Papers

Machine-learning detectors for power-system cyberattacks are themselves attack surfaces, and quantum machine learning has been proposed for them. We benchmark fidelity-kernel SVMs and variational classifiers against six tuned classical models on public power-system attack data (Mississippi State/ORNL), across white-box, transfer, decision-based black-box, and poisoning attacks. Our headline finding is methodological: the benchmark's answers are set by the evaluator's choices before the models. Eight choices – six in the evaluation protocol, two in the tuning the benchmark itself runs – each reversed or moved a conclusion at fixed models. The largest is the split: the row-level protocol scores 0.905 macro-F1 where holding whole source files out leaves 0.594, and in the capped matched-dimensionality regime the quantum arm sits within noise of chance with the classical arm 0.024 above it. A fidelity kernel looks most robust until attacked directly (retention 0.886 to 0.064); a mis-fitted surrogate manufactures a 10x asymmetry; an unseeded black-box attack moves 75% between restarts. A positive control explains the accuracy null: the labels, not the pipeline. We give the control that catches each choice and release the seeded benchmark.

While binary quantization theoretically promises extreme compression and acceleration for Large Language Models (LLMs), existing research often overlooks the necessity of specialized hardware kernels, thus failing to unleash the full acceleration potential due to persistent reliance on expensive floating-point arithmetic or runtime dequantization overheads. To bridge this gap, we propose FluxBin (Flexible LUT-based Ultra-low-bit eXecution with Binary bases), an algorithm-kernel co-design that synergizes post-training quantization with a highly optimized CUDA kernel. Algorithmically, we introduce Decoupled Row-Column Binary Decomposition to enhance representational capacity while maintaining hardware efficiency, complemented by a Hessian-guided saliency-aware hybrid bases that preserve critical information. At the kernel level, we implement a Lookup Table Building Approach with Scale Fusion to reduce floating-point arithmetic, featuring a Virtual Columnar Mapping that transforms irregular, sparse, and salient matrices into dense execution. Extensive evaluations demonstrate FluxBin achieves up to $5.92\times$ speedup and $10.19\times$ energy savings across diverse model architectures, delivering comparable accuracy to heavily fine-tuned methods. This effectively enables the deployment of 70B-scale models on one single A100 GPU with a $4\times$ memory reduction. Code is available at https://github.com/nicyyyy/FluxBin.

Compositional Concept Generalization (CoCoGen), the ability to systematically recombine learned primitives in novel contexts, is a key challenge for multimodal learning. In this work, we provide a solution using a compositional model of meaning that separates nouns from relations and uses tensors and variational quantum circuits to train them on data. This model enables us to employ a multi stage training paradigm, one that first learns object representations from single-object image-caption pairs, then subsequently transfers these to the relational stage where object parameters are frozen and optimisation is only applied to relational components. This design explicitly enforces compositional factorisation at the circuit, ensuring that relations are learned as transformations over stable primitives. The training paradigm is tested on the CLEVR dataset developed specificially for CoCoGen. For text, we work with vector representations of nouns and higher order tensor representations of relations using a set of different ansatz. For images, we work with quantum encodings of image embeddings dervied from Open AI's Vision Language tool CLIP and contrast amplitude encoding, which preserves the original embedding geometry, with angle encoding, which introduces nonlinear feature transformations. Our results show that multi-staged training combined with structured encodings significantly improves out of distribution relational generalisation, while using orders of magnitude fewer trainable parameters than classical baselines. We find that performance gains arise from the interaction between representation and encoding, with nonlinear quantum encodings enhancing the separability of compositional structure. These findings demonstrate that structured quantum representations and staged learning provide an effective framework for compositional generalisation in multimodal quantum machine learning.

Production paged-serving engines apply uniform paging granularity to the KV cache, even though the two regions of a multi-agent workload have opposite storage requirements: a long shared prefix demands contiguity, while the per-request suffix demands fine-grained allocation. We present GraniKV, a KV-cache layer that allocates the shared prefix in a contiguous HOT pool and the suffix in a token-level COLD pool, combined with a per-step dispatcher which selects the appropriate backend among dual backends for each regime (compute-, memory-, or communication-bound). To the best of our knowledge, GraniKV is the first system to apply asymmetric paging granularity to the KV cache of a production paged-serving engine. At $L_p{=}16$\,K shared tokens GraniKV reaches $\mathbf{2.16\times}$, $\mathbf{1.98\times}$, and $\mathbf{1.57\times}$ output-token throughput over the production baseline on Llama-3.1-8B/TP=1, Qwen-2.5-14B/TP=2, and Qwen-2.5-32B/TP=4. The gain decomposes: cascade attention integration contributes the majority at saturation; the asymmetric storage layer adds $1.05$–$1.15\times$ end-to-end while being what makes the batched-GEMM prefix backend possible at all. Under heterogeneous multi-agent serving with distinct prompts of different lengths, the attribution inverts: GraniKV sustains $\mathbf{1.95\times}$ while batch-global cascade collapses to parity — the storage layer alone carries the win in the regime that motivates the paper.

Weight-only post-training quantization (PTQ) enables the deployment of large language models under tight memory budgets, but accuracy often collapses at 2-3 bits. Existing backpropagation-free PTQ optimizers have two limitations: group decisions ignore the correction that the remaining continuous suffix can absorb, and discrete refinements typically keep the affine quantization grid fixed. We introduce SCHUROPT, which analytically eliminates the suffix's optimal continuous response, yielding an exact groupwise quadratic with Schur-complement curvature. It then alternates closed-form row-wise scale/zero-point refitting with coordinate descent over integer codes. With the GPTQ objective fixed, SCHUROPT improves mean zero-shot accuracy on 2-bit Qwen3-4B by 11.88 percentage points (pp). At higher precision, however, tighter reconstruction does not consistently improve end-model metrics. SCHURQUANT therefore combines SCHUROPT with quantized-prefix teacher reconstruction, reference-weight regularization, residual-add targets, and teacher-decision token weighting. Across eight Llama and Qwen models, SCHURQUANT achieves the highest mean zero-shot accuracy among the evaluated backpropagation free PTQ baselines, outperforming the strongest baseline by 9.65 pp at 2 bits.

Anomaly detection in dynamic graphs underpins financial fraud analysis, intrusion detection, and platform integrity, where automated decisions require human-interpretable justifications. StrGNN, the strongest performer in recent benchmarks, produces no explanation: when an edge is flagged, the analyst receives only a score. Explanation metrics are undefined for StrGNN because no attribution vector exists. This paper closes that gap. We present X-StrGNN, a post-hoc explanation layer that wraps a trained, frozen StrGNN and emits, for every flagged edge, dual attributions: a structural attribution identifying which contextual interactions in the enclosing subgraph drove the decision, and a temporal attribution identifying which historical snapshot carried the signal. Both attributions are multiplicative masks identically one in the unexplained pass, so the layer is an exact pass-through: detection is preserved to machine precision, verified rather than asserted (Delta AUC = 0.0000, Delta AP = 0.0000, Delta P@100 = 0.0000). Attribution costs 0.66 ms per edge, making explanation of an entire alarm list feasible. We conduct the first controlled design study of attribution strategies for this architecture, comparing gradient attribution, per-instance mask optimisation, and amortised parameterisation under one protocol, one budget, and three seeds. X-StrGNN attains the highest stability (0.913) at 268x lower cost than per-instance optimisation, and its temporal attribution (1.601 against a measured random floor of 0.973) is separably better than its ablated control, while per-instance optimisation - the most expensive strategy - falls below that floor. Code, protocol, and per-seed measurements are released.

For $m\geq 2$, let $c_p(m)$ be the all-dimensional best constant in $$ \left\|\sum_{k=1}^m A_k\right\|_p \leq c_p(m)\left\|\sum_{k=1}^m |A_k|\right\|_p. $$ Tang and Zhang conjectured an explicit formula for every finite $p>1$. We disprove the conjecture with two explicit real $2\times 2$ rank-one matrices at $p=3/2$. The comparison is certified by seven strict rational inequalities and, in particular, places the attained ratio above $207/200$, while the conjectured constant lies below $207/200$. On the positive side, we prove the conjectured sharp bound for every family of rank-at-most-one summands when $2\leq p<\infty$, and classify all equality cases. We also prove the corresponding endpoint statement for $p=\infty$. Finally, for arbitrary complex matrices, we establish the conjectured sharp constant in the case $m=2$, $p=4$.

Video models are increasingly used to predict what happens next in a scene, yet the metrics commonly used to compare their outputs say little about whether the predicted objects move correctly. Motion, geometry, identity, background stability, and visual similarity can fail independently, but whole-frame scores often mix these errors together. We introduce RigidBench, a simulator-grounded benchmark that compares a generated continuation with a reference rollout from the same initial frame and motion description. Its five rigid-body tasks vary objects, materials, viewpoints, and indoor and outdoor scenes, with per-frame masks, depth, 6-DoF trajectories, and contacts available for scoring. We evaluate eight models on the same 100 examples with ten measurements that keep these aspects separate. The resulting rankings depend strongly on what is measured: no model leads on all ten, and across model means, higher SSIM accompanies larger 3D trajectory error (r = 0.89). RigidBench also includes 5,000 training videos with exact simulator state, which we use to fine-tune and analyze Wan 2.2 TI2V-5B. Full fine-tuning reduces 3D trajectory error by about 20% with almost no change in SSIM, while teacher-forced probes and targeted interventions show that object position is represented throughout Wan's diffusion transformer and used by its denoising computation.

Machine unlearning (MU) aims to remove the influence of specific training data while preserving model utility. As the name suggests, MU can be viewed as the inverse of learning, using gradient-based updates to reduce the influence of a forget-set by counteracting the previously learned behavior. Recently, Muon, a gradient descent variant, has been introduced. Muon applies spectral magnitude normalization to encourage exploration of rare directions and demonstrates promising performance. Inspired by Muon, we adopt the spectral view for unlearning and propose Spectral Saliency Unlearning (SSU). SSU thresholds weak singular components and updates only those directions supported by a confident unlearning signal. We further provide theoretical justification for this thresholding approach from the perspective of the forgetting-retention trade-off. Experiments across image classifiers, diffusion models, and LLMs demonstrate SSU's effectiveness.

We present L3Cube-IndicQuest v2, a large-scale gold-standard multilingual question-answering benchmark for evaluating the India-specific factual knowledge of Large Language Models (LLMs). The benchmark comprises 3,471 curriculum-grounded English question–answer pairs spanning nine domains, curated from educational curricula, competitive examination materials, and domain-specific reference books. We introduce a practical hybrid construction strategy that combines context-grounded LLM-based question generation and validation with semantic deduplication and human verification, enabling scalable creation of benchmark data while preserving annotation quality. The benchmark is translated into 19 Indic languages, yielding a publicly released multilingual dataset of 69,420 question–answer pairs across 20 languages. We evaluate six LLMs under three protocols: LLM-as-a-judge and two deterministic lexical criteria, exact-substring and word-overlap matching. All three produce almost the same model ranking, showing that the results do not depend on the choice of judge. The frontier commercial model leads by a wide margin, and among open-weight models Gemma4 31B outperforms the Indic-specialised Sarvam 30B in every evaluated Indic language.

Linear attention models eliminate the quadratic prefix computation and context-growing KV cache of softmax attention by replacing pairwise token interactions with recurrent state updates. However, existing decoding implementations often materialize and write back the full recurrent state after every generated token, making state maintenance a major source of memory traffic, especially for models with large states and many heads. This paper presents DeltaLog, a recurrent-state decoding scheme that reduces this overhead without changing the model semantics. Specifically, DeltaLog represents the recurrent state as a dense base state together with a bounded log of recent compact updates. Most decode steps append only compact update factors to this log, while periodic merge steps fold the accumulated updates back into the dense base state. Thus, the model observes the same dense state as in eager decoding, but most full-state write-backs are replaced by lightweight append operations. We implement DeltaLog for GDN, KDA, and RWKV6 and integrate it into a prototype serving stack. Across these models, DeltaLog accelerates the recurrent-state update kernel by up to $1.86\times$, reduces profiled recurrent-state write traffic by up to $7.83\times$, and achieves $1.05$–$1.20\times$ end-to-end serving speedups over dense recurrent baselines.

A multimodal system may begin inference while holding only some of its inputs and may acquire the rest at a cost. With adaptive acquisition, the policy determines which inputs are ultimately observed, so we state risk control conditional on that terminal input pattern. Conditional calibration typically assumes that the grouping map is fixed independently of the calibration sample, a condition that policy-induced grouping does not satisfy. We characterize when pattern-conditional risk control remains valid and give two finite-sample constructions: threshold-free routing with calibration applied at the terminal pattern, and simultaneous validation of complete policy-pattern pairs, which lets calibration data select the deployed policy. A counterexample shows that validity proved for a calibration-independent grouping map need not transfer once the policy makes the terminal group calibration-dependent. We call the resulting method RouteCert. On a clinical electrocardiogram task with a staged, cost-ordered lead protocol, the deployed policy answers 71.2% of held-out patients, with an observed 7.4% disagreement with the cardiologist's diagnosis at 48.8% of the prespecified ordinal cost of acquiring each stage, and all three acquisition stages are validated separately. On masked multimodal benchmarks, validating pointwise at each terminal pattern holds observed worst-pattern selective risk, measured against the full-information reference decision rather than the true label, at 0.034, where a pooled design reaches 0.145 against a 0.10 cap, at a comparable answered fraction (0.350 vs 0.342); under the budget-matched simultaneous comparison, the answered fraction falls to 0.305.

Vision-Language Models (VLMs) have demonstrated strong performance in multimodal understanding and generation. However, fine-tuning of VLMs typically relies on centralized data, which raises privacy concerns in certain domains (e.g. healthcare). Federated Learning (FL) provides a natural solution by enabling model training without sharing raw data. However, applying FL to VLM instruction tuning is highly challenging. VLMs have substantial parameter scales, and in real-world scenarios, clients exhibit significant heterogeneity in tasks, modalities, and model architectures. Existing methods mainly focus on simplified settings and are unable to handle such multi-dimensional heterogeneous scenarios. In this work, we study federated instruction tuning under joint heterogeneity in tasks, modalities, and model architectures. We propose UniFed-VLM, a unified federated instruction tuning framework for VLMs that addresses multiple types of heterogeneity. It consists of two key components: 1) Federated Compensated Subspace Aggregation (FedCSA), which performs subspace-aligned aggregation of parameter-efficient adapters with dynamic weighting and compensation to mitigate heterogeneity-induced conflicts; 2) Two-stage Collaborative Distillation (TCoD), which enables effective knowledge transfer across heterogeneous models via a Mutual Distillation Adapter (MDA) and a mixture-of-experts-based distillation strategy. We conduct experiments on multiple benchmark datasets, and the results show that UniFed-VLM achieves stronger average performance across diverse tasks compared with existing FL methods. The source code is available at: https://github.com/wangpengyu2004/UniFed-VLM.

Task guided agents demonstrate strong performance in a wide range of complex tasks. However, most existing task representation algorithms are tailored to specific contexts and struggle to generalize across diverse scenarios. Moreover, they typically depend on gradient signals from reinforcement learning controllers to update their weights, which can degrade both representation quality and learning efficiency. To overcome these limitations, we propose LOTUS, a temporal logic inspired universal task representation framework that can be seamlessly integrated into any RL algorithm to enhance agent performance across diverse task settings. Specifically, we design a novel task representation architecture capable of modeling relationships and extracting task semantics from LTL formulas. We further introduce a more effective update mechanism that treats the LTL encoder as a policy, thereby improving representation capacity. To enhance stability and robustness, LOTUS leverages the bisimulation metric, which provides theoretical guarantees for LTL representation, including behavioral equivalence, optimality fidelity, and trajectory robustness. Experimental results show that LOTUS outperforms most existing methods in learning efficiency, generalization capability, and representation quality. Specifically, LOTUS accelerates convergence over 20% in single-task scenarios, achieves a 15%-45% higher success rate in unseen manipulation tasks, and improves generalization performance over 25% in complex multi-task environments with increased sub-goal depth or conjunctions. The corresponding code, videos, and appendix are available at: https://lotus-website.github.io/.

A consistent concept of the current time is important for temporal reasoning, yet how language models represent the current time is not well understood. We contribute two tasks that probe the current year in conceptually distinct ways: an associative task, which infers the current year from verb tense, and a declarative task, which directly queries for the current year. Both tasks estimate current years within one year of the post-training data cutoff of instruction-tuned language models. For base models, predictions on the associative task serve as a strong proxy for the pre-training data cutoff, with an average error of only 10 months across 13 models. However, their internal mechanisms diverge: the associative task uses mechanisms similar to factual recall, while the declarative task lacks consistent causal pathways. This divergence poses a challenge for updating the current year in language models. None of prompting, SFT, or weight editing succeed in shifting the associative and declarative years simultaneously. Prompting updates the declarative year (94.6% success across 351 target years) but leaves the associative year nearly unchanged (1.7% success). Year-shifted SFT also fails to shift the associative year, matching the target year in only one of eight models. Weight editing, while effective for both tasks individually, does not generalize across both. Overall, our results show that the current year is not consistently encoded in language models: The associative notion, deeply ingrained in linguistic structures learned in pre-training, uses different causal mechanisms and resists the same modifications that easily shift the declarative notion learned in post-training.

Encrypted traffic classification is vital for network security, yet real-world deployments are inherently sensitive to rare but high-loss errors such as misclassification of malicious traffic. The encrypted traffic foundation model, as a promising general-purpose technique, can achieve impressive overall performance. However, employing standard objectives such as empirical risk minimization often overlooks high-risk tail events, and commonly used performance metrics hardly reflect robustness limitations in risk-sensitive scenarios. Directly applying robust optimization objectives, such as conditional value-at-risk, to post-training is computationally prohibitive for large models, as identifying high-loss samples exhausts substantial computation. To this end, we propose Pre-Evaluation Robust Optimization (PERO), an efficient robust post-training framework for encrypted traffic foundation models. PERO employs a lightweight proxy to estimate sample-wise risk and selects a subset of high-risk samples to update the foundation model, decoupling risk estimation from expensive large-model optimization. Extensive experiments on typical encrypted traffic datasets show that PERO achieves competitive or superior robustness and average performance compared to outstanding robust post-training methods, while significantly reducing computational and memory costs.

Finding representative waveforms in long time series has scientific and practical value in many domains, as it enables summarization and visualization of large time series datasets, and downstream tasks like classification and forecasting. We present here QSMP, a method to find representative waveforms in long time series through a density-guided clustering of time series subsequences. Our method makes a novel connection between Quick Shift, a mode-seeking algorithm, and the Matrix Profile, a time series similarity-search data structure, to adapt Quick Shift to the clustering of subsequences in long time series, with a space complexity that is superior to the state-of-the-art method. Our experiments on synthetic and real datasets show that QSMP can be a valuable tool to summarize and visualize long time series by finding representative waveforms.

Modern deep networks are trained through long update trajectories, yet their temporal organization remains less systematically characterized than architectures, losses, or optimizers. We study short-horizon predictability as a measure of temporal redundancy: where, when, and under which training conditions recent updates contain information about near-future parameter motion. We combine three complementary probe families, displacement-direction, subspace-residual, and predictor-based probes, with convention-aware, null-calibrated group-level readouts, and apply them to multi-pass vision training on CIFAR and public Pythia pretraining checkpoints. Across both regimes, vector-like tensors such as normalization parameters and biases (auxiliary parameters) exhibit simpler short-horizon dynamics than matrix-like feature-transforming weights (bulk parameters), whose predictable behavior concentrates in localized, time-varying pockets. Agreement within and across probe families, and with independent trajectory diagnostics, indicates that these measurements capture intrinsic trajectory structure, while probe differences distinguish complementary forms of temporal organization. Controlled CIFAR comparisons further show that architecture and training recipe systematically modulate the measured structure. A Pythia-70M case study further exposes a sequence of role-, depth-, and scale-dependent events, including bulk ESA falling below the random sign-agreement level and the emergence and redistribution of predictable qkv pockets across layers. These results position short-horizon predictability as a retrospective, parameter-resolved diagnostic of training dynamics.

The problem of networked information aggregation, studied in Kearns et al. (2026), involves a group of learners situated on the vertices of a directed acyclic graph $G$, each learning a linear predictor $\widehat Y$ for a fixed random variable $Y$ given access to a local feature, as well as the predictors learnt by its parents. Learning proceeds iteratively, with learners ordered according to a topological sort of $G$. The main quantity of interest is the error incurred by the current learner, constrained to this flow of information, with respect to the best linear predictor using all the features seen so far. When the studied error is the MSE, i.e., $\mathbb{E} (\widehat Y - Y)^2$, Kearns et al. (2026) show that the error is at most $O(1/\sqrt{D})$ along a path of length $D$. They also obtain a hard instance where the MSE is lower bounded by $\Omega(1/D)$, leaving the correct order open. In this work, we resolve this central open problem, and obtain a family of worst case problem instances with a MSE lower bound of $\Omega(1/\sqrt{D})$. By exploiting invariances in the structure of the learnt predictors, our analysis generalizes to all convex loss functions $\ell(\widehat Y, Y)$ satisfying regularity conditions which include strong convexity in a ball around the origin, and that the ideal predictor minimizing the population loss is positively correlated with the label. We show that networked information aggregation on a gaussian instance in our worst case family incurs an $\ell$-error lower bounded by $\Omega(1/\sqrt{D})$ with respect to this ideal predictor. We demonstrate that a variety of common losses satisfy these regularity conditions. In particular, the logistic loss satisfies them, and hence our analysis also closes the gap between the upper and lower bounds in Bateni et al. (2026).

The population structure of an inbred population of 781 people on Norfolk Island in the Pacific, 318 of which are descendants of the original Mutineers of the Bounty, is analyzed phenotypically using shape from stereo retinal fundus photographs. Three-dimensional optic nerve head (ONH) shape is reconstructed from stereo pairs by a multi-scale stereo matching algorithm. Using deep neural network, the shape of ONH, which is under genetic control, is decomposed into a set of hierarchical features through self-taught learning. Features captured at different levels are selected according to their discriminant power in identifying the two populations. The prediction accuracy is evaluated with stratified cross validation. Given the selected feature set, individuals are grouped into k hierarchical clusters and cluster membership fractions are determined for k=2,3,4,5,6,7. Population structure analysis on the basis of phenotypes through image analysis allows heritability and linkage analysis, including founder effects from English and Polynesian ancestors, potentially leading to new genetic risk factors for glaucoma and other ONH-related eye diseases.

Distributional changes can be invisible to means and covariances yet appear in skewness, asymmetric interactions, or other third-order structure. We develop a nonparametric change-point method that retains every degree-at-most-three coefficient of a density while avoiding direct density estimation. For observations in $[-1,1]^d$, we construct a symmetric order-three Legendre feature tensor $H_3(X)\in\Sym^3(\R^{d+1})$ such that $A(f)=\E_fH_3(X)$ is an exact isometric encoding of the degree-three density projection: $\|A(f)-A(g)\|_{\F}=\|P_3(f-g)\|_{L^2}$. Instead, fixed tensor contractions are degree-three polynomial chaoses with $\psi_{2/3}$ tails. The two terms have the characteristic order-three tensor scaling and match the powers in sharp concentration results for simple random tensors. For a coordinate-orthogonal specialization, the bound improves to $\sqrt{\log d}$ and enables a prefix-sum implementation in hundreds of dimensions. We derive the exact population tent shape and localization margin, introduce a seeded shortest-interval algorithm with a padded local recentering step, and prove exact recovery by induction: null recursive segments remain inactive, every undetected change retains a balanced isolating interval, and the shortest active seed contains exactly one change before recentering. A two-way cross-fitted scalar refinement attains $O_{\Pp}(\kappa^{-2})$ localization in the small-jump regime, matching a Le Cam lower bound on a pure cubic family whose degree-two projection jump is exactly zero. Reproducible experiments at $d\in\{20,50,100,200\}$ and a three-change $d=100$ sequence demonstrate the intended high-dimensional regime without materializing a $(d+1)^3$ tensor.

Identification of IoT device types from passive traffic is increasingly used for security management in enterprise and ISP networks. However, the performance of machine learning-based classifiers gradually degrades under concept drift as device behavior evolves. Therefore, maintaining classification performance requires periodic retraining with newly labeled deployment traffic. The operational challenge is determining how much and which deployment traffic instances to label for maintaining classification performance. We show that these two decisions should be treated separately. While retraining solely on instances selected by a drift detector is prone to systematically overlooking parts of the emerging behavioral space, uniformly sampled deployment traffic captures more representative behavioral changes. Instead, drift detection is more effective at determining the amount of deployment traffic that should be labeled. We make three contributions. (1) We conduct a two-year longitudinal study of IoT traffic and characterize how behavioral evolution manifests across device classes and how retraining with newly labeled traffic restores classification performance. (2) We develop a conformity-based drift detector that captures class-conditional behavioral models directly from raw traffic features and provides feature-level explanations of behavioral evolution. (3) We demonstrate that adjusting the traffic labeling rate according to the observed behavioral evolution, combined with uniform traffic sampling, maintains classifier performance more effectively than detector-guided sample selection and is beneficial to managing the traffic labeling effort. We further show that this strategy performs comparably to confidence-guided adaptation while providing feature-level explanations. Our evaluation uses 3.8 million IPFIX flow records collected from 21 IoT types over more than 2 years.

Attention mechanisms have driven machine learning for a decade, from neural machine translation to language models that do general-purpose reasoning. This survey covers four connected threads: their formulation for sequence-to-sequence tasks, adaptation to computer vision, efficiency innovations that address the quadratic bottleneck, and advances in interpretability. We define three criteria: efficiency, expressiveness, and interpretability, and compare twenty-one methods using an EEI scoring framework. Scores come from a single rater with an assumed +/-1-point perturbation range. A deterministic Monte Carlo analysis with 200,000 samples shows that, under this perturbation model, rank changes of more than one position occur in 67-70% of samples on average. A rank-matched null model reproduces a similar stability profile, so the results support coarse tier-level comparisons rather than fine-grained rankings. The survey traces attention from Bahdanau-Luong alignment through the Transformer and into vision architectures. It reviews fixed and learned sparse attention, linear attention, IO-aware exact algorithms including FlashAttention, and state-space alternatives including Mamba. It also covers induction heads, superposition, and the attention-SSM duality. We further provide a structured narrative review, a benchmark synthesis with cross-study caveats, a five-problem research gap analysis, and a 2015-2026 evolution timeline. We conclude by framing attention research as an expansion of the efficiency-expressiveness-interpretability frontier and identifying future directions including unified efficiency benchmarks, learned routing for hybrid architectures, length generalization, and scalable mechanistic interpretability.

Dynamic Mixture-of-Experts Serving allocates k replica GPUs among m experts as workloads change. At each round, the online algorithm sees the current workload, chooses integral replica counts, and pays bottleneck service cost plus replica movement. It does not know future workloads. Huang, Lou, and Xiao gave an O(sqrt(log k))-competitive randomized algorithm for this problem. We prove a deterministic O(1)-competitive algorithm. For every number of experts and every k>=1, the algorithm satisfies ALG_det <= 10 C_PB OPT + (5 C_PB + 8) k + 16, where C_PB is the absolute constant from Chasing Positive Bodies at resource augmentation one and covering sparsity two. Consequently, CR_det(k)<=10 C_PB for every k>=1, so CR_det(k)=Theta(1). The multiplicative factor does not depend on the number of experts, replica budget, horizon, or workload values. Thus randomization is not needed for the asymptotic guarantee. The proof has two layers. A finite tangent envelope, summable positive resets, and a nonexpansive balanced projection reduce reciprocal-max service costs to a deterministic exact-budget fractional path. A new deterministic rounding theorem converts every such path to integral allocations with service distortion three and movement bounded by the fractional movement plus 6k. The complete reduction, rounding theorem, causal composition, and quantified main theorem are machine-checked in Lean 4 relative to the positive-body result as the sole scientific source premise. The theorem concerns the allocation model above. It does not include network topology, shared-edge congestion, or routing decisions.

Large language models increasingly rely on sampling as a driver of their own improvement, making the fidelity of their learned distributions more critical than ever. Yet, not all distributions are equally easy to learn. In this work, we identify a curse of ambiguity: in large language models, and more broadly in all neural networks that produce discrete probability distributions, the more ambiguous a next-token distribution is, the harder it is to learn accurately. Through an extensive theoretical analysis, we trace this curse to architectural and learning roots. More ambiguous distributions require more capacity to be stored, larger embeddings to be represented, more steps to be fitted, and amplify token-sampling noise. We validate these findings on synthetic tasks with controlled ground truth and observe the same signatures in language models trained on real data. Our results provide a new perspective on the statistical capabilities of large language models and a practical framework for when to trust their output distribution.

Mobile money has widened financial access across Sub-Saharan Africa and enlarged the surface for money-laundering and terrorism-financing (ML/TF) activity in ecosystems dominated by high-volume, low-value transactions. Rwanda is a case in point: several million active mobile-money users, telecom-led wallets on the MTN and Airtel networks, and a Financial Intelligence Centre (FIC) supervising transaction streams whose scale exceeds static rule-based monitoring. This paper develops and evaluates a transaction-monitoring framework aligned to the Rwandan AML/CFT regime under (i) extreme class imbalance (~0.1% prevalence), (ii) scarce and delayed labels, and (iii) bounded investigator capacity. Using SAML-D, a synthetic dataset of 9,504,852 transactions with 17 laundering typologies, we engineer account-centric behavioural features (rolling velocity, net-flow directionality, counterparty diversity, burstiness) and benchmark supervised classifiers (Logistic Regression, Random Forest, LightGBM), unsupervised anomaly detectors (Isolation Forest, Local Outlier Factor), a dense autoencoder, and a late-fusion meta-learner. Evaluation is operational: PR-AUC, recall at a calibrated ~90%-precision point, recall at top-K%, and alerts per 10,000. On the chronologically held-out test period, LightGBM attains PR-AUC = 0.0469, capturing 64 laundering cases at precision ~0.89 with 0.51 alerts per 10,000; the fusion stacker reaches PR-AUC = 0.0477 at precision ~0.91 and 0.46 alerts per 10,000, recovering 59 true positives. We map score bands to Rwanda-relevant analyst workflows and STR/SAR escalation, and outline a staged path from synthetic prototyping to real-data validation with the National Bank of Rwanda and FIC. The contribution is operational: a governance-aware pipeline and evaluation protocol calibrated to the constraints of an African mobile-money regulator, not a new algorithm.

Building approximate nearest neighbor (ANN) indexes at billion scale is often dominated by expensive global clustering or graph construction, making time-to-index a first-order systems concern. We present NeuRoute, a learned hashing index that turns short binary codes into an effective routing primitive for large-scale vector search. NeuRoute trains a lightweight neural network encoder with a selective similarity-preserving objective to produce well-balanced binary addresses. During construction, NeuRoute organizes vectors into buckets by their codes and performs bucket-local clustering in the encoder's low-dimensional space to form centroids. At query time, NeuRoute exploits the encoder logits as an uncertainty signal: it uses deviation-to-threshold scores to prioritize uncertain-bit perturbations for query-adaptive multi-bucket probing, scores bucket-local centroids by their distances to the query to form a compact candidate cluster set, and applies centroid-stage gating with heap-quality-driven early stopping to prune low-value clusters before exact refinement. On billion-scale benchmarks, NeuRoute achieves strong accuracy-throughput trade-offs with fast index construction: on BigANN-1B it reaches $90.3\%$ Recall@10 at 2,414 QPS and is $1.7\times$ faster than OPQ+IVF-PQ (refine) at comparable accuracy, while completing end-to-end training+construction in under an hour on both BigANN-1B and Deep1B-1B. These results show that logit-guided neural routing can make hashing competitive as a lightweight ANN indexing framework at billion scale. Source code and artifacts are available at https://github.com/XingqiaoWang/NeuRoute.

Prior embedding models for sequential recommendation typically operate within a homogeneous action space, limiting their ability to capture cross-surface behavioral signals spanning distinct behavioral domains. We present SAGA, a generative action embedding model that encodes multi-surface user interaction sequences across a Financial Service organization's ecosystems, from checkout, peer-to-peer (P2P) transactions, in-app engagement, email to account actions, into a unified user representation for downstream recommendation tasks. Central to SAGA is a per-field tokenization schema that decomposes each action event into multiple field-level tokens (e.g. product, interaction, surface), enabling field-level attention and per-field training objectives that fused single-token approaches cannot support. Through an offline ablation study on loss formulation, tokenization granularity and training data scope, we isolate the contribution of each design choice. A downstream model integrated with SAGA-generated user embeddings delivers the strongest overall click and conversion lift across diverse downstream touchpoints, compared to all ablated and alternative architectures.

The rapid adoption of large language models has enabled the development of clinical multi-agent systems (MAS) capable of integrating multimodal patient data and supporting increasingly complex clinical decision-making. However, the deployment of these systems in real-world healthcare settings raises critical ethical concerns related to safety, fairness, accountability, transparency, and patient trust. While numerous organizations, including the World Health Organization, the National Academy of Medicine, and the FUTURE-AI consortium, have proposed ethical frameworks and governance principles for healthcare AI, these efforts remain largely conceptual. To address this challenge, we present ETHOS (Ethics and Trust through Hierarchical Oversight System), a modular ethics framework designed as a governance meta-agent that can be integrated with any existing multi-agent system without requiring changes to its underlying architecture. ETHOS translates stakeholder-informed ethical requirements into executable runtime oversight through a layered governance approach consisting of deterministic checks, contextual reviews, and a final ethics critic. These components continuously evaluate intermediate reasoning steps and final outputs, enabling the system to identify ethical risks, request revisions, or suppress responses that fail predefined safety and trustworthiness criteria. We demonstrate ETHOS within a hepatology clinical decision-support MAS. Results show that ETHOS improves decision reliability by detecting incomplete, inconsistent, or out-of-scope evidence and appropriately increasing abstention when safe recommendations cannot be supported. By embedding ethical governance directly into system operation, ETHOS provides a practical and auditable mechanism for transforming high-level AI ethics principles into deployable safeguards.

Encoder-based code representation models remain widely deployed for discriminative tasks such as clone detection and code classification, where their small size and low inference cost are decisive. Their robustness, however, is fragile: under invariant programs, semantically equivalent code written in different syntactic forms, learned representations degrade substantially even though program behavior is unchanged. We present an empirical study of this robustness gap across four encoder baselines, two downstream tasks, and four datasets, together with a minimal code-only continued pretraining recipe that closes much of it. Our method, invariant pretraining (InvPT), applies semantic-preserving transformations to the corpus and combines masked language modeling with multi-positive supervised contrastive learning that treats all augmentations of the same source function as positives, mixing self-contrast pairs (same code, different masks) with invariant-contrast pairs (transformed code) for positives of varying difficulty. Unlike prior contrastive code encoders, InvPT does not require paired natural-language data. Across our evaluation, InvPT improves robustness on transformed test sets by up to 11 percentage points on clone detection and 19 on code classification while matching or improving standard accuracy, and our ablations isolate multi-positive invariant contrast as the main source of the gains. Our aim is not a new objective but a careful measurement of where encoder robustness breaks and how far a simple, code-only recipe can recover it.