg](), [deep-learning](), [deployment](), [evaluation](), [experiments](), [frontend](), [fusion](), [graph](), [hardware](), [nvidia](), [optimization](), [overhead](), [performance](), [pytorch](), [research]() ### AI overview The paper analyzes deep learning compiler architecture across frontend, middle-end, and backend stages, highlighting the middle-end's role in connecting model representations to hardware-specific optimization. In PyTorch/TorchInductor experiments on an NVIDIA RTX A6000, benefits varied by workload: fusion-friendly graphs achieved up to 8.21x speedup, while other workloads improved less. Compilation overhead could take thousands of executions to amortize. Dynamic-shape compilation cut cumulative shape-transition costs by about 70%, with a modest steady-state performance penalty. ### Source excerpt arXiv:2611.04273v1 Announce Type: new Abstract: As deep learning (DL) models become increasingly complex and hardware architectures grow more heterogeneous, DL compilers play a critical role in optimizing computational efficiency and enabling scalable deployment. This paper presents a systematic analysis of DL compilers, covering their architectural evolution, compilation workflows, optimization techniques, and emerging research directions. We organize DL compiler architectures into frontend, middle-end, and backend stages, emphasizing the role of the middle-end in connecting high-level model representations with hardware-specific optimizations. We further examine representative DL compilers and recent optimization techniques across different compilation stages. Beyond architectural and technical analysis, we conduct an empirical evaluation of compiler behaviors to investigate the effectiveness of graph optimization, compilation overhead, and dynamic-shape handling. Experiments using PyTorch/TorchInductor on an NVIDIA RTX A6000 demonstrate that compiler benefits depend strongly on workload characteristics, with speedups of up to 8.21\times for fusion-friendly computation graphs, while other workloads exhibit limited improvements. Our results further reveal that compilation overhead can require thousands of executions to amortize, and dynamic-shape compilation reduces cumulative shape-transition costs by approximately 70\% at the expense of modest steady-state performance degradation. By combining systematic architectural analysis with quantitative experimental evidence, this work provides insights into the design, optimization, and practical deployment of modern DL compilers. ## Universal Hardness From Masking Channels DevFeed: [Universal Hardness From Masking Channels]() Original publisher: [Read original article]() Author: Parikshit Gopalan, Manoj Prabhakaran, Pranay Tankala Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [computational complexity]() Tags: [algorithm](), [boolean](), [channel](), [characterization](), [coin](), [complexity](), [convex](), [cs-cc](), [cs-lg](), [distribution](), [duality](), [framework](), [function](), [learning](), [machine-learning](), [masking](), [random](), [randomness]() ### AI overview The paper generalizes the hardcore lemma by treating input-dependent masking channels as a way to characterize learning hardness. A channel adds randomness to a function's outputs until they are indistinguishable from random noise to a hypothesis class; its fidelity measures how much masking is possible. Using convex duality, the authors connect the maximum achievable fidelity to the difficulty of learning non-Boolean functions under convex losses. They derive multiclass variants, describe subtleties for larger output alphabets, and construct a universal masking channel for multiple notions of hardness. They also relate masking channels to omnipredictors and pseudoentropy, giving a tight pseudoentropy characterization using the MaxEnt algorithm. ### Source excerpt arXiv:2611.05296v1 Announce Type: new Abstract: We prove a broad learning-theoretic generalization of the celebrated hardcore lemma concerning the average-case complexity of computing a Boolean function (Impagliazzo, 1995). Our starting point is the observation that the classic notion of a hardcore measure, which reweights a function's input distribution, is equivalent to an input-dependent masking channel that injects randomness into the function's output distribution until it appears fully pseudorandom. The analogue of density of the measure is the fidelity of the channel. From this viewpoint, we develop a general framework for certifying hardness that allows for non-Boolean functions $f$ and a wide range of notions of hardness besides the classic 0-1 loss, by choosing the right notion of fidelity via convex duality. We show the difficulty of learning $f$ using a class $\mathcal{C}$ to minimize a convex loss is tightly characterized by the maximum fidelity possible for a masking channel, which measures the least amount of noise the channel must add to hide the structure of $f$ and make it $\mathcal{C}$-indistinguishable from random noise. Using our framework, we derive a number of multiclass generalizations of the hardcore lemma, and reveal some unexpected subtleties in going to larger alphabets. We construct a universal masking channel that applies to many natural notions of hardness simultaneously. Such channels are dual to omnipredictors, which optimize multiple losses at once (Gopalan, Kalai, Reingold, Sharan, and Wieder, 2022). We observe a symmetry between masking channels and generalizations of the hardcore lemma using pseudoentropy. We present a tight characterization of pseudoentropy using the MaxEnt algorithm from machine learning. A long line of research has shown that learning and hardness are two sides of the same coin; our framework gives a crisp and general formulation of this principle via convex duality. ## Bilinear Hulls for Support Splitting: Characterization, Optimization, and Applications DevFeed: [Bilinear Hulls for Support Splitting: Characterization, Optimization, and Applications]() Original publisher: [Read original article]() Author: Yusuke Aikawa, Keita Ishizuka, Tomoki Moriya Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [Cryptography]() Tags: [abstract](), [algorithm](), [all](), [application](), [applications](), [arxiv](), [based](), [best](), [can](), [characterization](), [code](), [complete](), [cost](), [cs-cr](), [cs-it](), [dimension](), [factor](), [family](), [field](), [field-characteristic](), [math-it](), [optimization](), [support]() ### AI overview The study finds that changing the inner product used to define the hull in Sendrier's Support Splitting Algorithm (SSA) can reduce its dominant cost by at most a factor of the field size q. SSA remains applicable only for bilinear forms whose standard-basis matrix is a linear combination of the identity and all-ones matrices; within that family, hull dimension falls by at most one, giving a dominant cost reduction from q^h to at best q^(h−1). The reduction is achieved for every self-dual code whose length is coprime to the field characteristic. The authors report that PEP-based SPECK admits this reduction, whereas LEP-based LESS and PKP-based PERK are outside the framework. A SageMath preprocessing implementation demonstrates feasibility at SPECK's recommended length n = 252. Larger savings would require invariants beyond bilinear forms. ### Source excerpt arXiv:2611.01337v1 Announce Type: new Abstract: Sendrier's Support Splitting Algorithm (SSA) is the standard solver for the Permutation Equivalence Problem (PEP); its cost is dominated by a term that grows exponentially in the hull dimension. We study, mathematically, how much this cost can be lowered by replacing the standard inner product, on which the hull is defined, with an alternative bilinear form. Two structural results give a complete answer: only a narrow two-parameter family of bilinear forms produces a hull on which SSA can still operate, namely those whose matrix in the standard basis is a linear combination of the identity and the all-ones matrix; and even within this family the hull dimension drops by at most one (with the reduction achieved on every self-dual code whose length is coprime to the field characteristic). Together, these results give a tight upper bound: since SSA on a code with hull dimension $h$ has cost dominated by $q^h$, replacing the standard hull with $\mathrm{Hull}_M$ lowers this dominant term from $q^h$ to at best $q^{h-1}$, so no bilinear-form generalization of the inner product can reduce the SSA cost by more than a factor of $q$, the field size. Going beyond this requires invariants outside the bilinear framework. As an application, we verify that PEP-based SPECK admits the factor-$q$ cost reduction, while LEP-based LESS and PKP-based PERK lie outside our framework. A SageMath implementation of the bilinear-hull preprocessing confirms feasibility at SPECK's recommended length $n = 252$. ## Introducing Cluster Efficiency to LLM Serving: Pooling Economical Accelerators over Fast Interconnects DevFeed: [Introducing Cluster Efficiency to LLM Serving: Pooling Economical Accelerators over Fast Interconnects]() Original publisher: [Read original article]() Author: Youhe Jiang, Pavle Glu\\v{s}ac, Binhang Yuan, Eiko Yoneki Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing]() Topics: [LLM optimization](), [Distributed Systems & Parallel Computing](), [LLM Techniques](), [long-context](), [Large Language Model](), [quantization]() Tags: [accelerators](), [bandwidth](), [blackhole](), [blackwell](), [capacity](), [characterization](), [chips](), [cluster](), [coding](), [coding-agent](), [collective](), [communication](), [compute](), [context](), [cs-dc](), [efficiency](), [fast](), [interconnects](), [llm](), [llm-serving](), [serving]() ### AI overview The study finds that fast, low-latency, high-bandwidth interconnects can preserve economical accelerators' cost advantage for long-context LLM serving, despite limited per-device capacity. On Tenstorrent Galaxy, a cluster of Blackhole chips, the serving stack uses integrated communication engines, pipelined ring transfers, communication-computation overlap, and device-side decode replay to reduce collective overhead. Across three dense models, BF16 and 8-bit configurations, and 8K-256K-token contexts, Galaxy delivers 1.07-2.24x the tokens per dollar of an eight-GPU RTX PRO 6000 Blackwell server under a three-year acquisition-and-electricity cost model. Emulating TCP-class communication on the same chips cuts throughput by 23.5-67.4% and removes the cost advantage in every evaluated configuration. ### Source excerpt arXiv:2611.02975v1 Announce Type: new Abstract: Long-context large language model (LLM) serving places substantial demands on compute, memory capacity, and memory bandwidth. Economical accelerators can offer these resources at lower cost, but their limited per-device capacity requires pooling across devices, making communication overhead critical. We study cluster efficiency, measured in tokens per dollar and per joule, and show how low-latency, high-bandwidth interconnects enable economical accelerators to retain their cost advantage. Guided by a characterization of coding-agent workloads and accelerator scaling, we realize this design on the Tenstorrent Galaxy, a cluster of Blackhole chips. The serving stack combines integrated communication engines, pipelined ring transfers, communication-computation overlap, and device-side decode replay to reduce collective overhead. Across three dense models in BF16 and 8-bit configurations with 8K-256K-token contexts, the Galaxy achieves 1.07-2.24$\times$ the tokens per dollar of an eight-GPU RTX PRO 6000 Blackwell server under a three-year acquisition-and-electricity cost model. Emulating TCP-class communication on the same chips reduces throughput by 23.5-67.4% and eliminates the cost advantage in every evaluated configuration. These results demonstrate how fast interconnects translate the per-dollar resource advantages of economical accelerators into cost-efficient long-context LLM serving. ## Diagonal Equivalence in State Space Models: Characterization, Retrieval Rank, and Selective Extensions DevFeed: [Diagonal Equivalence in State Space Models: Characterization, Retrieval Rank, and Selective Extensions]() Original publisher: [Read original article]() Author: Dennis Rochau Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Mamba](), [LLM Techniques](), [CUDA Tile]() Tags: [2](), [3](), [about](), [abstract](), [approach](), [arxiv](), [attention](), [best](), [blocks](), [but](), [cases](), [changes](), [characterization](), [class](), [cs-lg](), [extensions](), [math-oc](), [models](), [rank](), [retrieval](), [space](), [stat-ml](), [state]() ### AI overview The paper characterizes when real or complex state-space models are equivalent to diagonal transitions in a fixed basis: their transition families must commute and each member must be diagonalizable over the chosen field. Under time-shared, token-local matrix maps, the resulting state-mediated kernels correspond to unnormalized scalar-mode linear attention with multiplicative masks. It distinguishes Mamba 3 MIMO's one value per local interface from SSD-MIMO's independent values for retrieval rank above one, and presents HybridMamba, which combines real and rotating state blocks and recovers Mamba 1-3 SISO configurations as exact cases. Across five large H100 bfloat16 workloads, fused kernels improve geometric-mean training time over official Mamba 1 and 3 in matched SISO cases; Mamba 2 is near parity at SSM-core scope. Memory comparisons depend on baseline and metric. Small-model recall results favor both designs on the hardest tier, but in single-seed ~500M-parameter language models, recall gains do not yield lower loss: results improve on Mamba 1/2 at 2B tokens but remain behind Mamba 3 SISO and MIMO through 10B tokens. ### Source excerpt arXiv:2611.03101v1 Announce Type: new Abstract: We classify real and complex state-space models (SSMs) with linear transitions, distinguishing fixed-basis diagonal equivalence from input-output equivalence and tracking coefficient-conditioning windows under coordinate changes. We characterize diagonal equivalence through commuting transition families whose members are diagonalizable over the chosen field. With time-shared token-local matrix maps, we identify this class's state-mediated kernels with unnormalized scalar-mode linear attention with multiplicative masks. Mamba 3's multiple-input, multiple-output (MIMO) construction lifts one value per local interface; SSD-MIMO routes independent values for local retrieval rank above one. HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style heads and Mamba 3's complex rotations and two-endpoint writes, mixing real and rotating state blocks with different decay laws. They recover Mamba 1-3 single-input, single-output (SISO) configurations as exact special cases. On five large H100 bfloat16 workloads, our fused GPU kernels improve geometric-mean training time over official Mamba 1/3 in matched SISO recoveries, with Mamba 2 near parity at SSM-core scope. Memory tradeoffs depend on the baseline and metric: saved-for-backward versus peak allocation. In small-model multi-token recall experiments, both designs outperform Mamba 1-3 on the hardest tier; higher routing ranks perform best in gated or phase-diverse configurations. In single-seed language models with about 500M parameters, both improve on Mamba 1/2 in test loss at 2B tokens, but the routing gains do not translate into lower language-model loss. Our best configurations approach Mamba 3 SISO through 10B tokens while remaining behind it and Mamba 3 MIMO in language-model loss. ## Optimal Subgroup Discovery at Every Support Threshold DevFeed: [Optimal Subgroup Discovery at Every Support Threshold]() Original publisher: [Read original article]() Author: Lincen Yang, Qi Huang, Niki van Stein, Thomas B\\"ack, Matthijs van Leeuwen, Zhong Li Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Combinatorial optimization](), [Data analysis]() Tags: [atom](), [benchmark](), [characterization](), [conditional](), [constraint](), [cs-lg](), [discovery](), [distribution](), [distributions](), [feature](), [generative](), [kl-divergence](), [learn](), [model](), [partition](), [point](), [post](), [results](), [search](), [select](), [support]() ### AI overview The paper characterizes subgroups with maximum Kullback-Leibler divergence at arbitrary support thresholds. Under a model where the feature space is partitioned into regions ("atoms") with homogeneous target distributions, each point on the deviation-support Pareto front consists of whole atoms and at most one fractional atom, reducing an uncountable search to a finite combinatorial one. The proposed method learns such a partition, then selects atoms greedily. On 19 benchmark datasets, it finds highly deviating subgroups across support thresholds and improves results from existing methods when used as post-processing. The evidence here is the abstract and does not provide further implementation or experimental details. ### Source excerpt arXiv:2611.03053v1 Announce Type: new Abstract: Subgroup discovery aims to identify regions of the feature space where a target variable deviates from its marginal distribution. A fundamental challenge in this problem is the trade-off between a subgroup's deviation and its support: small subgroups can be highly deviating but statistically meaningless, while large ones cannot deviate much by construction. Existing methods tackle this challenge by collapsing the two into a single scalar quality measure, implicitly committing to one particular trade-off and making it difficult to recover the most deviating subgroup under a user-specified support constraint. We instead provide the first theoretical and structural characterization of optimal subgroups with maximum Kullback-Leibler (KL) divergence under arbitrary support constraints. Under a partition-based generative model, where the feature space decomposes into regions of homogeneous conditional distributions of the target variable called atoms, we prove that every point on the deviation-support Pareto front is a union of whole atoms together with at most one fractional atom. This reduces an uncountable search to a finite combinatorial problem. Guided by this theory, we propose a simple two-step recipe: learn a homogeneous partition, then select atoms greedily. Across 19 benchmark datasets, it finds the most deviating subgroups at a range of support thresholds, and reliably improves subgroups returned by existing methods when applied as a post-processing step. These results empirically validate the sufficiency and necessity of our theoretical contributions. ## Surrogate Losses for Fairness-Aware Classification: Fisher Consistency and Calibration DevFeed: [Surrogate Losses for Fairness-Aware Classification: Fisher Consistency and Calibration]() Original publisher: [Read original article]() Author: Changcheng Song, Yi Yang, Xiangyu Chang, Xiaoling Lu, Yufeng Liu Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [responsible-ai](), [Machine learning](), [text-classification]() Tags: [application](), [characterization](), [classification](), [consistency](), [cs-lg](), [experiments](), [fairness](), [loss-functions](), [objectives](), [optimization](), [parameter](), [stat-ml]() ### AI overview The paper studies when surrogate-loss optimization yields decisions that are optimal under fairness-aware risk for demographic parity and equal opportunity, in both attribute-blind and attribute-aware settings. It defines fairness-aware Fisher consistency and characterizes it through derivative conditions, showing that consistency depends on how individual losses interact within the fairness-regularized objective. A calibration inequality connects excess surrogate risk to excess fairness-aware risk. Experiments examine how violations of the consistency conditions affect risk and how the regularization parameter affects whether optimization is well-posed. In a consumer credit default application, the abstract reports that appropriately calibrated fairness regularization substantially reduced demographic disparities while maintaining predictive performance. ### Source excerpt arXiv:2611.01613v1 Announce Type: cross Abstract: Fairness-aware classification seeks to balance predictive accuracy and fairness across demographic groups. Although surrogate losses facilitate the optimization of fairness-aware objectives, it remains unclear when minimizing these losses yields optimal decisions under the original fairness-aware risk. This paper investigates the statistical validity of surrogate-based fairness-aware classification under demographic parity and equal opportunity, considering both attribute-blind and attribute-aware settings. We introduce fairness-aware Fisher consistency and characterize it through conditions on the derivatives of the surrogate losses. The characterization shows that consistency depends not only on the individual loss functions but also on their interaction within the fairness-regularized objective. We further establish a calibration inequality that relates the excess fairness-aware risk to the excess surrogate risk, providing a quantitative connection between surrogate optimization and the original classification objective. Numerical experiments illustrate how deviations from the consistency conditions affect fairness-aware excess risk and how the regularization parameter influences the well-posedness of surrogate optimization. An application to consumer credit default prediction demonstrates that appropriately calibrated fairness regularization can substantially reduce demographic disparities while maintaining predictive performance. ## Serving Agentic Robot Fleets: Workload Characterization, KV-Cache Thrashing, and Admission Control DevFeed: [Serving Agentic Robot Fleets: Workload Characterization, KV-Cache Thrashing, and Admission Control]() Original publisher: [Read original article]() Author: Sida He Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing]() Topics: [vllm](), [LLM optimization](), [AI Infrastructure](), [Cache](), [Performance Testing](), [Agent Harness]() Tags: [admission](), [agentic](), [cache](), [characterization](), [control](), [cs-dc](), [cs-ro](), [kv-cache](), [robot](), [robot-fleets](), [serving]() ### AI overview The measurements show that robot-fleet serving faces different capacity limits in its vision-language brain and VLA policy. Across 80 LIBERO-PRO episodes profiled on H800 GPUs, brain contexts reached 215k tokens; the median GPT-6-Astra request took 7.78 seconds, versus a 1.0-second gap between environment steps. One VLA GPU met a 1% starvation target for one robot with a 0.25-second action deadline, and supported at least 32 robots with a 1.0-second deadline. In sustained vLLM brain replay, raising concurrency from 20 to 32 cut throughput from about 2.4 to 0.6 requests per second as repetitive prefill displaced generation. A static active-episode cap restored 2.24-2.25 requests per second and increased completed trace-program throughput from 1.30 to 5.00-5.55 episodes per minute. These results show that capacity planning and admission control depend on the workload and action deadline. ### Source excerpt arXiv:2611.01404v1 Announce Type: cross Abstract: Agentic robots combine a vision-language brain, a tool-using harness, and a vision-language-action (VLA) policy. Serving a fleet of these agents introduces two coupled resource constraints: the brain repeatedly revisits long, growing key-value (KV) cache states, while the VLA executes action chunks under tight control deadlines. We characterize 80 LIBERO-PRO episodes generated by GPT-6-Astra, RPent, and pi0.5 (1,862 brain turns and 1,582 VLA inference chunks), profile both layers on H800 graphics processing units (GPUs), and replay the resulting agent programs on a vLLM serving engine. Brain contexts reach 215k tokens; the median Astra request takes 7.78 s, while the intervening environment-step gap is only 1.0 s. Under a 0.25 s action-chunk deadline, one VLA GPU meets the 1% starvation target for one robot; under a 1.0 s deadline, its measured capacity reaches at least 32. In sustained brain replay, increasing concurrency from 20 to 32 robots reduces throughput from about 2.4 to 0.6 requests/s as repetitive prefill displaces generation. A static active-episode cap restores 2.24-2.25 requests/s and raises completed trace-program throughput from 1.30 to 5.00-5.55 episodes/min. Our measurements expose the workload-dependent shift between action deadlines and brain KV residency, and establish practical principles for capacity planning and admission in sustained agentic robot fleets. ## LLM Fine-Tuning Workloads: Characterization, Challenges, and Opportunities for Serverless AI DevFeed: [LLM Fine-Tuning Workloads: Characterization, Challenges, and Opportunities for Serverless AI]() Original publisher: [Read original article]() Author: Radu Nicolae, Daniele Lotito, Alexandru Iosup Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing]() Topics: [Fine-tuning](), [LLM optimization](), [Large Language Model](), [Serverless](), [AI Infrastructure](), [performance-engineering](), [Performance Testing]() Tags: [batch](), [characterization](), [clusters](), [cs-dc](), [dataset](), [efficiency](), [failure](), [gpu](), [llm-fine-tuning](), [moe](), [optimization](), [roce](), [scaling](), [serverless](), [throughput]() ### AI overview An analysis of 30,920 LLM fine-tuning experiments finds that feasibility and performance vary sharply with workload and deployment. Failure rates rise from 52% for batch sizes 16-64 to 95% at batch size 1,024, while throughput grows logarithmically with batch size. Scaling is near-linear within a node; for inter-node runs, RoCE raises efficiency to 0.72 versus 0.26 without it. The reported system-level RoCE optimization reaches up to 2.9x improvement, compared with 1.7x for the application-level Fast MoE optimization. The study covers four GPU types and setups from 1 to 32 GPUs, and its dataset includes both successful and failing configurations. ### Source excerpt arXiv:2610.40983v1 Announce Type: new Abstract: Serverless execution of Large Language Models (LLMs) is becoming important, reflecting their wide societal use. An increasingly present stage in LLM pipelines is fine-tuning, which specializes trained models and improves inference performance. Serverless operation of LLM fine-tuning workloads shifts operational concerns from the user to the provider. For these new workloads, providers face important operational complexities, such as performance issues, resource waste, and even service failure due to misconfiguration. However, optimizing such workloads is still at an early stage, and, relevant to this work, relatively little is currently known about how LLM fine-tuning workloads run in practice. In this work, we address the knowledge gap by characterizing the execution of LLM fine-tuning workloads on clusters equipped with different GPUs, running LLM models at various scales. We analyze IBM's LLMFineTuningBench, a large-scale, diverse profiling dataset of 30,920 fine-tuning experiments that records both successful and failing configurations. We provide quantitative insight into throughput (expressed in tokens/s), scaling, failure, and potential for performance optimization for four commonly used GPU types, in setups of 1 to 32 GPUs. We find that (1) job feasibility (run or fail) depends on workload properties, with failure ratios rising from 52% for small batch sizes (16-64) up to 95% for the largest (1,024); (2) throughput grows logarithmically with batch size; (3) scaling is near-linear when the workload is deployed intra-node, and optimizations such as RoCE substantially improves efficiency for inter-node deployment (e.g., 0.72 vs. 0.26 efficiency without RoCE); and (4) system-level optimizations can outweigh application-level ones (RoCE, up to 2.9$\times$, vs. Fast MoE, 1.7$\times$). ## Hierarchical Cooperative GFlowNets DevFeed: [Hierarchical Cooperative GFlowNets]() Original publisher: [Read original article]() Author: Pedro Dall'Antonia, Juan Belieni, Ana Carolina Erthal, Saullo H. G. de Oliveira, Eliezer de Souza da Silva, Diego Mesquita Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [model architecture](), [Combinatorial optimization]() Tags: [abstract](), [aggregation](), [approximation](), [architecture](), [asynchronous](), [characterization](), [combinatorial-optimization](), [coverage](), [cs-ai](), [cs-lg](), [discovery](), [distribution](), [experiments](), [exploration](), [flow](), [generation](), [generative](), [hierarchy](), [learning](), [mechanism](), [model](), [multimodal]() ### AI overview Hierarchical Cooperative GFlowNets (HCG) coordinates multiple samplers to reduce overlap and improve exploration of compositional search spaces. It lets the highest-ranked GFlowNet explore the full space while progressively steering lower-ranked models toward regions not covered by higher-ranked ones. A shared classifier partitions the space without coupling the models' parameters, allowing asynchronous training. The authors characterize objectives that shift probability mass away from designated regions and derive a normalizer-weighted ensemble rule that recovers the reward-proportional target distribution under stated assumptions. Experiments on multimodal distributions, sequence generation, combinatorial optimization, and biological sequence design report improved mode coverage and target approximation. ### Source excerpt arXiv:2610.33968v2 Announce Type: replace Abstract: Generative Flow Networks (GFlowNets) are trained to sample objects from compositional spaces with probability proportional to a reward function, but ineffective exploration can cause learned samplers to concentrate on a few high-reward regions. An intuitive remedy is to train multiple GFlowNets independently in parallel and combine their samples. However, this does not prevent multiple models from collapsing onto the same high-reward areas, leaving large regions of the state space unexplored. With this in mind, we propose Hierarchical Cooperative GFlowNets (HCG), an approach that dynamically partitions the search space to coordinate discovery without relying on hand-crafted boundaries. HCG imposes an ordering in which the highest-ranked model may explore the entire space, while lower-ranked models are progressively pushed toward regions not covered by higher-ranked ones. We provide a theoretical characterization of how specific learning objectives can shift probability mass away from designated areas, instantiating this mechanism within our architecture. A shared classifier coordinates the hierarchy by partitioning the search space, keeping the GFlowNets independently parameterized and enabling asynchronous training. Furthermore, we derive a normalizer-weighted aggregation rule for the ensemble and prove that it recovers the target reward-proportional distribution under the stated assumptions. Finally, experiments across multimodal distributions, sequence generation, combinatorial optimization, and biological sequence design show that HCG improves mode coverage and target approximation, translating increases in the number of coordinated samplers into broader, complementary exploration. ## From MS/MS evidence to adaptive spectroscopic characterization: Reinforcement learning guided by candidate molecular structures DevFeed: [From MS/MS evidence to adaptive spectroscopic characterization: Reinforcement learning guided by candidate molecular structures]() Original publisher: [Read original article]() Author: Bowen Gao, Lei Zhu, Yiying Wang, Wenjie Yu Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [Machine learning]() Tags: [acquisition](), [adaptive](), [characterization](), [cost](), [cs-lg](), [dataset](), [evaluation](), [groups](), [infrared](), [measurement](), [network](), [nuclear](), [performance](), [ranking](), [reinforcement-learning](), [structure](), [support](), [uncertainty]() ### AI overview The structure-informed deep Q-network (SI-DQN) uses MS/MS evidence, candidate rankings, and score-weighted chemical and structural descriptors to choose whether to request IR or NMR measurements and when to stop. On 79,462 held-out samples with simulated spectra and fixed candidate sets containing the correct structure, it raised Top-1 accuracy from 80.10% to 93.78% for positive-ion starts and from 70.44% to 93.31% for negative-ion starts, averaging 0.227 and 0.340 additional measurements per sample. Compared with acquiring all four additional measurements, it cut measurement counts by 94.3% and 91.5%, while complete acquisition reached 97.72% accuracy. Relative to a rule-based policy, SI-DQN reduced a composite loss of ranking error and acquisition cost by 19.3% and 25.5%. These results are from simulations, with equal normalized acquisition costs. ### Source excerpt arXiv:2610.40264v1 Announce Type: new Abstract: Small-molecule identification from tandem mass spectrometry (MS/MS) may require complementary measurements. Herein, we propose a structure-informed deep Q-network (SI-DQN) that uses existing MS/MS evidence to decide whether further characterization is needed, which measurement to request, and when to stop. The policy combines spectral representations and candidate rankings with score-weighted descriptors of the candidates' chemical features and structural differences. The policy was evaluated on 79,462 held-out samples with simulated MS/MS, infrared (IR), and nuclear magnetic resonance (NMR) spectra from the Multimodal Spectroscopic Dataset. Evaluation used fixed hard candidate sets containing the correct structure and equal normalized acquisition costs. Averaged across three training runs, SI-DQN increased Top-1 accuracy from 80.10% to 93.78% for positive-ion MS/MS starts and from 70.44% to 93.31% for negative-ion starts. The corresponding mean additional measurement counts were only 0.227 and 0.340 per sample, respectively. Complete acquisition achieved 97.72% accuracy with four additions, whereas SI-DQN reduced the additional measurement count by 94.3% and 91.5%. Relative to a rule-based policy, SI-DQN reduced the composite loss combining terminal ranking error and acquisition cost by 19.3% and 25.5%. Adding structural information to DQN further improved accuracy while reducing requests across all three training runs. At comparable ranking uncertainty, SI-DQN increased request rates relative to DQN more strongly in groups with high structural disagreement, where larger reductions in composite loss were also observed. Together, these results support using the ambiguity among candidate structures after MS/MS analysis to guide further characterization, providing a quantitative basis for balancing identification performance against subsequent measurement effort. [Next page]()