# d Published articles for d. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Stochastic Autoregressive Learning DevFeed: [Stochastic Autoregressive Learning]() Original publisher: [Read original article]() Author: Ilan Doron-Arad, Idan Mehalel, Elchanan Mossel Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Language models](), [Large Language Model]() Tags: [2](), [2025](), [abstract](), [analogue](), [arxiv](), [autoregressive](), [base](), [bernoulli](), [bounded](), [can](), [chain-of-thought](), [class](), [comparison](), [cs-lg](), [d](), [dependencies](), [dimension](), [distribution](), [distributions](), [e2e](), [learning]() ### AI overview The paper studies PAC learning for a generator that repeatedly samples Bernoulli next tokens from an expanding prompt over M steps. It compares base supervision, which provides one-step samples, CoT supervision, which reveals full trajectories, and end-to-end supervision, which reveals only each trajectory's final token; learning is measured under KL divergence. At the same accuracy scale, these tasks have no universal sample-complexity ordering: both CoT relative to base and end-to-end relative to CoT can exceed M/ε by arbitrarily large factors. With adjusted scales, CoT learning at ε is bounded by base learning at ε/M², while end-to-end learning is bounded up to logarithmic factors by (M/ε) times CoT sample complexity at Θ(ε). The abstract says these dependencies and scales are essentially tight and notes a study of d-dimensional logistic functions. ### Source excerpt arXiv:2608.07224v2 Announce Type: replace Abstract: Motivated by LLMs, which generate outputs by iteratively sampling from next-token distributions, we introduce a PAC-learning model for binary stochastic autoregressive learning. This generalizes the deterministic autoregressive learning framework of Joshi et al., COLT 2025. In our model, one fixed generator assigns a Bernoulli next-token distribution to every prompt string. Starting from an input prompt, a token is sampled and appended to the prompt; the same generator is then applied again to this expanded prompt; this procedure is repeated for $M$ steps. Three forms of supervision are considered: base one-step samples, chain-of-thought (CoT) samples that reveal full random trajectories of length $M$, and end-to-end (e2e) samples that reveal only the final token of length $M$ trajectories. For a generator class, we study the minimum number of samples $m_{base}(\varepsilon),m_{CoT}(\varepsilon), m_{e2e}(\varepsilon)$, resp., required to learn the one-step probabilities in the base model, and the final-token probability in the CoT and e2e models, under KL divergence. We show that stochastic autoregressive learning fundamentally differs from the deterministic theory. At scale $\varepsilon$, there is no universal comparison between the three learning tasks: both $m_{CoT}/m_{base}$ and $m_{e2e}/m_{CoT}$ can be made simultaneously arbitrarily larger than $M/\varepsilon$, the natural analogue for the existing deterministic results. Nevertheless, after altering scales, for every class, CoT learning at scale $\varepsilon$ is upper-bounded by base learning at scale $\varepsilon/M^2$, whereas e2e learning at scale $\varepsilon$ is upper-bounded, up to logarithmic factors, by $(M/\varepsilon) m_{CoT}(\Theta(\varepsilon))$. These dependencies and scales are essentially tight. We complement these bounds by studying dimension $d$ logistic functions in our model. ## Optimizing Computational-Statistical Runtime for Wasserstein Distance Estimation DevFeed: [Optimizing Computational-Statistical Runtime for Wasserstein Distance Estimation]() Original publisher: [Read original article]() Author: Peter Matthew Jacobs, Jeff M. Phillips Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [computational complexity](), [Algorithms](), [Operations research and optimization]() Tags: [2](), [3](), [abstract](), [algorithms](), [alpha](), [also](), [arxiv](), [can](), [collecting](), [compress](), [consider](), [cost](), [cs-cc](), [cs-lg](), [d](), [data](), [develop](), [distributions](), [error](), [expectation](), [optimizing](), [runtime](), [stat-ml]() ### AI overview The paper proposes a Sample-Sketch-Solve method for estimating squared Wasserstein distance from samples with additive error ε in expectation. It sketches samples on a regular Cartesian grid, which the authors say can compress data without increasing asymptotic error, especially for α-Hölder smooth distributions, while regularizing the problem for faster exact algorithms. For distributions on (0,1)^d with 0 < α < 1, the stated runtime is ε^−max(2,(d+1+o(1))/(1+α)). The authors report optimal Θ(ε^−2) runtime when d=2 and α>1/2, and nearly optimal runtime as α approaches 1 when d=3. The analysis treats sample collection as having O(1) computational cost. ### Source excerpt arXiv:2605.20122v2 Announce Type: replace-cross Abstract: Squared Wasserstein distance is a frequently used tool to measure discrepancy between probability distributions. This distance is typically computed between empirical measures of size $n$ from two underlying random samples. Unfortunately, even in lower dimensional Euclidean space problems $\left( d \in \{2,3\} \right)$, algorithms for Wasserstein distance computation with approximate or exact precision guarantees scale poorly in the runtime as a function of $n$ and the desired precision. In response, we consider the computational-statistical runtime, where the goal is to estimate from samples the Wasserstein distance between potentially smooth measures up to $\epsilon$-additive error in expectation with respect to the sampling; we allow $O(1)$ computational cost for collecting a sample. Towards this, we develop a Sample-Sketch-Solve paradigm where we introduce a regular cartesian grid sketch of the samples. We show that (especially under $\alpha$-H\"older smooth distributions) this can compress the data without increasing asymptotic error, and also regularizes the structure which enables faster exact algorithms. Ultimately, we approximate $W_2^2(P,Q)$ within $\epsilon$ error in $\epsilon^{-\max(2,\frac{d+1+o(1)}{1+\alpha})}$ time for $0 < \alpha < 1$ H\"older smooth distributions $P,Q$ on $(0,1)^{d}$; an optimal $\Theta(\epsilon^{-2})$ for $\alpha > 1/2$ when $d=2$ and nearly optimal as $\alpha \to 1$ when $d = 3$. ## On the Feature Evolution of Temporal Difference Learning DevFeed: [On the Feature Evolution of Temporal Difference Learning]() Original publisher: [Read original article]() Author: Shangzhe Li, Dechen Zhang, Difan Zou, Weitong Zhang Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Representation learning](), [thompson sampling]() Tags: [acquisition](), [algorithm](), [analysis](), [bounded](), [budget](), [commit](), [concurrent](), [continuation](), [copy](), [cs-lg](), [d](), [error](), [evolution](), [exp](), [exploration](), [family](), [feature](), [features](), [global](), [learning](), [temporal]() ### AI overview For a constructed family of smooth stochastic horizon-H MDPs with unknown quadratic structure, the paper analyzes how finite-width two-layer ReLU critics acquire and propagate predictive features through TD updates. Simultaneously updating hidden weights and the output head gives strictly lower finite-sample prediction risk than matched head-only SGD, with a 2(d+1)-fold larger population quadratic response. The analysis also finds that an imperfect learned continuation can supervise an earlier zero-reward stage; coupled updates with one SGD step per target copy yield a prediction-error transient of exp[-Ω(K/(Hd))], subject to action-independent shifts and sampling and finite-width errors. On a bounded-reward signed-action MDP family, an exploration-then-commit algorithm achieves stated regret Õ(d³H^(1/3)T^(2/3)) using Õ(Hd) global updates. ### Source excerpt arXiv:2611.04158v1 Announce Type: new Abstract: Temporal-difference (TD) learning trains neural representations using bootstrapped targets generated by evolving value estimates, yet how predictive features emerge and propagate under incomplete neural updates remains poorly understood. We investigate this problem using finite-width two-layer ReLU critics on a constructed family of smooth stochastic horizon-$H$ MDPs with unknown quadratic structure and $d$-dimensional observations. We first show that an informative TD update acquires the quadratic Bellman signal and yields a representation admitting a bounded predictive readout. The simultaneous hidden-weight and head update achieves a strict finite-sample prediction-risk improvement over matched head-only SGD, with a $2(d+1)$-fold larger population quadratic response. We then establish that an imperfect learned continuation can already provide informative supervision to an earlier zero-reward stage. By tracking coupled neural updates with one SGD step per target copy, without requiring intermediate Bellman regressions to be solved, we obtain a uniform prediction-error transient $\exp[-\Omega(K/(Hd))]$, up to action-independent shifts and explicit sampling and finite-width errors. On the constructed signed-action MDP family whose total reward is bounded by 1, we show that an exploration-then-commit algorithm can achieve $\widetilde{O}(d^3H^{1/3}T^{2/3})$ regret over $T$ episodes, using $K=\widetilde{O}(Hd)$ global updates. Our analysis connects bootstrapped feature acquisition, concurrent neural refinement, and finite-budget policy performance. ## Sweep Less, Steer More: Using Activation Steering Geometry for Reduced Hyperparameter Search DevFeed: [Sweep Less, Steer More: Using Activation Steering Geometry for Reduced Hyperparameter Search]() Original publisher: [Read original article]() Author: Nishkal Hundia, Swastik Agrawal, Navita Goyal, Sarah Wiegreffe Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Hyperparameter optimization]() Tags: [3](), [abstract](), [activation](), [activation-steering](), [analysis](), [approach](), [arxiv](), [behavior](), [behaviors](), [build](), [but](), [choice](), [compute](), [concepts](), [cs-ai](), [cs-cl](), [cs-lg](), [d](), [less](), [more](), [search](), [steering](), [using]() ### AI overview The discriminability index d′ can reduce layer-selection sweeps for difference-in-means steering: across five behaviors and three models, it reliably identifies performant layers in multiple-choice question answering and also works for layer selection in the tested open-ended, multi-token generation settings. Predicting success for an individual example from a generated token's position on the DIM line is more limited: it works on high-d′ MCQA datasets but is inconsistent across behaviors, models, and training-token choices in open-ended generation. The results caution that findings from single-token MCQA do not necessarily transfer to open-ended generation. ### Source excerpt arXiv:2611.01496v1 Announce Type: new Abstract: Representation steering methods like difference-in-means (DIM) offer a lightweight approach to controlling LLM behavior, but their effectiveness varies widely across concepts, datasets, and layers, making principled deployment costly without exhaustive hyperparameter sweeps. We build on the discriminability index $d'$, a compute-cheap geometric measure of how well contrastive training activations separate along the steering direction. Across five behaviors on 3 models, we demonstrate uses beyond its established role as a dataset-level steerability predictor. First, in the multiple-choice question answering (MCQA) setting, we show that $d'$ reliably identifies performant layers at which to steer, eliminating the need for a layer-wise sweep. Second, we extend this analysis to open-ended multi-token generation, and find that $d'$ remains successful for layer selection across different settings. Lastly, we test whether the position of a generated token's representation on the DIM line predicts steering success on that individual example. In the MCQA setting, this works for high-$d'$ datasets but is inconsistent across behaviors, models, and training-token choices in open-ended generation. This highlights that methods developed and conclusions drawn in single-token MCQA settings do not necessarily transfer to open-ended generation. ## Proximity Gaps for Gabidulin Codes and Applications DevFeed: [Proximity Gaps for Gabidulin Codes and Applications]() Original publisher: [Read original article]() Author: Songsong Li, Chaoping Xing, Chen Yuan, Ruiqi Zhu Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [Cryptography]() Tags: [abstract](), [adapting](), [all](), [alpha](), [applications](), [arxiv](), [based](), [c](), [central](), [code](), [coding](), [coding-theory](), [cryptographic](), [cryptography](), [cs-cr](), [cs-it](), [d](), [delta](), [either](), [error](), [error-correcting-codes](), [math-it](), [proximity]() ### AI overview Every linear rank-metric code over $\mathbb F_{q^m}$ has a proximity gap up to $(d-1)/(3n)$, with error at most $q^{e+1}/q^m$ for $e=\lfloor\delta n\rfloor$; Gabidulin codes improve the threshold to $(d-1)/(2n)$ with error $10q^{n-1}/q^m$. The abstract reports that this Gabidulin threshold is tight for an infinite family of constant-rate codes, and gives a matching lower bound on error up to a constant factor. Applications adapt Ligero to build an IOPP for interleaved Gabidulin codes and a commitment scheme for $q$-linearized polynomials, described as the first PCS framework based on rank-metric codes. ### Source excerpt arXiv:2609.09838v2 Announce Type: replace-cross Abstract: Proximity gaps are central to the soundness of interactive oracle proofs of proximity (IOPPs) and polynomial commitment schemes (PCSs). An $[n,k,d]$ linear code $C\subseteq\mathbb F^n$ has a $\delta$-proximity gap with error $\epsilon$ if, for every $u_0,u_1\in\mathbb F^n$, either all points on $\ell_{u_0,u_1}=\{u_0+\alpha u_1:\alpha\in\mathbb F\}$ are $\delta$-close to $C$, or at most an $\epsilon$ fraction are. Although proximity gaps for Hamming-metric codes are well understood, their rank-metric counterparts remain largely unexplored despite their applications in coding theory and cryptography. In this work, we study proximity gaps for linear rank-metric codes and their cryptographic applications. First, we show that every $[n,k,d]$ linear rank-metric code $C$ over $\mathbb F_{q^m}$ admits a proximity gap for every $\delta\le(d-1)/(3n)$, with error at most $q^{e+1}/q^m$, where $e=\lfloor\delta n\rfloor$. For Gabidulin codes, we improve the gap to $(d-1)/(2n)$ with error $10q^{n-1}/q^m$. These two proximity gaps match those for general linear Hamming-metric codes and Reed--Solomon (RS) codes, respectively. We prove the $(d-1)/(2n)$ bound is tight by constructing an infinite family of constant-rate Gabidulin codes and affine lines $\ell_{u_0,u_1}$ on which a $1-o(1)$ fraction of points are $d/(2n)$-close to the code, while $u_1$ is at least $3d/(4n)$-far from it. At $\delta=(d-1)/(2n)$, a further counterexample gives a lower bound of order $q^{n-1}/q^m$ on $\varepsilon$, matching our upper bound up to a constant factor. As applications, we construct an IOPP for interleaved Gabidulin codes by adapting the Ligero IOPP for interleaved RS codes. We then adapt the Ligero-based PCS for ordinary polynomials to obtain a $q$-linearized polynomial commitment scheme. To our knowledge, this is the first PCS framework based on rank-metric error-correcting codes. ## Tight Stochastic Sequential Covers for General VC Classes with Unknown i.i.d. Features DevFeed: [Tight Stochastic Sequential Covers for General VC Classes with Unknown i.i.d. Features]() Original publisher: [Read original article]() Author: Changlong Wu, Wojciech Szpankowski Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning]() Tags: [2](), [abstract](), [adversarial](), [also](), [approximation](), [arxiv](), [bits](), [boolean](), [class](), [classes](), [cost](), [counts](), [covers](), [cs-lg](), [cube](), [d](), [features](), [general](), [i](), [stat-ml](), [vc](), [with]() ### AI overview For binary hypothesis classes, finite VC dimension alone supports tight distribution-blind stochastic sequential covers when features are unknown and i.i.d. The paper gives an exact cover with log-size O(d log T + log(1/δ)), matching the stated worst-case horizon dependence up to constants, and derives O((d+1) log T) expected worst-case log-loss regret. Its construction weights splits by how many Boolean subcubes meet each realizable trace; a coordinate-deletion inequality bounds their total cost using VC dimension. The authors also report an exact polynomial-time implementation for planar affine halfspaces. This result does not extend to pseudo-dimension generally: a fixed pseudo-dimension-one sigmoid class needs Ω(log² T) bits for uniform sequential approximation at accuracy 1/T. ### Source excerpt arXiv:2611.00788v1 Announce Type: new Abstract: We study hybrid online learning with stochastic features and adversarial labels when the feature distribution is unknown. Prior work gave stochastic sequential covers of log-size $O(d\log^2 T)$ for a general VC class and logarithmic bounds only under stronger star-type structure. We show that finite VC dimension alone is enough: every binary class of VC dimension $d$ admits a distribution-blind exact cover with log-size $O(d\log T+\log(1/\delta))$, matching the worst-case $\Omega(d\log(T/d))$ dependence on the horizon. The key ingredient is a new local cube weight that counts Boolean subcubes incident to each realizable trace. Unlike the equal splits used in earlier constructions, we split mass unevenly according to these cube weights; a coordinate-deletion inequality shows that VC dimension controls the total cost of these splits. This yields, to our knowledge, the first tight $O((d+1)\log T)$ expected worst-case log-loss regret for general VC partitions with unknown i.i.d. features, and an exact polynomial-time implementation for planar affine halfspaces. We also show that the binary phenomenon does not extend from VC dimension to pseudo-dimension: a fixed pseudo-dimension-one sigmoid class requires $\Omega(\log^2 T)$ bits for uniform sequential approximation at accuracy $1/T$. ## Predicting Human-Gain Curves under Local Proxy Optimization from Repeated Ratings DevFeed: [Predicting Human-Gain Curves under Local Proxy Optimization from Repeated Ratings]() Original publisher: [Read original article]() Author: Moonwon Choi, Kisung Nam, Seunggeun Lee Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [ai alignment](), [human feedback](), [dpo](), [lora]() Tags: [abstract](), [alignment](), [analysis](), [arxiv](), [branch](), [bridge](), [but](), [can](), [changes](), [clusters](), [conditional](), [correlation](), [cs-ai](), [d](), [datasets](), [dpo](), [error](), [local](), [optimization](), [proxy](), [under]() ### AI overview PACE uses repeated human ratings to estimate how human-perceived quality may change as a fixed proxy is optimized gently. It models the local gain at KL radius d as A√d + Bd, with A capturing first-order proxy-human alignment and B its leading second-order correction. The authors report lower held-out gain-prediction error than A-only, zero-gain, and reliability-only baselines across several evaluation datasets and a GPT-4o judge proxy experiment. PACE is competitive with flexible proxy-conditional calibration given the same pre-pool information; a LoRA-DPO experiment offers a modest link to actual optimization. Source coefficients help transfer across datasets, while target calibration improves predictions. The diagnostic is limited to local, one-sided optimization along the positive tilt branch of a fixed item pool. ### Source excerpt arXiv:2610.41114v1 Announce Type: new Abstract: Optimizing a proxy objective can improve human-perceived quality, but the resulting gain can vary with optimization strength. Average proxy-human correlation alone does not reveal how human-perceived quality changes as optimization strength increases. We propose PACE (Proxy Alignment and Curvature Expansion), a pre-optimization diagnostic that uses repeated human ratings to predict the human-gain curve induced by a fixed proxy. The method models proxy optimization as a KL-regularized exponential tilt. At KL radius $d$, the local gain is approximated by $A\sqrt{d}+Bd$, where $A$ measures first-order proxy-human alignment and $B$ captures the leading second-order correction along the proxy-optimization path. We estimate $A$ and $B$ from repeated human ratings on calibration items and test the resulting gain-curve prediction on held-out evaluation items. Across SummEval, analysis-eligible WMT MQM clusters, an OpenMEVA-MAGS ROC/WP repeated-rating slice, and a GPT-4o judge proxy experiment, PACE consistently reduces held-out gain-prediction error relative to $A$-only, zero-gain, and reliability-only baselines. With the same pre-pool information, it is competitive with flexible proxy-conditional calibration, and a LoRA-DPO experiment provides a modest bridge to actual optimization. Transfer experiments across datasets and evaluation slices show that source coefficients are useful priors, but target calibration further improves prediction. The resulting diagnostic is local and one-sided: it characterizes gentle optimization along the positive tilt branch of a fixed item pool. ## When is Fine-Tuning Scalable? Certified Recovery from Noisy Data and Linear-Cost Fine-Tuning on Structured Task Shifts DevFeed: [When is Fine-Tuning Scalable? Certified Recovery from Noisy Data and Linear-Cost Fine-Tuning on Structured Task Shifts]() Original publisher: [Read original article]() Author: Anastasis Kratsios, Tin Sum Cheng, Bum Jun Kim, Daniel Roy Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Fine-tuning](), [Machine learning]() Tags: [abstract](), [alignment](), [all](), [alpha](), [approximation](), [arxiv](), [attention](), [based](), [c](), [continuous](), [convex](), [cost](), [cs-lg](), [cs-na](), [cs-ne](), [d](), [data](), [distribution](), [fine](), [fine-tuning](), [linear](), [math-na](), [math-pr](), [recovery](), [stat-ml](), [structured]() ### AI overview The abstract presents a model-specific two-stage procedure for a single-query, single-head transformer that certifies recovery from noisy observations without assuming training reaches an empirical risk minimizer. With high probability, it recovers the source map to arbitrary accuracy in fixed finite L^p norms; for Lipschitz targets, the rate is minimax-optimal up to logarithmic factors. For scalar Pfaffian source maps, the values matrix has a factor with at most O(ε⁻¹) scaling. If the source and shifted task share identified geometric structure, adaptation requires fine-tuning only the A values factor, a convex problem with at most O(ε⁻¹) parameters. This contrasts with the stated worst-case O(ε⁻ᵈ) source-training scaling. The claims are limited to the setting and conditions described in the abstract. ### Source excerpt arXiv:2509.00924v2 Announce Type: replace-cross Abstract: Transformers often transfer across related tasks with little fine-tuning; we study this question in noisy nonparametric regression for continuous maps $f,g:[0,1]^d\to\mathbb{R}^D$, representing the source and shifted tasks, using a single-query, single-head transformer $\mathcal{T}(\cdot)=V\operatorname{Softmax}(-\alpha(\cdot,K))$ with $\ell^{\infty}$-based alignment scores $\alpha$. Rather than assume that gradient-based training reaches an empirical risk minimizer, we give a model-specific two-stage procedure \textbf{certifying} source-task recovery and out-of-distribution (OOD) adaptation. With high probability, the resulting attention head recovers $f$ to arbitrary accuracy $\varepsilon>0$ in every fixed $L^p([0,1]^d)$ norm, $1\le p<\infty$, from noisy observations. For Lipschitz targets, its rate is minimax-optimal up to logarithmic factors without assuming oracle access to an asymptotic ERM. If $D=1$ and $f$ is Pfaffian than the values matrix factors into $AC$ with $A$ scaling at-most linearly $\mathcal{O}(\varepsilon^{-1})$ in the reciprocal approximation error $\varepsilon>0$. Importantly, our training procedure encodes all our identified (potentially) shared geometric structure from $f$ only into the values matrix factor $C$ and the keys matrix $K$ allowing for rapid access, and all $f$-specific structure only into $A$. Our main result concerns the OOD phase; where we identify geometric structures, which if shared by the source $f$ and OOD $g$ task; then only only the values matrix factor $A$ needs to be fine-tuned to guarantee an $L^p$-recovery of $g$. Critically, fine-tuning only $A$ is a convex problem which involving an \textit{at-most linear} number of parameters $\mathcal{O}(\varepsilon^{-1})$; dramatically cutting down on the worst-case scaling of $\mathcal{O}(\varepsilon^{-\mathbf{d}})$ unavoidable in the source-training phase. ## Hidden-State Privacy Has an Empty Middle DevFeed: [Hidden-State Privacy Has an Empty Middle]() Original publisher: [Read original article]() Author: Alexander Okezue Bell Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Security & Privacy](), [LLM security](), [Language models]() Tags: [2](), [3](), [abstract](), [adaptive](), [architecture](), [arxiv](), [budget](), [but](), [class](), [co-design](), [cs-ai](), [cs-lg](), [d](), [design](), [edge](), [f](), [first](), [fixed](), [full](), [hidden](), [privacy](), [state]() ### AI overview Among 1,536 tested Gaussian covariances for single-layer hidden-state release, none achieved both moderate utility and privacy against an adaptive retrieval attacker. A Fisher-ball lower bound further shows that every full-rank Gaussian release with O(1) Fisher utility has a direction whose Mahalanobis signal grows linearly with hidden width. The diagonal inverse-Fisher mechanism is uniquely minimax-optimal among diagonal releases at the stated first-order KL budget and alone kept worst-attacker top-1 at or below 0.001 across a 32 model-layer grid, but it lies at a privacy/utility edge. A generalized-eigen mechanism's 13x Pareto reduction under Euclidean retrieval vanished under adaptive Mahalanobis retrieval, where it reached 100% top-1. The abstract also reports that a split-memory transformer trained from scratch outperformed same-budget GPT baselines on its stated privacy measure, with a language-modeling loss penalty; it frames architecture or release co-design as a direction beyond Gaussian mechanism design. These are abstract-level claims; the supplied text notes that v4 was withdrawn. ### Source excerpt arXiv:2605.24042v4 Announce Type: replace-cross Abstract: Of $1{,}536$ Gaussian release covariances we tested for single-layer hidden-state privacy, zero achieve both moderate utility and moderate privacy against an adaptive retrieval attacker. We prove a complementary Fisher-ball lower bound: every full-rank Gaussian release at $O(1)$ Fisher utility admits a direction whose Mahalanobis signal grows linearly in hidden width, ruling out uniform Gaussian safety in the class and matching the empirical empty middle. The diagonal inverse-Fisher release $\Sigma^\star_{\mathrm{diag}}(\mathcal{K}) = (2\mathcal{K}/d)\,\mathrm{diag}(1/F_{ii})$ is the unique minimax-optimal diagonal mechanism at first-order KL budget $\mathcal{K}$ and the only release with worst-attacker top-1 $\le 0.001$ at every point of a 32 model-layer grid, but it sits on a privacy/utility edge rather than filling the middle. A generalized-eigen mechanism reaching $13\times$ Pareto reduction under Euclidean retrieval collapses to $100\%$ top-1 under the adaptive Mahalanobis attacker, and a full-trajectory sequence inverter recovers $94\%$ of clean GPT-2 prefixes but $0\%$ under $\Sigma_{\mathrm{diag}}$. A split-memory transformer trained from scratch reaches $G_{\mathrm{Mah}} \in [20, 33]$ at 90M and maintains a $6$--$24\times$ advantage over same-budget GPT baselines from 30M to 1B at a fixed-token language-modeling loss penalty; pretrained models top out at 9.3. These results reframe hidden-state release from mechanism-design within the Gaussian class to architecture or release co-design. ## Boundary Geometry Determines Optimal Shallow ReLU Approximation under Weighted Coefficient Constraints DevFeed: [Boundary Geometry Determines Optimal Shallow ReLU Approximation under Weighted Coefficient Constraints]() Original publisher: [Read original article]() Author: Jia-Qi Yang, Lei Shi Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Neural Network](), [Artificial Intelligence]() Tags: [2](), [abstract](), [all](), [also](), [approximation](), [arxiv](), [c](), [can](), [change](), [classes](), [convex](), [cs-lg](), [d](), [dimension](), [domain](), [error](), [fixed](), [growth](), [lower-bounds](), [math-fa](), [math-mg](), [relu](), [stat-ml](), [under]() ### AI overview For weighted variation classes on convex bodies, the optimal polynomial rate of shallow ReLU approximation can depend on boundary geometry, even with dimension, error norm, and weighting fixed. The work gives upper and lower bounds for arbitrary convex bodies and identifies optimal exponents for planar bodies and their products with polytopes. For C² bodies, it establishes exponents over a dimension-dependent range of Lq norms; with positive principal curvatures, this extends to every finite q >= 2. The domain dependence is captured by a growth exponent for small boundary caps. The authors also give sharp weight thresholds for preserving the classical rate, improve weights on Euclidean balls while retaining the classical L² rate, and show that finite-network targets can be compressed using only their original neurons without increasing the weighted coefficient sum. ### Source excerpt arXiv:2610.36854v1 Announce Type: cross Abstract: We study how boundary geometry affects $L^q$ approximation by shallow ReLU neural networks on convex bodies $K\subset\mathbb R^d$. The target classes are unit balls of weighted variation spaces that can enlarge classical spaces by incorporating domain geometry into constraints on output coefficients. We establish upper and lower bounds for arbitrary convex bodies and determine the optimal polynomial approximation exponents for all planar convex bodies and their products with polytopes, including cubes and circular cylinders. For $C^2$ convex bodies, we determine the optimal exponents in a dimension-dependent range of $q$ extending beyond $2$. When all principal curvatures are positive, the result extends to every finite $q\ge2$. For these classes, the dependence of the optimal exponent on the domain is captured by a single geometric growth exponent associated with small boundary caps. We also identify sharp weight thresholds for retaining the classical approximation exponent and improve previous weights on Euclidean balls while preserving the classical $L^2$ approximation rate. When the target is given as a finite ReLU network, the upper bounds can be realized by a compressed network using only the original neurons, without increasing the weighted coefficient sum. These results show that boundary geometry can change the optimal exponent even when dimension, error norm, and weighting rule are fixed. ## What Should an Information Bottleneck Encoder Learn? Separating Representation Learning from Rate Control DevFeed: [What Should an Information Bottleneck Encoder Learn? Separating Representation Learning from Rate Control]() Original publisher: [Read original article]() Author: Hassan Tavakoli, Bella Bose Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Representation learning](), [Machine learning](), [Neural Network]() Tags: [2](), [about](), [abstract](), [alignment](), [arxiv](), [beta](), [budget](), [certificate](), [cifar-10](), [clock](), [cnn](), [compression](), [conditional](), [control](), [cross](), [cs-it](), [cs-lg](), [d](), [data](), [information](), [learn](), [learning](), [math-it]() ### AI overview The paper separates information-bottleneck representation learning from rate control: for jointly Gaussian data, the informative directions remain fixed across operating points while scalar gains vary. Beyond Gaussian models, it relates the separation gap to a representation's sufficiency deficit; for finite inputs there is a worst-case lower bound, and for finite label alphabets a uniform-in-budget upper bound. The certificate also extends to changed target distributions and pooled domains, with an added conditional-alignment term. In experiments on MNIST, SVHN, CIFAR-10 and STL-10, a frozen CNN with a separate compressor-decoder pair for each budget used 7.7x fewer estimated training MACs and about half the wall-clock time of end-to-end VIBs, with higher budget-averaged accuracy. Its advantage was concentrated at strong compression, where VIBs collapsed. At weak compression, VIB had higher cross-entropy and calibration error; surrogate Lagrangian comparisons varied by domain and evaluation set. ### Source excerpt arXiv:2610.36255v1 Announce Type: new Abstract: The information bottleneck (IB) trades compression of an input $X$ against prediction of a target $Y$ through a multiplier $\beta=1/\lambda$, yet each operating point is usually served by training, or fully conditioning, a new encoder. For jointly Gaussian data, we show that the IB solution factorizes as $A(\beta)=D(\beta)W$: the information-bearing directions $W$ do not depend on $\beta$, and $\beta$ acts only through scalar gains. We ask whether this separation of \emph{representation} from \emph{rate control} survives beyond Gaussian models, treating the budget as a given operating parameter rather than a quantity to be optimized. For a budget-free representation \(F(X)\), the \emph{separation gap} is controlled by the \emph{sufficiency deficit} $\epsilon_F=I(X;Y\mid F(X))$: We show there is at least $\epsilon_F$ in the worst case for finite inputs, and it is bounded uniformly in $\lambda$ by a function of $\epsilon_F$ for finite label alphabets. The same certificate extends to a new target distribution and to pooled multi-domain data, where a conditional-alignment term appears. On MNIST, SVHN, CIFAR-10 and STL-10 (10 seeds, eight prespecified budgets), a frozen CNN representation with one compressor--decoder pair per budget reduces estimated training multiply--accumulate operations (MACs) by $7.7\times$ and wall-clock time by about $2\times$ relative to end-to-end VIBs, and attains higher budget-averaged accuracy ($65.6\%$ versus $57.3\%$ per domain; $67.1\%$ versus $57.1\%$ pooled), with the advantage concentrated at strong compression, where the VIBs collapse. At weak compression ($\lambda\le0.036$), VIB has higher cross-entropy and ECE; its surrogate Lagrangian is higher per domain, indistinguishable from the pooled VIB in the macro average, and lower on the pooled test set. ## Combinatorial Characterizations of the Learnability of Conditional Distributions DevFeed: [Combinatorial Characterizations of the Learnability of Conditional Distributions]() Original publisher: [Read original article]() Author: Mihir Dhanakshirur, Ambuj Tewari Published: 2026-10-26T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Combinatorial optimization]() Tags: [abstract](), [also](), [arxiv](), [characterization](), [class](), [classes](), [conditional](), [cs-lg](), [d](), [dimension](), [dimensions](), [distribution](), [distributions](), [function](), [gamma](), [graph](), [i](), [input](), [leading](), [learner](), [learning](), [stat-ml]() ### AI overview The paper studies PAC learnability of classes mapping inputs to distributions over a finite label space, with performance measured by excess risk under bounded proper losses satisfying two assumptions. It introduces the fat_γ Natarajan dimension and shows that finiteness at every positive scale is necessary and sufficient for learnability. A fat_γ Graph dimension and inequalities relating the two dimensions yield an equivalent characterization. The abstract also states that a distribution-valued class is learnable exactly when each coordinate class is learnable in the probabilistic concept model. ### Source excerpt arXiv:2610.31890v1 Announce Type: cross Abstract: We study the PAC-learnability of function classes that map inputs to probability distributions over a finite space. The learner observes i.i.d. input-label pairs, with each label sampled from the conditional distribution of an unknown target hypothesis in the class. We measure learning through excess risk under a bounded proper loss that satisfies two mild assumptions. We introduce the \textbf{fat$_\gamma$Natarajan dimension}, which generalizes both fat$_\gamma$ and Natarajan dimension, and prove that its finiteness at every positive scale is necessary and sufficient for learnability. We also introduce a \textbf{fat$_\gamma$Graph} dimension extending the classical \gra dimension and establish inequalities comparing the two dimensions at different scales, leading to an equivalent characterization of learnability. Finally, we show that a distribution-valued function class is learnable if and only if each of its coordinate classes is learnable in the probabilistic concept model. ## Generalization properties of the token perceptron: the role of embedding and vocabulary DevFeed: [Generalization properties of the token perceptron: the role of embedding and vocabulary]() Original publisher: [Read original article]() Author: Dario Bocchi, Matteo Negri, Flavio Nicoletti, Francesco D'Amico Published: 2026-10-26T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Language models](), [Embeddings](), [Large Language Model](), [mlp]() Tags: [2](), [abstract](), [alpha](), [arxiv](), [call](), [capacity](), [classes](), [classification](), [closure](), [compute](), [cond-mat-dis-nn](), [cond-mat-stat-mech](), [consider](), [continuous](), [cross](), [cs-lg](), [d](), [dimension](), [embedded](), [embedding](), [perceptron](), [properties](), [stat-ml](), [token]() ### AI overview The token perceptron model shows that embedding geometry changes token prediction from regression to inequality-constrained classification. With continuous spherical embeddings, normalized labels fix the teacher output direction; for dimension d>1, the teacher is recovered exactly above α_rec = d²/(d−1), and spherical cross-entropy reaches that threshold at large inverse temperature. Random-association storage capacity scales as d² for margins of order 1/√d. With finite-vocabulary embeddings, each label identifies a Voronoi cell and imposes inequalities. An approximate spectral closure for Bayes-optimal error indicates that learning first resolves directions with the vocabulary's largest variance, even for an isotropic teacher. ### Source excerpt arXiv:2610.32788v1 Announce Type: cross Abstract: Large language models are trained to predict tokens from a finite vocabulary, each represented by a vector in a $d$-dimensional embedding space. Therefore, token prediction is a classification problem whose classes have a geometry determined by the embedding. We study how this geometry affects learning in a minimal solvable model: a linear predictor that maps a sequence of $N$ embedded tokens to a $d$-dimensional output through one $d\times d$ matrix per position, trained in the teacher--student setting. We call this the \textit{token perceptron}. We compute its generalization error with the replica method, taking $N\to\infty$ at fixed number of examples per position $\alpha=P/N$ and keeping $d$ and the vocabulary size $D$ finite. As a first reference case, we consider continuous embeddings on the sphere, which isolate the role of the embedding dimension. The natural load is then $\alpha/d$, and for $d>1$ a normalized label fixes the direction of the teacher output, so the problem becomes a noiseless regression: the teacher is recovered exactly above $\alpha_{\rm rec}=d^2/(d-1)$, and spherical cross-entropy reaches this threshold at large inverse temperature. In the same setting the storage capacity of random associations scales as $d^2$ for margins $O(1/\sqrt{d})$. We then turn to vocabulary embeddings, where a label only identifies the Voronoi cell of the teacher output, so that each example imposes inequalities rather than equations, as in the classical perceptron. We compute the Bayes-optimal error for a given embedding matrix through an approximate spectral closure and show that the optimal learner first resolves the directions in which the vocabulary has the largest variance, even when the teacher is isotropic. ## Autoregressive Physics-Informed Neural Operator Surrogates and Joint 4-D Seismic-Production Bayesian History Matching DevFeed: [Autoregressive Physics-Informed Neural Operator Surrogates and Joint 4-D Seismic-Production Bayesian History Matching]() Original publisher: [Read original article]() Author: Clement Etienam, David Blom, Hari Pattathil, Tsubasa Onishi, Harpreet Sethi, Lalson Vincent, Suraj Pawar, Kishan Ramesh Kumar, Pandu Devarakota, Faruk O. Alpak Published: 2026-10-26T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Neural Network](), [Artificial Intelligence]() Tags: [2](), [abstract](), [accelerated](), [adaptive](), [alpha](), [arxiv](), [autoregressive](), [backpropagation](), [bayesian](), [bias](), [black](), [but](), [cs-lg](), [d](), [history](), [matching](), [neural](), [operator](), [physics](), [physics-informed](), [production](), [stat-ml]() ### AI overview The framework makes Bayesian history matching for a 100,188-cell synthetic model feasible within one GPU-hour by replacing the forward simulator with a fast autoregressive Fourier neural operator surrogate. It trains the rollout with a black-oil residual physics penalty, truncated backpropagation, and noise injection, and derives 4-D seismic ΔVp through a closed-form Gassmann transform. Assimilation proceeds in two stages: aREKI fits production data, then ES-MDA assimilates production and seismic data jointly. To address rank deficiency and unconstrained directions in the high-dimensional inversion, permeability and porosity are represented in a lower-dimensional chart using either a truncated cosine transform or a variational auto-encoder decoder. On held-out data, the surrogate reports R² above 0.999 for pressure and saturation and a 0.579-second rollout, about 2,500 times faster than the forward model. The abstract also states theoretical bounds on rollout error, contraction under a small residual defect, and truncation bias; results are demonstrated on a synthetic model. ### Source excerpt arXiv:2610.31677v1 Announce Type: cross Abstract: We present a surrogate-accelerated Bayesian history-matching framework, demonstrated on a realistic synthetic model ($N=100{,}188$ cells). The forward map $\mathcal{G}$ is replaced by a Fourier neural operator $\mathcal{G}_\theta$ trained in a closed loop, $\hat{x}_{n+1}=\mathcal{G}_\theta(u,\hat{x}_n)$, with the discrete black-oil residual $R(x_{n+1},x_n)$ as a physics-informed neural operator (PINO) penalty, truncated backpropagation and noise injection. We assimilate production data and 4-D seismic using adaptive regularised ensemble Kalman inversion (aREKI), and synthesise $\Delta V_p$ inside the rollout via a closed-form Gassmann transform of the surrogate, under a two-stage schedule (aREKI on production, then an ensemble smoother with multiple data assimilation, ES-MDA, on both). To temper the ill-posedness, we invert $(\log K,\phi)$ not in $\mathbb{R}^{2N}$, where $N_e\ll 2N$ leaves a rank-deficient covariance and a null space the data never inform, but on a chart $u=\Psi(\tilde m)$, $\tilde m\in\mathbb{R}^{d}$, $d\ll 2N$, given by a Parseval-exact truncated cosine transform or a variational auto-encoder decoder. The ensemble then spans its own search space; the directions the data cannot constrain are cut to those the chart can express, and every member decodes to a geologically plausible field. For the rollout we prove an exposure bound $W_1\le\sum_j L^{\,n-1-j}\mathbb{E}e_j$, a physics-penalty contraction $\|D_x\mathcal{G}_\theta\|\le\alpha<1$ under a small residual defect, and an $\mathcal{O}(\rho^K)$ truncation bias. A Loewner-order inequality $\mathcal{I}_p\preceq\mathcal{I}_p+\mathcal{I}_s$ explains why joint assimilation is no less informative. On held-out data, $\mathcal{G}_\theta$ attains $R^2>0.999$ on pressure and saturation and rolls out in $0.579$ s, a $\sim2500\times$ speed-up, bringing an $N_e=190$-member, $N_a=10$-iteration joint inversion within one GPU-hour. ## Transient Amplification in Cascaded Gradient Methods: Exact Minimax Laws and Tracking without Resonance DevFeed: [Transient Amplification in Cascaded Gradient Methods: Exact Minimax Laws and Tracking without Resonance]() Original publisher: [Read original article]() Author: Ahanaf Hasan Ariq, Davide Wiest Published: 2026-10-26T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [catboost]() Tags: [2](), [3](), [abstract](), [accelerated](), [arxiv](), [block](), [blocks](), [c](), [can](), [code](), [coupling](), [cs-lg](), [cubic](), [d](), [dimension](), [effective](), [eigenvalues](), [equals](), [gradient](), [math-oc](), [minimax](), [stat-ml](), [tracking]() ### AI overview Cascaded leader-follower gradient methods can have large transient amplification even when asymptotically stable, so eigenvalues alone do not predict finite-time gain. For the stated block-matrix assumptions, the excerpt gives an exact dimension-free worst-case amplification at each horizon, attained by an aligned two-dimensional quadratic; the largest transient scales on the order of 1 + c/(1−γ). With martingale noise, the aligned system's stationary follower variance has cubic inverse-gap scaling. For quadratic followers, coupling is related to sensitivity R; with an isotropic leader and positive semidefinite follower map, amplification is bounded by 1 + ||R|| independently of spectral gaps. A look-ahead Newton update removes follower memory and is minimax-optimal among linear updates that track the leader exactly. Experiments report that accelerated or inexact follower solvers can exceed the sensitivity bound when their effective maps have negative eigenvalues. ### Source excerpt arXiv:2606.04031v3 Announce Type: replace Abstract: Leader-follower gradient iterations, including bilevel and two-time-scale methods, can exhibit large finite-time excursions despite asymptotic stability. For cascaded iterations with block matrix J = [[A,0],[C,D]], symmetric diagonal blocks of norm at most gamma < 1, and coupling norm at most c, we prove an exact, dimension-free minimax formula: at every horizon t, the worst-case amplification equals the spectral norm of [[gamma^t,0],[c t gamma^(t-1),gamma^t]], attained by an aligned two-dimensional quadratic. Consequently, the worst transient is on the order of 1 + c/(1-gamma), and eigenvalues alone cannot predict finite-time gain. Under martingale noise, the aligned system has stationary follower variance sigma^2 c^2 (1+gamma^2)/(1-gamma^2)^3, establishing cubic inverse-gap scaling. For quadratic followers, the coupling is linked to follower sensitivity R by C = (I-D)R. When the leader is isotropic and the follower map is positive semidefinite, amplification is at most 1 plus the operator norm of R, independently of the spectral gaps. A look-ahead Newton update eliminates follower memory and is minimax-optimal among linear updates that track the leader exactly. Experiments validate the bounds and show that accelerated or inexact follower solvers can exceed the sensitivity bound when their effective maps have negative eigenvalues. Code and experiment scripts: https://github.com/Axym-Labs/transient-amplification ## SABLE: Scalable Affine Bayesian Latent-space Exploration DevFeed: [SABLE: Scalable Affine Bayesian Latent-space Exploration]() Original publisher: [Read original article]() Author: Erik Bodin, Lisa Maria Kreusser, Donney Fan, Henry Moss, Geoff Pleiss, Nicolas Durrande Published: 2026-10-23T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Artificial Intelligence](), [Model Development]() Tags: [3](), [abstract](), [acquisition](), [arxiv](), [bayesian](), [benchmarks](), [budget](), [but](), [can](), [complex](), [component](), [continuous](), [cost](), [cs-ai](), [cs-lg](), [d](), [design](), [designs](), [dimension](), [dimensions](), [exploration](), [space](), [stat-ml]() ### AI overview SABLE targets Bayesian optimisation when a generative model's latent dimension greatly exceeds the evaluation budget. It uses a Bayesian affine surrogate and decomposes each query into an in-span component optimised with expected improvement and an off-span component controlled by a scalar norm. The authors report that affine EI's implicit choice of this norm can underexplore; SABLE instead sets it with a finite-horizon schedule without confining search to a learned subspace. Recompression retains O(M) pseudo-observations, with refitting cost O(M²D + M³); the abstract reports inference, prediction and acquisition decisions taking tens of milliseconds. On synthetic benchmarks and latent-space tasks, SABLE is reported to achieve high sample efficiency and reduce decision time by several orders of magnitude at large budgets. ### Source excerpt arXiv:2610.31495v1 Announce Type: new Abstract: Complex objects such as engineering designs and proteins are often difficult to optimise because their design spaces exhibit high-dimensional structure and objective evaluations are expensive. Generative models provide continuous latent representations, but their latent spaces can have thousands to tens of thousands of dimensions, creating a regime where the latent dimension $D$ greatly exceeds the evaluation budget. We introduce SABLE, a Bayesian optimisation method for high-dimensional latent spaces with an isotropic distribution and limited budget. SABLE uses a Bayesian affine surrogate for prediction and acquisition, treating higher-order structure as surrogate discrepancy. After $N$ observations, the slope update has dimension at most $N-1$, yielding a decomposition of each query into an in-span component optimised by expected improvement (EI) in at most $N$ dimensions and an off-span component governed by a scalar norm. We show that the off-span norm implicitly selected by affine EI can lead to underexploration, and SABLE instead prescribes it using a finite-horizon schedule, without restricting search to a learned subspace. A recompression scheme reduces information to $\mathcal O(M)$ pseudo-observations, bounding refitting cost by $\mathcal O(M^2D+M^3)$ while keeping parameter inference, prediction, and acquisition decisions to tens of milliseconds. Across synthetic benchmarks and latent-space optimisation tasks, SABLE achieves high sample efficiency and reduces decision time by several orders of magnitude at large budgets. ## SECURE-AI: Multi-Signal Uncertainty Quantification for Trustworthy Acoustic Cough Detection DevFeed: [SECURE-AI: Multi-Signal Uncertainty Quantification for Trustworthy Acoustic Cough Detection]() Original publisher: [Read original article]() Author: Sejal Solkhan, Feiya Xiang, Mayur Sanap, Geet Khatri, Edgar Lobaton Published: 2026-10-21T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Plover](), [Embeddings](), [benchmark overfitting machine learning](), [Raspberry Pi RP2350]() Tags: [about](), [abstract](), [ai](), [all](), [arxiv](), [audio](), [benchmarks](), [but](), [can](), [cohen](), [cs-ai](), [cs-lg](), [cs-sd](), [d](), [data](), [deployment](), [detection](), [discrimination](), [eess-as](), [embedding](), [secure-ai](), [signal](), [uncertainty](), [uncertainty-quantification]() ### AI overview SECURE-AI wraps a frozen MobileNetV3 cough classifier with Monte Carlo Dropout, Temperature Scaling, and embedding-space Mahalanobis distance to estimate when predictions may be unreliable. On a held-out split of 12,161 Coswara recordings, Monte Carlo Dropout assigned lower uncertainty to errors than correct predictions, while Temperature Scaling improved calibration but not discrimination. Mahalanobis distance was the strongest signal: it separated true from false cough detections (AUROC 0.86) and false negatives from true negatives among predicted non-coughs (AUROC 0.82). At a descriptive threshold selected using labeled test data, keeping 90% of true-positive detections removed about two-thirds of false positives and raised precision from 59.9% to roughly 80%. However, a threshold tuned on validation did not transfer: F1 fell from 82% to 34% on test, so the authors emphasize threshold-independent AUROC. Most false positives were related to speech. The pipeline runs on Raspberry Pi hardware costing $30-$90. ### Source excerpt arXiv:2610.23937v1 Announce Type: new Abstract: Acoustic models for cough detection achieve high accuracy on curated benchmarks but fail silently in deployment, where speech, environmental noise, and audio unlike the training data induce confident misclassifications. In health monitoring, a confidently wrong prediction is worse than no answer at all. We present SECURE-AI (Self-aware Estimation of Cough Uncertainty for Reliable Evaluation of AI), a multi-signal uncertainty quantification framework that wraps a frozen MobileNetV3 cough classifier and estimates when its predictions should not be trusted, combining Monte Carlo Dropout, Temperature Scaling, and Mahalanobis distance in the embedding space. On a held-out split of 12,161 Coswara recordings, the probability-space methods are individually insufficient: Monte Carlo Dropout assigns lower uncertainty to incorrect than to correct predictions, and temperature scaling improves calibration but not discrimination. The strongest signal is embedding-space distance: Mahalanobis distance separates true positives from false positives (means 44.7 versus 63.4; Cohen's $d = 0.94$; AUROC 0.86) and, among predicted non-coughs, false negatives from true negatives (AUROC 0.82), so the same signal can flag both false alarms and missed coughs. At a descriptive operating point selected from labeled test data, retaining 90 percent of true-positive detections removes about two-thirds of false positives and raises precision from 59.9 to roughly 80 percent. A rejection threshold tuned on one split does not transfer to another (validation F1 of 82 percent falls to 34 percent on test), so we report threshold-independent AUROC as the main measure. An error taxonomy shows that failures are systematic, with most false positives related to speech. The pipeline runs on Raspberry Pi hardware costing 30 to 90 US dollars. Probabilistic calibration alone is thus inadequate for reliable cough detection. ## PolyGraph: Plan-Buffered CUDA Graphs for Adaptive Speculative Decoding DevFeed: [PolyGraph: Plan-Buffered CUDA Graphs for Adaptive Speculative Decoding]() Original publisher: [Read original article]() Author: Junchao Yang Published: 2026-10-21T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [CUDA Graphs](), [Model Optimizer]() Tags: [adaptive](), [batch](), [buffer](), [build](), [capture](), [cs-lg](), [cuda](), [cuda-graphs](), [custom](), [d](), [decoding](), [draft](), [fixed](), [gpu](), [gpu-memory](), [graph](), [graphs](), [grid](), [plan](), [polygraph](), [speculative-decoding]() ### AI overview PolyGraph adapts CUDA Graph replay to changing speculative-decoding shapes by capturing one maximum-grid graph and updating a 28-byte GPU-resident plan before each replay. Custom Triton kernels skip tiles outside the active shape while graph topology, allocations, layouts, pointers, and launch grids stay fixed. In a trace-driven prototype using Qwen2.5-3B dimensions on one NVIDIA H800, it replaced 64 shape-specific graphs with one, matched warmed-baseline P50 step latency (17.7 vs. 17.8 ms), and cut measured pre-serving graph-build time from 3,915 to 93 ms. Against maximum-draft padding, its P99 advantage was negligible with no padding waste and reached 18.7% at 100% waste. The evidence is limited to synthetic traces on one GPU; it demonstrates a bounded work-shape mechanism for custom kernels, not a production serving system. ### Source excerpt arXiv:2610.23861v1 Announce Type: new Abstract: Adaptive speculative decoding changes the active batch size B and draft length d from one decode step to the next, whereas a CUDA Graph replays a fixed launch configuration. The resulting verification width M = B d forces serving systems to pre-capture many shapes, capture new ones on the serving path, or pad every step to the largest draft. We present PolyGraph, a plan-buffer mechanism for custom Triton kernels that avoids this choice. PolyGraph captures one maximum-grid graph and writes the active shape to GPU memory before each replay. Kernels skip tiles beyond that shape, while graph topology, allocations, layouts, pointers, and launch grids remain fixed; the update transfers only 28 bytes. On one NVIDIA H800, a trace-driven prototype with Qwen2.5-3B dimensions replaces 64 shape-specific graphs with one, matches the warmed baseline in observed step latency (17.7 vs. 17.8 ms P50), and reduces measured pre-serving graph-build time from 3,915 to 93 ms. Fixed-draft controls show that the benefit comes from avoiding padded payload: the P99 difference from MaxPad is negligible at zero padding waste but reaches an 18.7% reduction at 100% waste. Our evidence establishes a bounded work-shape mechanism for custom kernels, not a production serving system, and is limited to synthetic traces on one GPU. ## dltHub Transformation is in public preview DevFeed: [dltHub Transformation is in public preview]() Original publisher: [Read original article]() Author: Matthaus Krzykowski Published: 2026-05-27T00:00:00Z Content type: release Language: en Sources: [DLT Hub]() Topics: [data-processing](), [Data pipelines](), [AI Integration Strategies](), [AI Context Management](), [elt]() Tags: [agent](), [agentic](), [d](), [data](), [data-engineering](), [pipelines](), [product]() ### AI overview dltHub announces the public preview of its Pythonic data transformation engine and toolkit. The engine runs transformations in supported warehouses, while editor workflows for Claude, Codex, and Cursor help users map source data to business concepts, generate canonical data models, and write transformation code. The post describes customer migrations and reports faster data work, alongside a workflow where humans review and approve generated artifacts. ### Source excerpt Announcing dltHub Transformation in public preview: a Pythonic, agent-native transformation engine and toolkit that turns raw pipelines into a canonical data model in weeks, not quarters. ## Is there space for 'craft' in the world of AI? DevFeed: [Is there space for 'craft' in the world of AI?]() Original publisher: [Read original article]() Author: Jamie Tanna Published: 2026-05-26T13:30:58Z Content type: opinion Language: en Sources: [Jamie Tanna]() Topics: [Artificial Intelligence](), [Large Language Model](), [Code](), [Programming]() Tags: [ai](), [code](), [d](), [large-language-models-llms](), [llm]() ### AI overview An opinion piece considers whether software engineering craft remains valuable as AI tools change how teams plan, write, explain, and ship code. The author supports a balanced approach that uses AI where it helps while retaining human understanding and attention to engineering trade-offs. ### Source excerpt My friend Alan wrote yesterday in a thread on Bluesky: In my end-of-year review I was described as a 'craft engineer' more than once. I don't disagree with this, but I have to wonder if this approach (one that has been comfortable for me for more than 10 years) is totally incompatible with modern AI tooling and a 'deliver it' mindset in a team? I want to split this across what I do; use the new tools and deliver quickly in my workplace, and not use AI at all for my personal work and use that for the 'craft' I can see these being complimentary to each other, but I am struggling with the change in my day job My current AI use at work is mostly limited to using the tools to collaborate on a plan and scope for a change, then writing the change we have agreed upon by hand. This is a happy medium for me, but it's not very conventional and I am noticeably slower than team mates on pure pace of shipping When we have things like a large state machine to determine lazy loading in a component so dense it would take me some time to really appreciate its nuance, offloading that to AI seems reasonable. But it also keeps me in a place of misunderstanding, and I really don't love that When AI can explain code back to us as well as write it, does that misunderstanding matter anymore? Is this just the next layer of abstraction? Having a think about this, I think the TL;DR is yes we should still have some level of "craft" in the world of Large Language Models (LLMs), but there's a balance - as ever - around how much we focus on the craft. As I've written about before, I'm a self-described "cautious skeptic" of AI and LLMs, and I'm trying to do more with AI where makes sense, which makes more of a "moderate" in the discussion, somewhere between AI maximalist and "never use AI". I think I'd argue I'm somewhere more towards the "craft" side of the spectrum, and have been someone who's had similar feedback in the past of being a bit too focussed on "doing the right thing". "engineer" I do ## Teach AI to understand your business with Omni Topics DevFeed: [Teach AI to understand your business with Omni Topics]() Original publisher: [Read original article]() Author: Jamie Davidson Published: 2026-02-19T16:00:00Z Content type: article Language: en Sources: [Omni Blog]() Topics: [Tutorial](), [semantic-layer](), [data-engineering](), [trust and safety](), [AI Chat](), [Machine learning]() Tags: [ai](), [ai-context](), [analysis](), [analytics](), [context-curation](), [customers](), [d](), [data](), [data-security](), [database](), [databricks](), [dbt](), [risks](), [users]() ### AI overview Omni's Topics are curated datasets that give AI analytics agents defined tables, joins, metrics, permissions, and business context. The article explains how these boundaries and semantic guidance can reduce incorrect queries and data exposure, and how teams can refine context, sample queries, and governed models over time. ### Source excerpt In Omni, we use curated datasets called Topics to help you minimize the risks of ungoverned AI. When you've got dozens, hundreds, even thousands of users asking questions, you need a mechanism like Topics to stop AI from going off the rails. ## Libghostty Is Coming DevFeed: [Libghostty Is Coming]() Original publisher: [Read original article]() Published: 2025-09-22T00:00:00Z Content type: release Language: en Sources: [Mitchell Hashimoto]() Topics: [Terminal](), [Library](), [Parsing](), [Sequences](), [cross-platform](), [API](), [Zig](), [C](), [GitHub Actions](), [Visual Studio Code]() Tags: [api](), [c](), [cross-platform](), [d](), [github](), [github-actions](), [hosting](), [library](), [parsing](), [terminal]() ### AI overview Mitchell Hashimoto announces that libghostty, an embeddable library extracted from Ghostty's terminal-emulation core, is beginning to take shape. The first planned component is a zero-dependency library with an API for parsing terminal sequences and maintaining terminal state. A Zig API is available for early testing, while the C API is still forthcoming. ### Source excerpt Libghostty Is Coming Over two years ago, in one of my first public talks about Ghostty, I shared my vision for : an embeddable library for any application to embed their own fully functional, modern, and fast terminal emulator. Libghostty is finally starting to take shape, and I'm excited to share more details about my plans for it. ## Introducing "Create Mode" for embedded analytics DevFeed: [Introducing "Create Mode" for embedded analytics]() Original publisher: [Read original article]() Author: Jamie Davidson Published: 2024-12-06T16:00:00Z Content type: article Language: en Sources: [Omni Blog]() Topics: [product analytics](), [AI Chat]() Tags: [analytics](), [build](), [business](), [conversion](), [create-mode](), [customer-engagement](), [customer-success](), [customers](), [d](), [dashboards](), [data](), [data-science](), [deal](), [engagement](), [feature](), [first](), [flexibility](), [functionality](), [go](), [insights](), [needs]() ### AI overview Omni's Create Mode lets customers of software products explore embedded analytics by creating and saving queries, workbooks, dashboards, and reports. Users can analyze data with AI, point-and-click tools, Excel calculations, or SQL, then share their findings. The article presents self-service reporting as a way to give customers more insight into product usage, while reducing manual report work for customer success teams and supporting engagement, retention, and potential upsells. ### Source excerpt Customers don't just want to see data. They want to interact with it, explore it, and get their own insights, without waiting for your team to make adjustments. With Omni's Create Mode, your customers can run analyses and build reports to get the answers they want - and they won't need your team to pull data for every new question. ## Getting comfortable with psql DevFeed: [Getting comfortable with psql]() Original publisher: [Read original article]() Author: Map Published: 2024-11-11T15:57:56Z Content type: tutorial Language: en Sources: [Craig Kerstiens]() Topics: [Query (disambiguation)]() Tags: [cli](), [commands](), [config](), [d](), [database](), [editor](), [index](), [options](), [pager](), [postgres](), [psql](), [query](), [table](), [tables](), [text](), [time](), [timing](), [toggle](), [variable](), [x]() ### AI overview The excerpt gives practical starting points for using psql: run `\?` to see utility commands, use `\d` to list database tables or `\d accounts` to describe a specific table, and toggle `\timing`, `\x auto`, and `\pset pager 0` to control query timing, output formatting, and the pager. It also notes that psql settings can be customized in the psqlrc file and that `\e` opens a query in the editor configured by `$EDITOR`. ### Source excerpt psql is a CLI editor that ships with Postgres. It's incredibly powerful for working with Postgres, and doesn't take too much of a learning curve to start to get comfortable so you can really feel like an expert working with your database. Just a rundown of a few things to get you started: Once connected in psql you can get an idea of all utility commands available with: \? A handy thing I use all the time is \d. \d will describe the tables within database. You can also add a table/index/etc. onto it to describe that specific table such as: \d accounts There are a number of options you can set in your psqlrc (config) file to customize your CLI experience. But you can also toggle those when directly working in psql. \timing will give you the time it took to run your query \x auto will autoformat your text output \pset pager 0 turns off your pager or 1 to turn it back on Oh and for editing a query in your editor of choice. Make sure you set your $EDITOR enviroment variable to the editor of your choice, though the only right choice is vim: \e Just a few things to get you started working with psql. Your browser does not support video [Next page]()