Papers

Deep generative models reproduce the observational distribution underlying the training data, which, in turn, inherit any spurious associations it contains. A common source is an unobserved confounder that simultaneously affects both a treatment attribute the user wants to control/edit (e.g., age, sex) and an observed outcome (e.g., an image) expected to vary in response. Existing causal inference methods resolve the resulting ambiguity by imposing structural assumptions strong enough to single out one interventional distribution. However, in practical domains like image generation, such assumptions are rarely warranted, and the interventional distribution is generally non-identifiable: distinct causal models can agree on the observed data yet imply different causal predictions. This paper proposes CauVaDE (Causal Variational Deep Embedding), a novel family of causal generative models in which the unobserved confounder collapses into a discrete latent cluster with bounded support without loss of generality. We prove that this family can approximate both observational and interventional image distributions with arbitrary accuracy in a canonical latent causal diagram. Building on this model representation, we introduce a novel causal bootstrap that varies the strength of confounder entropy to generate a diverse family of causal explanations, each defining a distinct interventional image distribution. Such a procedure produces distinct interventional images while maintaining a good fit to the observed data.

Despite their widespread use, the role of reward models in shaping reinforcement learning is poorly understood. Reward models offer a tempting promise: they automatically estimate response quality in the absence of verifiers or human judges. Unlike "verifiable rewards" which typically produce binary scores, reward models typically produce continuous scores, allowing them to be sensitive to fine-grained differences in responses. However, we show this apparent strength is a serious weakness: many popular reward models are oversensitive, assigning different scores to equally good responses. Theoretically, we show that seemingly perfect reward models can be highly oversensitive; empirically, this oversensitivity can lead to bad policies. In place of existing notions of "reward model accuracy," we propose evaluating reward models using distinct measures of "discriminative ability" and "specificity" (the complement of oversensitivity). As a solution, we describe a training-free algorithm that uses Monte Carlo dropout on any neural reward model to produce discrete reward clusters. Theoretically, we prove there exist discretizations that reduce oversensitivity at minimal expense of discriminative ability; empirically we show, in both controlled and natural RL settings, that discretizing rewards leads to less reward hacking and better policies than training on the original rewards.

We study what Lorentz-equivariant jet taggers learn internally, using equivariance tests, linear probes and grade ablations across five models including L-GATr, L-GATr-slim and LLoCa-T. Linear probes show that equivariant models suppress frame-dependent pseudorapidity to zero while encoding jet mass and N-subjettiness strongly. Grade ablations on L-GATr reveal that bivector channels are negligible for top-quark tagging while vector-like channels are dominant but seed variable, consistent with the network exploiting multiple representational pathways. These results characterize which physical features and algebraic grade structures carry discriminative information in equivariant taggers and may inform future development of such models.

Accurate 3D seismic velocity modeling through seismic travel-time tomography using both active- and passive-source data provides critical underpinning models for seismicity monitoring and hazard assessment. Because travel-time tomography is an inherently ill-posed inverse problem, UQ of the estimated models using Bayesian methods is also important for reliable downstream interpretations and analyses. However, Bayesian inference for 3D tomography based on conventional grid-based representations faces the “curse of dimensionality” and severe computational bottlenecks. Consequently, rigorous Bayesian UQ for margin-wide 3D travel-time tomography has remained largely unexplored. In this study, we propose a meshless 3D Bayesian travel-time tomography method that combines PINNs with a neural representation of the velocity structure, enabling tractable and data-efficient Bayesian inference through function-space particle-based variational inference. To efficiently integrate passive-source data into the Bayesian estimation of the velocity structure, we conduct analytical marginalization treating uncertain source parameters as nuisance parameters, with passive-source relocation carried out in post-processing. We validated the capability of our approach for 3D problems through synthetic experiments. Furthermore, we applied the method to a real-world dataset from marine active-source surveys and natural earthquakes off the Kii Peninsula, Nankai Trough. Our probabilistic 3D ensemble successfully resolves key geological features and provides data-consistent uncertainty maps. The posterior mean hypocenters shifted mainly in the vertical direction by 10-15 km, consistent with a previous relocation result. Finally, the neural representation drastically reduces storage requirements for the entire ensemble velocity model, highlighting the scalability and data efficiency of the proposed framework.

We introduce RocketPFN, a training-free pipeline for time series classification that combines random convolutional feature extraction (Rocket) with in-context classification via a pretrained tabular foundation model (TabPFN v2.5). On 92 UCR datasets (30-resample protocol), RocketPFN matches HC2, the strongest published method on the archive, in mean accuracy (both 0.900, Wilcoxon p=0.50), with no training on the target data and a median inference time of 30 seconds per fold. It also significantly outperforms every individual classifier in the HC2 ensemble. On UEA (20 datasets) the difference is likewise not statistically significant. A separate comparison concerns TSC foundation models: when paired with the same downstream classifier, MOMENT, Mantis, and MantisV2 are all significantly outperformed by RocketPFN using fewer extracted features and no learned parameters (p<0.001 in each case). This holds even when the encoders were pretrained on corpora that include the UCR training samples. We propose this two-stage pipeline as a reference point for evaluating zero-shot TSC foundation models.

High-resolution turbulence modeling is essential for scientific computing, but remains constrained by the cost of direct numerical simulation and the scarcity of full-resolution data. Existing scientific compressors reduce storage but typically operate on per-frame representations, whereas learned compressors yield compact latents that are often resolution-dependent and weakly aligned with the physics of turbulence. This raises the need for a compression framework that reduces data size, preserves physical diagnostics, and transfers from low-resolution training fields to high-resolution test fields without retraining. In this paper, we propose Physics-Preserving Latent Compression (PPLC), a patch-local latent compressor for three-dimensional turbulence. Motivated by inertial-range scale similarity, PPLC treats fixed-size patches as transferable units and applies a shared variational autoencoder independently of the global grid size. It combines exact mean preservation, zero-mean fluctuation encoding, an invertible Haar wavelet front-end, shift-consistency regularization, and overlap-aware reconstruction. Instantiated on forced isotropic turbulence, PPLC is trained only on stride-downsampled 256^3 fields and transfers zero-shot to 1024^3 fields. Experiments show that PPLC improves the balance between reconstruction accuracy and physical fidelity over classical and learned baselines, keeping diagnostics such as dissipation, enstrophy, energy spectra, and incompressibility closer to the ground truth. Beyond turbulence compression, PPLC offers a general strategy for physics-preserving latent representations that support data-efficient scientific surrogate modeling.

A Prior-data fitted Network learns the posterior predictive induced by its training prior; bringing this paradigm to multivariate time-series classification therefore calls for a synthetic generator that produces complete labelled datasets with temporal structure. We introduce a causal prior that synthesizes each dataset from a randomly sampled DAG over typed nodes across two modalities (tabular attributes and time series), natively producing multivariate, multi-class TSC datasets with cross-modal causal structure across channels, timesteps and labels, a regime not addressed by existing synthetic priors. To validate the prior, we finetune TabPFN v2.5 with minimal adaptations and evaluate on 75 UCR/UEA datasets within TabPFN's operating regime. Finetuning on our generator significantly outperforms both the unmodified upstream model and a tabular-only ablation of the same prior (Wilcoxon signed-rank $p=3.0\times 10^{-8}$ on ROC-AUC), isolating the contribution of the cross-modal temporal structure.

Recently, world models have emerged as a promising paradigm for building intelligent agents by learning predictive models that estimate future environment states conditioned on observations and actions. In particular, JEPA-style latent world models provide an efficient alternative to pixel space prediction by learning action-conditioned dynamics in compact representation spaces. However, existing latent world models typically rely on one-step prediction and must be recursively rolled out for long-horizon planning, which leads to compounding errors and a mismatch between training objectives and downstream planning tasks. To address this limitation, we propose Variable-length Latent World Models (VLWMs), a framework that learns to predict future latent states conditioned on action sequences of variable lengths. Instead of training only on one-step transitions, VLWMs directly model temporally extended dynamics, allowing the same predictor to evaluate action plans over different horizons. We further introduce a curriculum training strategy that progressively expands the action horizon, stabilizing optimization from short-range dynamics to long-range prediction. At test time, we design planning methods tailored to VLWMs to better exploit their variable-length predictive capabilities. Experiments on long-horizon control tasks show that VLWMs significantly improve latent space world models, achieving 13% average improvement over the state-of-the-art LeWM across different datasets, with especially large gains on tasks requiring extended planning. These results suggest that VLWM provides a simple yet effective paradigm for improving long-horizon prediction and planning in latent world models.

Plugging predictions of unknown parameters into downstream optimization problems, often referred to as the “predict-then-optimize” paradigm, has long been a standard approach in decision-making under uncertainty. However, improved predictive accuracy does not, in general, translate into improved decision quality. This disconnect has motivated growing interest in decision-focused learning (DFL) within the operations research community. This tutorial reviews recent developments in DFL and highlights key methodological insights, with a particular focus on stochastic linear programming as the downstream decision-making problem. We discuss why several widely used tools in traditional statistical learning are not directly suited to decision-focused settings and must be rethought, including (i) data collection strategies driven purely by predictive uncertainty and (ii) distributional distance measures such as the Wasserstein distance. We summarize properties of DFL that distinguish it from conventional predictive modeling and provide insights into the development of new decision-focused tools.

We present AdaPrivate-TS, a differentially private contextual bandit algorithm that combines Thompson Sampling with batched zCDP composition. Our key insight is that differential privacy noise inflates the posterior covariance in a structured way: adding Gaussian noise $N(0,\sigma^2 I)$ to $b$ yields sampling covariance $v^2 A^{-1} + \sigma^2 A^{-2}$, which Thompson Sampling interprets as increased uncertainty rather than pure corruption. Under event-level privacy (protecting individual interactions) with stochastic contexts, we prove that the privacy cost is only $O(\sqrt{d}\,\log T/\sqrt{\rho})$, logarithmic in $T$, because parallel composition amortizes noise across batches. Additionally, we explore privacy amplification via Poisson subsampling, which can reduce effective noise at stringent privacy budgets. Experiments on synthetic and real-world datasets demonstrate: (1) AdaPrivate-TS achieves 93-99% of non-private performance at $\varepsilon \in [0.5, 5]$, outperforming UCB by 0.5-3.7% and up to 18% with tuned adaptive exploration at extreme $\varepsilon$; (2) privacy amplification provides additional 2-5% gains at low $\varepsilon$; (3) on MovieLens and Jester, AdaPrivate-TS achieves the best overall performance among event-level baselines, dominating at $\varepsilon \geq 2$; (4) under DP-SVD private features, TS's advantage over UCB grows to +11%, confirming noise-as-uncertainty is not limited to reward privacy. We provide rigorous proofs for privacy guarantees under interactive zCDP composition and comprehensive evaluation including convergence curves, 12-seed CIs, and DP-SVD feature ablation.

Machine learning models are increasingly used to model chemical process systems, yet they often lack principled uncertainty quantification and mechanisms to enforce physical constraints. We propose a probabilistic neural network framework that guarantees satisfaction of linear equality constraints within a given tolerance, while capturing aleatoric uncertainty. Compared to state-of-the-art methods, our formulation demonstrates improved predictive accuracy, uncertainty calibration, and adherence to constraints on reduced data. It also demonstrates competitive performance, but with significantly faster training times when evaluated on large data regimes. We evaluated this on two batch reactor case studies, enforcing mass balances.

The Morse index of a free boundary minimal surface is encoded in its Jacobi-Steklov spectrum, and we test how faithfully a physics-informed neural network (PINN) reproduces that spectrum on a problem whose answer is already known in closed form. The benchmark is the critical catenoid in the unit ball $\mathbb{B}^3$, where it is well known that the Morse index equals $4$ and the nullity equals $2$. Separating the angular variable reduces the eigenvalue problem to a family of one-dimensional Robin problems on $[-T,T]$, one for each Fourier mode. A network that enforces the parity of each mode by construction, and carries the eigenvalue as a trainable parameter, returns the three eigenvalues below the stability threshold to within $10^{-6}$ to $10^{-4}$ of their exact values, with PDE residuals of order $10^{-4}$; assembling them recovers the index $4$ and the nullity $2$. We then track the spectrum along a one-parameter homotopy joining a flat reference operator to the catenoid Jacobi operator and identify the crossings at which the index changes. Since the critical catenoid is rigid, a fact we prove, this homotopy deforms operators rather than surfaces. We close by explaining how the same pipeline, with its one-dimensional solver replaced by a two-dimensional one, is poised to address genuinely geometric families in ellipsoidal balls, where the boundary curvature is no longer constant, and the Morse index is not yet known.

Detecting online polarization remains a critical challenge, particularly in multilingual and multicultural contexts where intergroup hostility is prevalent. The problem is particularly challenging due to the data scarcity for these tasks in the low-resource languages. Identifying such phenomena has become an active area of research and is addressed in SemEval-2026 Task 9: Multilingual, Multicultural Online Polarization Detection. To address this problem we propose an architecture that leverages LaBSE embeddings - an unconventional choice typically reserved for retrieval tasks, to obtain strong cross-lingual learning which enhances scores in low-resource language by a score up to 0.2 macro F1. Furthermore, we provide a comprehensive ablation study evaluating the performance of diverse encoder models in the Qwen model family within a retrieval-based prompting framework. Our code will be soon available at https://github.com/carrycurious/PolarMind.

Batch inference has become a central mode of AI computation, yet existing inference engines still rely on execution models designed for interactive serving. When scaled to millions of sequences, batch workloads reveal two fundamental requirements: the ability to handle extreme inter- and intra-sequence load variation that emerges only at runtime, and the ability to sustain high utilization across large fleets of GPUs. Existing systems fail to meet these requirements, losing substantial fractions of achievable throughput. We introduce a new architectural foundation for batch inference: the sequence coroutine compute model, which represents each sequence as a fine-grained, event-driven coroutine. This model exposes expressive primitives that allow the runtime to reorganize work dynamically, enabling larger expert-level batches, mitigating stragglers, reallocating work across devices, and maintaining utilization even on cost-effective or memory-constrained GPUs. Building on this abstraction, we implement BatchGen, a production-ready system that uses the coroutine model at cluster scale. On a 128-GPU cluster, BatchGen reduces batch completion time by up to $2.3\times$, and on memory-constrained accelerators it outperforms the strongest offloading baseline by up to $9.6\times$. We will open-source BatchGen at https://github.com/batchgen-project/batchgen

Long-term records of the Martian atmosphere based on general circulation models and reanalysis of atmospheric state variables are important to understand the diurnal, seasonal, and climatological changes of the planet. Atmospheric dynamics of the Martian atmosphere are strongly influenced by the characterization of dust lifting, solar insolation, and spatial variations in topography. We present ARCO-Mars, a unified Analysis-Ready Cloud-Optimized dataset providing integrated access to three independent Mars atmospheric reanalysis products: EMARS, MACDA, and OpenMARS spanning over Mars Years 24-35. These reanalyses assimilate thermal infrared retrievals from the MGS/TES, ODY/THEMIS, and MRO/MCS instruments, providing both two and three-dimensional surface and atmospheric state variables, including temperature, winds, surface pressure, and dust optical depth. The dataset is stored in Zarr v3 format and hosted on HuggingFace, enabling efficient cloud-based access without requiring local storage of the full archive. We compare the state variables between the three reanalysis products to identify systematic differences, attributed to differences in data assimilation and general circulation models. ARCO-Mars provides a community resource for Mars atmospheric science, numerical weather prediction validation, and machine learning applications, including weather forecasting and data assimilation.

Phishing is a multi-modal threat. We present a hybrid pipeline that scores each modality with its own engine and fuses the results. Three engines are built, deployed, and independently benchmarked: a four-stage URL stack (Domain Guard, lexical model, threat intelligence, and an asymmetric L2 fusion sidecar); a generalization-hardened DistilBERT NLP classifier whose held-out real-phishing recall rises from 0.8% to 87.3%; and a threat-intelligence synchronizer with end-to-end OpenTelemetry instrumentation confirming 1:1 message conservation. A decision-level fusion stage, characterized on a 10,677-email whole-system benchmark, reaches F1=0.914 with a calibrated probabilistic-OR over URL, header, and phishing-probability channels while reducing held-out real-spam false positives to 3.6%. Because that benchmark uses proxy URL and header channels and an operating point still needing recalibration, we present it as a preliminary integrated result. For deployable detection, the limiting factor is how well a model generalizes, not how accurately it scores data drawn from its own training distribution.

Piecewise-affine neural networks (e.g., with ReLU or LeakyReLU activations) implement continuous piecewise-affine maps, and the number of affine regions provides a natural proxy for expressive capacity. However, the gap between theoretical region capacity and the affine regions realized after training remains insufficiently understood. We study this gap from two complementary perspectives. First, we give a rigorous, architecture-dependent theorem for affine line-segment probes: for multilayer perceptrons with piecewise-affine activations, the number of affine pieces realized along an affine line-segment probe is upper bounded by an explicit product of layer-wise width terms (and activation breakpoint factors). This yields a neuron-threshold lower bound for representing target functions with prescribed one-dimensional piece complexity, formalizing the minimal region budget required for complex signals. Second, we exactly enumerate affine regions realized within bounded 2D and higher-dimensional domains under controlled task complexity. Under fixed architectures and training protocols, increasing input–label complexity yields trained solutions with markedly fewer realized regions in the evaluation domain, even though worst-case architectural capacity is unchanged; we call this reduced region usage expressivity saturation. Moreover, in the most challenging regimes, 2D visualizations show that region-usage collapse often coincides with degraded decision boundaries. Finally, we visualize the training dynamics of affine-region partitions and decision boundaries, revealing a consistent refinement process during optimization.

We consider the parameter estimation problem in logistic regression with Gaussian design: the estimation of a fixed unknown parameter $\theta^*\in \mathbb{R}^d$ ($\|\theta^*\|_2\ge 1$) from $n$ i.i.d. samples $\{(x_i,y_i)\}_{i=1}^n$, where $x_i\sim N(0,I_d)$ and $y_i|x_i \sim {\rm Bernoulli}(1/(1+\exp(-x_i^\top \theta^*)))$. Our main aim is to characterize the finite-sample estimation performance and convergence behavior of gradient descent (GD) on the maximum likelihood objective (i.e., the logistic loss). Under small $O(1)$ stepsize and $0$ initialization, we show that GD linearly converges to a small neighborhood of $\theta^*$ achieving an $\ell_2$ error of order $O(\sqrt{\|\theta^*\|_2^5d/n})$. This substantially goes beyond existing theoretical results that lack non-asymptotic estimation error rate and exhibit much slower parameter convergence. We also establish a faster local linear convergence to the same statistical error under a large $\Theta(\|\theta^*\|_2)$ stepsize. The main technical component is to show that the gradient of the logistic loss satisfies a certain approximate invertibility condition (AIC). To that end, we uniformly control the deviation of the gradient from its population counterpart by covering and peeling arguments, and then show that the population GD is a contraction by a delicate analysis based on the eigenvalues of population Hessian matrices. Finally, we build upon the recent work Matsumoto and Mazumdar (2025) and devise a novel efficient estimator that attains a sharper rate in high dimensions. This indicates that the existing non-asymptotic guarantees exhibit sub-optimal dependence on $\|\theta^*\|_2$, and that in many regimes $\Theta(\sqrt{\|\theta^*\|_2d/n})$ is the tight estimation error rate. Numerical examples are provided to corroborate our theoretical results.

Language models can generate plausible rationales for their predictions, but these explanations may not faithfully represent the model's internal reasoning. We propose verifier-coupled reasoning, a framework that inserts inline claims into reasoning traces and trains an auxiliary consistency head to predict programmatic verifier outputs from rationale-span hidden states. The central finding is a gap between decodability and faithfulness: consistency training reliably makes verifier information decodable from rationale representations, but decodability does not guarantee faithful generation. In LeanCheck (formal theorem proving), rationale-only and proof-only pooling achieve perfect directional separation under counterfactual conflict. In KataGo (Go engine), commentary spans encode 10-way win-rate buckets at 81% accuracy. Yet in a code setting, the model achieves 98.6% coupling while its generated explanations remain unfaithful: fluent prose with correct structured claims, but describing unrelated algorithms; a controlled pretrained-vs-from-scratch comparison shows the gap is not capacity-driven. Synthetic activation patching confirms causal influence (73-89% vs. 31% baseline), FEVER reveals that evidence-only pooling isolates genuine evidence sensitivity at the cost of raw accuracy, and per-claim analysis shows that consistency loss disproportionately benefits fine-grained claims over binary ones. These results establish that consistency losses are effective diagnostics and representation-shaping tools, but not sufficient conditions for faithful reasoning.

Imitation learning has emerged as a powerful paradigm for learning visuomotor policies, but its generalisation and stability are limited by the scale and quality of demonstration data needed. A promising direction is to leverage more abundant but heterogeneous data sources, which differ in action space and often lack action labels altogether. Existing co-training approaches that combine heterogeneous data sources rely on heuristic and hand-engineered alignment techniques. In contrast, we argue that action representations should be grounded in prediction: actions that produce the same effect on the environment should share the same representation, regardless of their sources. To this end, we instantiate this principle by using a grounded latent-action world model (GLAM), a pair of generative models with a shared latent action space across data sources that is grounded by predicting future observations consistently across sources. This latent action space is used to train downstream behavioural cloning (BC) policies which map observations to latent actions and decode them back to robot actions, providing a paradigm for learning from heterogeneous data. Empirically, we demonstrate that GLAM successfully learns an aligned latent action space that facilitates action transfer across data sources with and without action labels. Across five manipulation tasks in simulation and in the real world, GLAM-aligned policies significantly outperform BC baselines and prior latent-action methods, achieving an average of +48% improvement in task success rate with the same data-scarce setting. Videos and code are available at https://viccccciv.github.io/glam/.

We describe our entry to the efficiency track of the Academic Text-to-Music (ATTM) Grand Challenge at ICME 2026. Beyond the challenge protocol's FAD-CLAP and CLAP score, we add a learned human-preference reward from TuneJury, a twin pairwise ranker trained over open music-preference datasets. The reward serves both as a training-time conditioning signal and as a sample-selection criterion. The pipeline combines five engineering decisions on a 120M-parameter FluxAudio-S backbone, four at training time and one at inference: (i) training-time reward conditioning that doubles as an inference-time CFG axis, (ii) a sweep over five score-conditioning architectures, where training and inference use different variants, (iii) expert iteration on the top decile, (iv) a short preference-tuning pass (CRPO) for audio-text alignment, and (v) inference post-processing via joint CFG, source separation, and loudness normalization. Per-stage decomposition on 100 Song Describer prompts shows training-time reward conditioning as a functional conditioning axis, expert iteration as the dominant contributor, the preference-tuning pass adding only noise-level gain, and the inference-time score scalar already saturated by the end of the chain.

Continuous optimization methods for learning Directed Acyclic Graphs (DAGs) operate on weighted adjacency matrices and are therefore limited to pairwise causal relationships. We propose a framework for learning Directed Acyclic Hypergraphs (DAHGs) from observational data, capturing joint parental influences that pairwise models cannot represent. Our approach rests on three components: (i) a generalized linear structural equation model (SEM) with multiplicative interaction terms whose non-zero weights correspond one-to-one with directed hyperedges; (ii) a weighted adjacency tensor representation whose acyclicity is characterized via nilpotency under the tensor t-product; and (iii) a differentiable acyclicity constraint derived through the Fourier decomposition of the t-product, which reduces tensor nilpotency to slice-wise matrix nilpotency and enables least-squares learning via the augmented Lagrangian method.

Large language models (LLMs) can readily reproduce conventional expressions, yet their ability to model gradient frequency distributions remains underexplored. We investigate this using linguistic binomials, such as men and women, where both word permutations are grammatically valid but exhibit distinct, cross-linguistic variations in conventionality. We formalize binomial ordering as a distributional alignment problem, and construct a multilingual dataset of 600 binomial pairs across 8 languages. With categorical and distributional metrics, we measure and compare the corpus-derived preferences with model-induced ordering probabilities of 6 open-weight LLMs. While models often behaviorally recover the dominant corpus-preferred order, particularly for strongly conventionalized pairs, they align less well with the exact corpus preference distributions. This suggests that apparent directional order overstates how faithfully LLMs capture the statistical nuances of language use. Sparse probing verifies that the concept of preference strength is partially encoded among middle-to-late layers, and steering along probe-derived directions alters model-induced ordering distributions, demonstrating that the statistical behavioral preference of LLMs can be mechanistically measured and manipulated via internal representations.

Large language models (LLMs) have been proposed as hyperparameter-optimization (HPO) advisors that "warm-start" search from prior knowledge, proposing strong configurations in very few evaluations. We test that claim under a budget-matched, multi-seed protocol on eight PMLB tabular benchmarks, comparing an LLM advisor (LLM-OptFlow) against four classical baselines (random search, Optuna-TPE, Gaussian-process Bayesian optimization, and successive halving) over one shared search space, with paired tests and bootstrap 95% CIs across 8 x 5 = 40 (task, seed) units. The finding is cautionary. The advisor's strong first point is not an LLM output at all: like prior LLM-HPO systems the loop is seeded with a fixed default configuration, evaluated before any model call, which alone reaches 88.7% mean best-CV, identical to within 0.01 pp across all seven advisor models tested. The LLM's own proposals add only +0.40 pp of cross-validation accuracy over that seed and nothing on held-out test (LLM-Default = -0.01 pp, p = 0.92). When the same seed is granted to classical search, the apparent lead collapses: against seeded random search it leads by +0.20 pp at 2 evaluations, is tied by 5, and is behind by 12 (-0.37 pp). Without the seed, classical search ties the advisor by 12 evaluations and beats it by 40 (+0.6 to +0.8 pp, p <= 1e-4). Two LLM-specific behaviors survive: a single-task exploration failure (vehicle), and a rule-based confidence filter that removes ~33% of wasted compute without changing accuracy. The recommendation is deflationary: on tabular HPO, seed classical search with a sensible default; an LLM advisor adds no measurable generalization benefit and is overtaken within a handful of evaluations. We release the harness and a script that reproduces every statistic.

Minimum Spanning Trees have been used in unsupervised learning, particularly in clustering tasks, due to their ability to recognize clusters by removing edges that are considered inconsistent in defining those clusters. This paper aims to study the use of Minimum Spanning Trees in supervised learning. Specifically, we propose a classification algorithm based on Minimum Spanning Trees. To improve its performance, we introduce a robust version of the method that is also computationally more efficient. We evaluate the effectiveness of our proposed method through an extensive simulation study. We also apply the proposed methodology to a real-world case study involving aircraft trajectories.

Block diffusion LLMs (block dLLMs) improve decode throughput by generating tokens in parallel, and are emerging as a next-generation alternative to autoregressive LLMs. At long contexts, however, the KV cache remains a major throughput bottleneck in both memory bandwidth and capacity. For autoregressive LLMs, dynamic sparse attention is often combined with KV offloading to alleviate both bottlenecks, keeping the full KV cache in CPU DRAM and fetching only a critical subset to the GPU. However, KV subset selection and retrieval over bandwidth-limited PCIe incur substantial offloading overheads that often stall the GPU. We present HERALD, the first dynamic sparse attention method for block dLLMs that supports KV offloading. HERALD exploits a property of block dLLMs: a critical KV subset selected once per block is reused across all of its denoising steps, so the next block's selection can be hidden behind the current block's denoising. HERALD realizes this with a two-stage selection algorithm co-designed with its execution system. While the GPU denoises the current block, the CPU speculatively searches the full KV cache for the next block with a single query merged from the current block's B queries, cutting the selection compute by Bx; the GPU then refines the candidates into the final top-K subset with the next block's real queries. At K=1024, HERALD achieves near-lossless accuracy relative to Dense on LongBench and reasoning tasks across four block dLLMs, and improves decode throughput by up to 3.70x over Dense GPU-only inference, 2.59x over MAGE, a state-of-the-art block-dLLM sparse attention method, and 8.38x over MAGE-Offload, its direct KV-offloading extension.

Data curation is a critical part of post-training pipelines for large language models, yet existing tools often treat ingestion, deduplication, synthetic generation, and quality filtering as separate stages. This fragmentation makes it difficult to audit pipeline decisions or understand why individual samples are rejected. CuratorKIT is an open-source Python library that covers this full lifecycle in a single configurable pipeline. The framework is composed of six source format readers and automatic schema detection, a pre-generation data hygiene layer for credentials, PII, and toxic content, eight LLM-powered generation tasks, three complementary quality gates with provenance-exact hallucination verification, structured adaptive recovery, and five training-ready export formats compatible with TRL, Unsloth, and AlignTune. Every pipeline decision is recorded in an append-only per-sample provenance chain, and rejected samples carry structured failure reasons rather than being silently discarded. CuratorKIT supports 100+ LLM providers through LiteLLM, exposes both a Python API and a YAML-driven CLI, and is designed for practitioners who need reproducible, auditable data pipelines at scale .

As agentic systems tackle increasingly complex multi-step tasks, evaluating their trajectories presents a major bottleneck - human annotation of a single trajectory on popular agentic benchmarks can take hours, making it difficult to scale evaluations for measuring performance or curating training data. This has driven widespread reliance on automated approaches such as LLM-as-a-judge (LLMJ) to critique agents at the process and outcome-levels at scale, however, the soundness of LLMJ critiques often goes unmeasured. Here, we introduce Counsel, the first public dataset of meta-evaluations for agentic tasks. Counsel consists of process-level critiques from open-weight LLMJs on two agent benchmarks: tau-bench (customer support agents) and DA-Code (coding agents), and human meta-evaluations of these critiques. Human annotators label critiques on each flagged error as "spot on", "correct location but poor reasoning", or "should not have flagged", achieving reliable inter-annotator agreement (Krippendorff's alpha of 0.78). The resulting dataset stratifies LLMJ critiques by human alignment across both error location within a trajectory and reasoning quality, serving as valuable data to calibrate, improve, or train LLMJs for agents. Comparing open-weight judges, we find that more capable judge models and more reasoning effort both enabled improved human agreement, with the strongest judge reaching ~88% agreement on location and ~65% on reasoning. Counsel is generated using open-weight models and is permissively licensed for broad community use, which we hope will enable rigorous study and improved alignment of LLM-based evaluators for agentic systems.

Person-level psychological assessment requires aggregating meaning across many messages from the same individual, a task that document-level training objectives were not explicitly designed for. We present a systematic, empirical comparison between architecturally matched traditional (a) base-transformers and (b) document-tuned-transformers (further contrastively fine-tuned at the document-level, sometimes referred to as "sentence transformers") under otherwise identical conditions. Comparing layer-wise and overall performance across two longitudinal mental health and psychological datasets, we find document-tuned models demonstrated a consistent improvement over base representations (increase in Pearson r of 13.4%, p=.015). Robustness analyses revealed document-tuned models remained more accurate under perturbations to word deletion, synonym replacement, typo injection, and back translation. Further, hedged language (e.g., `usually') was more characteristic of outcomes in document-tuned embeddings while abundance (e.g., `lot') was more characteristic of base-transformers, suggesting document-tuned models may better capture uncertainty. These results suggest representation choice impacts mental health prediction, document-tuned models often being more adept.

Large Language Models (LLMs) have demonstrated strong capabilities in code generation, but their outputs frequently contain syntax or type errors that result in compilation failures. Constrained decoding has been proposed as a solution to mitigate compilation errors by construction, improving functional correctness as a byproduct. However, previous works overlook a critical aspect of constrained decoding: the alignment between constrainer (e.g., types), language model and the target specification language (e.g., TypeScript). Misalignment is caused by the constrainer being incomplete–rejecting programs that belong to the target–or unsound–allowing programs that are not part of the target. The bias created by incompleteness distorts the language model distribution, and can be detrimental for code generation. We evaluate this hypothesis using seven language models, two target languages, two constrainers, enforcing types and syntax during decoding, and we study how language models react to varying levels of incompleteness. On three benchmarks, when the constrainer is incomplete, unconstrained decoding significantly outperforms constrained decoding in terms of functional correctness. Incompleteness pushes the model into low-probability regions of the program space, causing the generation to frequently time out, and reducing functional correctness by up to 97%. These contributions make the community aware of the negative effects of misalignment in constrained decoding, and provide quantitative insights on how to design constrainers that are beneficial for code generation systems with formal guarantees.