Papers

Reliable uncertainty evaluation is essential for assessing the trustworthiness of machine learning (ML) predictions, particularly in safety-critical applications. This paper explores conceptual analogies between uncertainty evaluation in ML and measurement science, drawing on the evolution of metrology from error-based approaches to the framework established by the Guide to the Expression of Uncertainty in Measurement (GUM). Particular attention is devoted to the distinction between process-level and result-level quality and to the limitations of the conventional classification of ML uncertainty into aleatoric and epistemic components. It is argued that this dichotomy, similarly to the traditional distinction between random and systematic measurement errors, is inherently model-dependent and provides an insufficient basis for a unified uncertainty evaluation framework. Further analogies between machine learning and measurement field of knowledge are identified concerning distribution shifts, anomalous operating conditions, and the validity of uncertainty evaluation. Building on the proposed considerations, the paper advocates a metrology-inspired approach based on the systematic identification, quantification, and propagation of uncertainty contributions. By extending the methodology already established in measurement science to ML, this approach could offer a potentially rigorous, consistent, and transferable framework for evaluating the trustworthiness of individual predictions.

Transformers can implement reinforcement-learning algorithms in context, yet how off-policy control emerges within Transformers remains poorly understood. We study Q-learning, a foundational model-free control algorithm that combines off-policy bootstrapping with greedy policy improvement. We prove that a two-layer attention-only Transformer can implement a finite-temperature form of Q-learning, where the temperature $\tau>0$ controls the approximation to greedy action selection, recovering window Q-learning as $\tau\to0$. We train the structured Transformer on randomly generated MDP tasks using output-only supervision. Experiments show that the theoretically predicted parameter structure emerges during training, including the internal greedy action-selection mechanism, and that the trained Transformer achieves performance comparable to Q-learning on unseen tasks. Our theoretical analysis further shows that the equivalent realizations form a solution manifold and that, under suitable task-diversity conditions, the training iterates converge locally to this manifold at an exponential rate. For the random-task model analyzed here, the local worst-case number of gradient steps needed to reduce the distance to the manifold by a factor $\epsilon$ is $\Theta\left(\tau^{-5}\log(1/\varepsilon)\right)$, quantifying the increasing optimization cost as the learned update approaches window Q-learning.

We study stochastic bandits without knowledge of reward scales or tail parameters. Almost-sure regret guarantees describe eventual learning but do not quantify expected performance or the risk of large regret in a finite run. We introduce $\texttt{EL-F}$, an anytime empirical-likelihood (EL) extension of the $g$-Forcing policy of Cowan and Katehakis (2020), and characterize its performance at three levels. First, for integrable rewards and non-decreasing, diverging exploration schedules $g$ with vanishing increments, $\texttt{EL-F}$ preserves the $g$-Forcing asymptotics for pseudo-regret $R(T)$: $R(T)/g(T)\to\sum_{i:\Delta_i>0}\Delta_i$ almost surely as $T\to\infty$, where $\Delta_i$ is the suboptimality gap of arm $i$. Second, on every sub-exponential instance, for non-decreasing $g(t)=\omega(\log t)$ satisfying a mild feasibility condition, expected pseudo-regret is at most $g(T)\sum_{i:\Delta_i>0}\Delta_i+C_{\underline{P}}$, where $C_{\underline{P}}$ is an instance-dependent constant independent of $T$. Moreover, we also establish finite-time tail bounds with super-polynomial decay in regret beyond the forced-exploration budget. For slowly varying $g$, we determine the exact exponent of linear regret deviations on two-armed Gaussian instances with common variance and establish a matching lower bound among policies satisfying the same expected-regret budget on every such Gaussian instance. These quantitative guarantees also cover the greedy variant obtained at zero EL radius. A fixed positive radius introduces local optimism while preserving the leading asymptotic regret. We establish coupled pathwise dominance over the greedy variant in two-armed bandits with a deterministic suboptimal arm. Experiments demonstrate reductions in prolonged suboptimal allocation compared with $g$-Forcing on challenging instances.

Anti-money laundering (AML) in financial networks requires modeling transaction patterns that span multiple institutions. In practice, however, banks cannot share account and transaction features due to strict privacy regulations. We propose a privacy-preserving graph neural network (GNN) inference framework tailored to this multi-bank setting. A central server holds the graph connectivity and pretrained model weights in plaintext, while each bank keeps its transaction features encrypted. Inference is then carried out entirely on encrypted features under the CKKS homomorphic encryption scheme, so that no plaintext transaction features or predictions are exposed to the server. The key to making this practical is an edge-centric GNN architecture, wherein message passing is expressed through operations on node–edge incidence matrices. This formulation not only preserves the edge-attributed multigraph structure of transaction networks but also reduces every aggregation step to plaintext–ciphertext matrix products. Since CKKS can evaluate only polynomial circuits, we also approximate each nonlinearity with a composite polynomial of the sign function, which we calibrate per activation site. On the IBM AML benchmark, our model achieves performance competitive with plaintext baselines, and with a consumer-grade GPU, classifies each queried transaction end-to-end in as little as 5.31 seconds, while per-transaction latency can be amortized to 1.65 seconds when optimized for throughput.

General-purpose robot manipulation policies learn short-horizon skills effectively, but degrade rapidly on long-horizon tasks. Existing benchmarks cannot isolate failures because task horizon is either too short or entangled with scene and object variations. We propose LADDER, a long-horizon robot manipulation benchmark scaling from 1 to 10 subgoals within the same environment over three suites (SimplerEnv-Long, RobotArena-$\infty$-Long, and RoboLab-Long), 145 tasks, eight categories, two embodiments, and 2,024 simulated episodes per system. We evaluate two orchestration paradigms: text subgoals for learned low-level policies (VoLo) and tool selection orchestration (CaP-X, GPT-Policy, PSP-Agent, TiPToP). Given the full instruction directly, low-level policies achieve 85.4% on single-subgoal tasks, but drop to near zero on extended tasks. While orchestration remedies this, it remains bounded by physical execution errors of the low-level policy rather than high-level planning and achieves only ~25%. Tool selection orchestration with GPT-Policy on GPT-6 Astra significantly outperforms learned-policy methods, but incurs significant time and cost budget and succeeds only ~62%. Real-world experiments across two embodiments confirm the collapse of the low-level policy and the gain from orchestration. Project website: https://ladder.insait.ai

This work investigates the application of the Membership Inference Test (MINT) to large language models (LLMs) fine-tuned with LoRA, aiming to determine whether specific samples were used during training and address concerns about data privacy, leakage, and transparency. Building on our previous LoRA-MINT framework, which relies exclusively on perplexity under a strict black-box setting, we propose an approach based on intermediate model representations, including activations and parameter-dependent signals. These representations enable a MINT classifier to distinguish training from non-training samples without requiring full access to model parameters. We evaluated the method on four LLMs (Qwen, Phi-4, DeepSeek-V3, and LLaMA 3.2-3B) fine-tuned on three datasets covering medical, mathematical, and academic reasoning tasks. Results show consistent performance across architectures and domains, achieving scores between 95% and 98%, with robustness to variations in model configurations and data distributions. The proposed approach enables efficient auditing of sensitive training data and supports data provenance and compliance with emerging privacy and AI governance regulations. Although our experiments focus on LoRA-based fine-tuning, the methodology can be extended to other LLM adaptation techniques and, more broadly, to other AI models.

High-fidelity three-dimensional turbulent flow fields are essential for scientific and engineering analyses, yet obtaining them through direct numerical simulation or high-resolution measurements remains expensive. Super-resolution reconstruction from coarse-grained data therefore offers a promising alternative, but remains fundamentally challenging because fine-scale structures are intrinsically three-dimensional and dynamically coupled across scales. Existing data-driven approaches still struggle to recover multiscale volumetric structures and their coherent temporal evolution. To address these limitations, we propose a turbulence-aware framework for time-resolved three-dimensional flow reconstruction. The framework couples a multiscale neural operator, designed to recover volumetric structures, with a state space model to promote temporal consistency in the reconstructed dynamics. In addition, Navier-Stokes residuals are incorporated as soft constraints to regularize the reconstruction toward physically plausible dynamics. A mixture-of-experts mechanism is also introduced to support unsupervised adaptation to unknown coarse-graining. Across evaluations on canonical turbulent systems and atmospheric reanalysis data, the proposed framework reduces reconstruction errors relative to learning-based and classical baselines, while better preserving multiscale spatial structure and coherent temporal evolution. These results validate the effectiveness of the proposed framework in reconstructing high-fidelity three-dimensional flow fields from coarse-grained data.

Tabular learning tasks often suffer from class imbalance, which is often addressed via weighting-based strategies. In this work, we compare the naive approach (unweighted) with the weighting-based rebalancing strategies. Somewhat surprisingly, we demonstrate that, asymptotically, as the predictor complexity increases with sample size, weighting strategies yield predictive performance identical to unweighted approaches across all evaluation metrics, up to a shift in the final threshold. To go beyond, we theoretically study the estimation of a linear regression function through weighted and unweighted CART trees with limited depths. In this setting, we prove that the points in the PR space for both strategies are distributed along the same curve, thus illustrating the similar nonasymptotic performance of the two strategies. Then, we numerically identify more complex synthetic settings where performance gaps, evaluated via fixed classifier complexity, favor weighting strategies. Thus, to guide practitioners, we introduce an indicator that aims at signaling whether a data set may benefit from weighting.

Although machine learning potentials (MLPs) generalize across configurational and chemical compound spaces, their accuracy is commonly believed to be limited by their training methods, with improvements beyond this limit achieved so far only by refitting to experimental observables. In this Perspective, we argue that MLPs can also be made more accurate than their training methods non-empirically by generalizing across a third space, that of quantum-chemical methods. We propose routes to method-transferable MLPs that predict at unseen methods, with working implementations in both density-functional and wavefunction-based theories showing that they can be more accurate than any of their training methods. Aiming beyond the accuracy of the training methods could also make universal MLPs more accurate than single-method models, even when the method-transferable MLPs remain less accurate than those QM methods. Method-transferable MLPs will shape the construction of training datasets and the evaluation of models, pushing the accuracy limits of chemical simulations.

Some observation-window-based KV cache compressors use prompt attention to identify important tokens but allocate uniform or predefined budgets across heads, overlooking input-dependent differences in budget demand. We propose ChampionKV, an input-adaptive head-wise budget allocator for long-context KV cache compression. It uses the maximum aggregated observation-window attention score as a lightweight empirical proxy for attention concentration, assigning larger budgets to heads with dispersed attention and smaller budgets to concentrated heads. An abnormal-head mask and a floor-budget mechanism improve robustness under weak attention responses. ChampionKV plugs into compressors such as SnapKV and PyramidKV without changing their token-selection procedures. On LongBench and RULER, both Champion variants consistently improve over their SnapKV and PyramidKV base compressors; the improvements are largest and statistically significant on Llama3.1-70B-Instruct, especially under tight budgets, whereas results on the tested 7B/8B models remain competitive but mixed. ChampionKV does not uniformly outperform Ada-KV, whose global cross-head Top-K selection retains an advantage in several settings. Controlled analysis further shows improved retained attention mass over the corresponding base compressors at negligible overhead.

Test-time scaling has enabled LLMs to solve increasingly complex reasoning, discovery, and optimization tasks. However, existing approaches often gather observations while seeking evidence for current hypotheses, then generate new hypotheses conditioned on those hypotheses and their associated evidence. This couples observation discovery to existing hypotheses, favoring local refinements over exploration of unknown hypotheses regions. Inspired by the hypothetico-deductive method (Hempel, 1945) and dual-space models of scientific discovery (Klahr & Dunbar, 1988), we introduce SHELDON, a scientific reasoning harness that separately represents and jointly explores hypothesis and observation spaces. By actively seeking new observations along with new hypotheses, SHELDON can discover observations that its current hypotheses do not explain. These unexplained observations expose gaps in its understanding, motivate new investigations, and guide the generation of hypotheses beyond refinements of existing ones. Across diagnosis, abstract reasoning, and optimization benchmarks, SHELDON outperforms evaluated baselines by up to 26.7 points. Additional analysis shows that SHELDON generates up to 4.5x more novel solutions.

Graph fraud explanation generation emerges as an important task, which requires jointly leveraging graph structure and content information to produce human-readable explanations for a given risk label. An intuitive approach is to adapt existing graph-enhanced large language models (LLMs), which inject graph structure into LLMs via graph-conditioned instruction tuning or structure-aware embeddings. However, existing methods face two key challenges in graph fraud explanation generation: a lack of evaluation protocols tailored to explanation quality, and a mismatch between the granularity of graph information they provide and what explanation generation requires. In this paper, we propose FRaud Explanation Generation via Agentic Graph Context NavigaTION (FRACTION), a framework for graph fraud explanation generation. To evaluate explanation quality, we introduce a task-specific rubric that assesses faithfulness, graph grounding, helpfulness, consistency, and conciseness, and validate it against human judgments. To focus on fine-grained details of relevant neighbors without reading the entire neighborhood, we equip a language model with graph tools that progressively disclose neighbor records. To train our graph agent, we propose a rubric-guided trajectory distillation strategy through which the agent learns both tool-use behavior and structured explanation generation. Experimental results on three fraud datasets demonstrate consistent improvements over state-of-the-art baselines under a human-aligned rubric.

On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher, mitigating the distribution mismatch of off-policy supervision. While prior work has primarily trained on problems a student can largely solve, we ask whether OPD can improve a student when training is restricted to problems it currently fails. We evaluate five representative OPD mechanisms with Qwen-3 and Phi-4 models under a unified competition-mathematics protocol. We find that OPD does not improve accuracy on this hard subset, with most mechanisms instead degrading performance despite continued decreases in training KL divergence. All eight configurations show a decline, and the two-sided sign test gives p=0.0078. In contrast, the same pipeline improves accuracy by 4.6 percentage points when training is not concentrated on failures, showing that the failure is conditional on the training regime rather than systematic. These findings suggest that OPD is effective when supervision is dominated by correct student prefixes, but can reinforce existing errors when training is concentrated on problems the student currently fails.

Diffusion language models (DLMs) enable flexible parallel generation by resolving tokens according to the model's evolving predictability rather than a fixed left-to-right order. This flexibility, however, introduces training choices absent from standard autoregressive teacher forcing: which reference tokens are visible when a target is supervised and how much generation space is allocated. Standard supervised fine-tuning (SFT) typically makes these choices through random masking, independently of inference. Our analysis shows how this training-inference mismatch can misdirect supervision. Random masking can conceal prediction errors or penalize uncertainty that decoding would otherwise resolve, while the attention-visible input length can change which response formulations the model favors, even when the underlying solution is unchanged. Motivated by these findings, we propose Predictability-Aware Canvas Selection and Supervision (PACS), an SFT method. PACS uses the model's own predictability to select each example's attention-visible input length and supervises reference tokens immediately before their confidence-guided revelation. Across reasoning and code-generation benchmarks, PACS consistently achieves the highest mean accuracy among the compared SFT methods, with gains of up to 12.84 percentage points over vanilla SFT on individual benchmarks. Code will be released upon publication.

Spiking Neural Networks (SNNs) are promising for energy-efficient neuromorphic computing, but they still suffer from catastrophic forgetting when learning tasks sequentially. This work investigates the integration of Experience Packing and Replay (EPR) as a memory-efficient continual learning strategy for deep SNNs under low-latency constraints. The proposed framework combines Batch Normalization Through Time stabilized VGG9 SNNs with Spike Activation Maps (SAM), which provide forward-only explanations based on spike activity. These explanations are temporally aggregated into spatial attention maps and used to select the most informative image patches for replay, instead of storing full images. Experiments on Split CIFAR10 and Split CIFAR100 show that SAM-guided EPR improves the accuracy-memory trade-off under tight memory budgets, and remains competitive while requiring substantially less storage.

Symbolic regression is the problem of finding an expression that fits a dataset. Solving it requires searching a combinatorial space of candidate expressions. Neural symbolic regression can amortize that search by training a model once on synthetic data and then predicting expressions for a new dataset. But existing models still underperform genetic programming on real, noisy data. We argue that the deficit lies in the training rather than in the paradigm, and present RASR (Reinforced Amortized Symbolic Regression), a method with two key changes to the training recipe. First, we design a synthetic prior that narrows the gap with real-world data: inputs come from the prior of a tabular foundation model; constants are sampled bottom-up, conditioned on the data; targets carry heteroscedastic and structured noise; and everything is standardized. Second, we split training into two stages. Supervised pre-training maximizes the token-level likelihood of a reference expression. This is followed by reinforcement-learning post-training that aligns the model with its inference-time use by rewarding held-out fit and penalizing expression size. On the black-box track of SRBench 2.0, RASR achieves the best mean test $R^2$ rank of all evaluated methods. Three RASR variants, trained with different size penalties, make up most of the accuracy-complexity Pareto front. On Strogatz, our method recovers more ground-truth expressions than any baseline, and on Feynman it is competitive with the best-performing methods.

Certified machine unlearning ensures the proximity between the unlearned model and a model retrained without the data to be forgotten. Extending this framework to sequential deletion requests, typically arriving as groups of variable size (e.g., all records of one user or of coordinated users), is challenging. For unlearning methods based on fine-tuning on the retain sets, this leads to either degrading performance or increasing computation as the number of unlearning requests grows. For independent successive requests, we instead certify each model release against a counterfactual unlearning trajectory in which the group being deleted never took part. With a fixed fine-tuning budget per deletion, we show that, for smooth strongly convex objectives, this approach keeps the performance loss bounded, uniformly in the number of requests. Experiments with multiclass logistic regression on FEMNIST show that the resulting algorithm needs fewer corrective steps than existing sequential methods and outperforms training-time differential privacy, particularly when group sizes are heterogeneous.

Plurality-voting ensembles aggregate the predictions of multiple base classifiers and output the most voted class, with ties typically broken at random. Despite the simple aggregation rule, attributing the impact of each base classifier to the ensemble's predictions is far from straightforward. That is, a single base classifier's impact depends on how it interacts with the votes of all other classifiers in the ensemble, for example, in resolving ties, and not only on its own predictions. We address the attribution problem using established tools from cooperative game theory, studying the Banzhaf and Shapley power indices for new games modeling plurality voting and random tie-breaking, for binary and multiclass classification settings. Computing power indices is, in general, intractable. Surprisingly, our methods, collectively named ENPOWER, enable the exact computation of Banzhaf and Shapley values in linear time for the binary classification setting and in low-degree polynomial time for the multiclass setting. The algorithms in ENPOWER rely on new combinatorial formulations of the Banzhaf and Shapley indices, yielding closed-form solutions for the binary case and ordinary generating function representations for the multiclass case. We validate empirically the methods in ENPOWER demonstrating their high applicability, providing attribution values not captured by existing methods. We apply the attribution of ENPOWER to: ensemble pruning and prompt-based ensembling of vision-language models.

We introduce STRUCTURES-P, a machine-learned all-electron orbital-free density functional for periodic CHNO systems. Electrostatics are evaluated analytically in an atom-centered Gaussian basis, while an equivariant network learns the kinetic and exchange-correlation energy from energies and gradients of non-ground-state densities. Minimizing the functional from a superposition of atomic densities recovers the Kohn-Sham ground state: on held-out crystals, the density error falls from 6.7% to 0.4% of the electron count and the energy error reaches 0.25 mHa per atom, half that of a same-architecture interatomic potential. A single functional can be trained for both isolated molecules and periodic systems, and the model generalizes to larger unit cells. It largely preserves relative-energy rankings and reproduces convex-hull vertices on the tested H-N composition slice. Together, these results demonstrate the viability of machine-learned all-electron orbital-free DFT for variational density optimization and energy evaluation in fixed-geometry solids.

The semantic entropy of a language model is the entropy of the meanings of its sampled answers. The meanings are never observed, and the entropy is estimated from the verdicts of a pairwise equivalence judge. Given the meanings of $n$ answers, a known channel declares each pair equivalent independently, with probability $q+\Delta$ within a meaning and $q$ across meanings. At most $K$ meanings occur, and $q$ is bounded away from $0$ and $1$. For a weak judge, with $n\Delta^2\to0$, the global minimax rate is set by distributions near a vertex of the simplex and overstates the difficulty elsewhere. We show that the difficulty is set by how many meanings can be rare. Over distributions with at most $s$ meanings below a fixed probability $\tau<1/(K-s)$, the minimax squared risk is of order $N^{-1/s}$ for $s\ge2$, $N^{-1}\log^2(e+N)$ for $s=1$ and $N^{-1}$ for $s=0$, capped at one, where $N=n^K\Delta^{2K-2}$. The matching lower bounds hold on every fixed polynomial window $n\Delta^2\le n^{-\kappa}$. One estimator attains these rates without knowing $s$ or $\tau$: it averages products of centered verdicts along all spanning trees of every $k$-set of answers, projects the resulting power sums onto the valid moment set and inverts them. The case $s=K-1$ gives the global rate $N^{-1/(K-1)}$ and the consistency threshold: with $\Delta=n^{-\gamma}$ the entropy is uniformly consistently estimable exactly when $\gamma<K/(2K-2)$. On polynomial windows, the same estimator attains the sharp local minimax constant on shrinking neighborhoods of interior points with distinct probabilities. Restricting the average to stars multiplies the constant by $K^{K-3}$.

Long-running agents retain KV caches across turns and sessions for hours or days. Compressing these persistent caches requires reducing storage while controlling reconstruction error. We present AlignPred, a codec that combines local alignment with predictive coding under a configurable KV reconstruction-error budget. Local orthogonal transforms align related heads or layers into a common coordinate frame, allowing AlignPred to combine token-wise prediction with same-token prediction across heads or layers. Residuals are quantized on an error-bounded lattice and entropy-coded, without learned predictors, encoder-side mode search, or large joint transforms. Across three attention mechanisms, AlignPred requires less storage than the evaluated baselines at comparable measured reconstruction error. Against the evaluated CPU codec implementations, it decodes at least 1.6 times faster while matching or improving encode time. AlignPred achieves near-native task scores at substantially compressed operating points on LongBench-Pro at 128K context. AlignPred's SGLang HiCache integration reduces persistent target-KV storage by 42-72% relative to native HiCache, with comparable latency.

Modern Transformers rely on many scale-control mechanisms, including normalization, residual scaling, attention temperatures, loss weighting, and adaptive updates. To explain how these mechanisms maintain scale consistency, we introduce Neural Dimensional Analysis (NDA), a unified framework based on dimensional consistency. We define neural units that specify how tensors rescale and which quantities can be added, just as physical units do in physics. Our rules match residual units, nondimensionalize attention and gate inputs, combine losses across units, and convert gradient units into parameter units. We implement NDA through dimensionless Unit Violation Indices (UVI) for diagnosing scale mismatches within Transformers, a unit-consistent multi-loss objective, and the Gauge-Canonical Transformer (GCT) with normalized coordinates and orthogonalized residual additions. In language-model experiments spanning 51M to 1B parameters, GCT remains stable during training across a broad sixteenfold learning-rate range, outperforming Pre-LN, LLaMA-style, and QK-Norm in 124M validation loss at tuned learning rates. GCT's normalized residual stream also shows improved robustness to quantization. At 1B parameters, calibrated per-tensor int8 quantization increases GCT's validation loss by only 0.009 nats, compared with up to 5.51 nats for the baselines. Vision and physics-informed neural network experiments further confirm NDA's benefits beyond language modeling. These results establish dimensional consistency as a neural network design principle.

Long-context inference makes the key-value (KV) cache a major memory bottleneck for large language models. We propose BELT, a low-rank KV cache compression framework. Its key-side objective targets attention distributions rather than raw keys or pre-softmax logits, yielding a quadratic surrogate with a closed-form low-rank solution. A learned global rank allocator uses differentiable soft selection during training and hard selection at deployment to meet a user-specified component budget. Experiments demonstrate improved model quality under aggressive compression and increased decoding throughput at comparable accuracy. Evaluated across three LLMs and multiple benchmarks, BELT achieves up to 70% low-rank KV cache compression while outperforming the strongest baseline by up to 23.3 points on LongBench, and up to 17.8x overall KV cache reduction when combined with quantization. With Triton-based GPU kernels and a quantized cache, BELT delivers up to 2.7x speedup for the attention module and 3.3x end-to-end decoding throughput.

Brick-and-mortar retailers face many different challenges that involve understanding thoroughly its products catalog and customer preferences. In particular, assortment optimization - proposing the ideal mix of products - and promotion planning hold a pivotal role in their strategy. By leveraging sales data, retailers can make informed decisions on which products to sell and how to manage inventory, based on customer preferences as well as regional and seasonal trends. It is especially crucial to capture interactions between products, in order to minimize the number of items that cannibalize each other's sales and to ensure that complementary products, which are often purchased together, are conjointly available and sold across all stores. In this paper, we propose a model of shopping basket that learns embeddings to represent interactions between products, prices and stores. Our model is built to uncover sales patterns from very large transaction datasets. In particular, the optimization loss is computed with random negative samples in order to overcome the computational bottlenecks that arise with large number of items. Our experiments on synthetic data show the efficiency of drawing such negative samples based on the actual assortment of available products, with better results than approaches from the literature. We also validate our approach by training and evaluating our model on a dataset composed of millions of transactions from Fortenova Group, one of the leading retail company in Southeast Europe. Our model showcases promising applications in the sector of retail, with enriched interfaces to efficiently support category managers.

Predictive service assurance converts forecasts of key performance indicators into alarms, and that conversion decides whether the assurance function is deployable. We present CALM-5G, an assurance layer that operates above any forecasting backbone and issues alarms with distribution-free, finite-sample guarantees. It calibrates on service-level violation events, scores violation onset directly, selects its calibration scheme out of sample, and reports the alarm duty cycle with every certified miss rate. Four corpora are evaluated separately, namely a production cellular network, commercial 5G traces, an open radio access network testbed and our 5G standalone testbed. Risk control on the regression-derived score yields duty cycles of 0.24 to 0.91. Onset targeting with cross-validated calibration reduces time-in-alarm by 56.3% to 89.7% on three corpora, with no significant change on the fourth. The realised miss rate stays within a two-standard-error tolerance on three corpora and within one standard error of it on the fourth. Under label shift, pooled batch calibration exceeds that tolerance structurally, by 10.9 standard errors on the production corpus. Online calibration restores compliance. Calibration schemes cannot be ranked on the split that fits their thresholds. Miss rate and duty cycle are certified jointly where within-entity dependence leaves sufficient calibration units.

We study parameter estimation when each sample from a continuous parametric distribution is accompanied by a binary quality label governed by the same unknown parameter. By imposing hard feasibility constraints, deterministic labels can accelerate estimation from the classical $n^{-1/2}$ rate of smooth quality-aware M-estimators to $n^{-1}$. Combining PAC guarantees with local error-to-parameter conversions, we establish this rate for both identifiable and finitely ambiguous quality classifiers, with likelihood information resolving the latter ambiguity. When the quality model is known only up to an additive constant, we construct threshold-free pairwise preferences from oppositely labeled samples and retain the accelerated rate despite the unknown threshold. We further show that pointwise qualities and pairwise preferences are incomparable: either can be strictly more informative than the other. Our assumptions hold for Gaussian, Laplace, and exponential families under thresholded log-likelihood rewards. Finally, a minimax lower bound establishes the optimal $n^{-1}$ dependence on sample size, up to logarithmic and dimension-dependent factors, while experiments confirm the benefits of constraint-based estimation.

Diffusion large language models (dLLMs) have attracted increasing research attention due to their bidirectional attention and arbitrary order decoding. However, there are two issues in dLLMs inference: the decoding difficulty is dominated by the tokens on the right side of the block, and the block-wise decoding inherently suffers from boundary contextual discontinuity. To tackle these limitations, we propose HaloBlock, a difficulty-aware adaptive halo mechanism for dLLM inference. We first leverage bidirectional cumulative entropy to characterize block decoding difficulty. Then, based on the bidirectional decoding difficulty, we adaptively allocate bidirectional halo regions on both sides of each block to mitigate boundary contextual discontinuity. The difficulty-aware design of HaloBlock smooths cross-block context propagation and lowers decoding difficulty for hard tokens, yielding better generation quality and faster convergence. Evaluated on LLaDA-8B-Instruct, Dream-7B-Instruct and LLaDA-1.5 across mathematical reasoning and code generation benchmarks, HaloBlock demonstrates consistent accuracy improvements and higher inference efficiency. Source codes are publicly available at https://github.com/PolyU-VCLab/HaloBlock.

Traditional time-series classification methods often operate as black boxes, providing a class prediction without indicating which time point or time interval within the series led to that prediction. To improve temporal interpretability, MILLET (Multiple Instance Learning with Temporal Explanation), which combines an InceptionTime encoder with Conjunctive Pooling, provides both classification and temporal explanations. However, MILLET contains approximately 424 K parameters, making it unsuitable for deployment on resource-constrained micro-controllers, and it achieves an overall time-series accuracy of 0.888 and an explanation score of 0.681 NDCG@n. We therefore propose SEA-Net, a lightweight architecture using small multi-scale depthwise-separable convolutional TSC encoders and improved Multiple Instance Learning (MIL) pooling heads. Our SEA-Net bottleneck encoder with Top-k pooling achieves 0.931 accuracy and 0.765 NDCG@n with only 41 K parameters. Another variant, using a gated encoder with Top-k pooling, achieves 0.955 accuracy and 0.750 NDCG@n with 62 K parameters. These results demonstrate that lightweight architectures can maintain strong classification performance and temporal explanations while substantially reducing model size. The proposed variants address specific limitations of the pooling strategy used by MILLET, providing a practical direction for interpretable time-series classification on resource-constrained edge devices.

Decisions that weigh several reward signals against each other require the joint distribution of the return vector. Categorical critics scale exponentially in the reward dimension, and quantiles have no canonical multivariate analogue. Multivariate distributional critics therefore represent each return distribution by M equal-weight particles trained with a discrepancy loss. Beyond their atoms, the particles provide no sampler and cannot resolve events rarer than 1/M. We propose Stein variational distributional RL (SVDRL). Its critic outputs Stein particles, which Stein variational gradient descent transports along a learned score field of the distributional Bellman target. We learn this field by bootstrapped denoising score matching across noise levels and tie it across the Bellman recursion by a consistency loss, which we prove is a non-expansive temporal-difference scheme in score space. The same field certifies the particles. Its reverse-diffusion sampler satisfies a supremal Wasserstein policy-evaluation bound that holds for bounded, possibly multimodal returns. The Stein particles inherit this bound up to a gap that we measure without ground truth, and the sampler answers tail-risk queries below the particle resolution at any sample size. Empirically, SVDRL is more accurate on rare events than particle critics with the same number of particles, and on par on MuJoCo continuous control.

Distributed learning with multiple local steps reduces communication but aggravates drift, and its practical performance often depends on carefully tuned hyperparameters under heterogeneous data and communication constraints. At the same time, recent Muon-type optimizers suggest that constraining updates within suitable norm balls provides a simple way to encode geometric priors into optimization. In this work, we propose NBCLT (Norm-Ball Clipped Local Training), a unified local-training framework for both federated learning (FL) and decentralized learning (DL). NBCLT combines a predict-and-correct local update rule with Euclidean projections onto flexible norm balls, thereby performing norm-ball clipping under flexible update geometries while retaining a simple implementation. In FL, its problem-parameter-free regime achieves $\mathcal{O}(\epsilon^{-2})$ communication complexity and $\mathcal{O}(\epsilon^{-4})$ total stochastic-gradient complexity without assuming bounded heterogeneity, when the per-round local-computation budget is scaled appropriately with the horizon. In DL, NBCLT is compatible with both static and time-varying mixing and preserves these accuracy exponents under the analogous computation scaling, up to fixed network constants and an additive initialization cost.