Research Papers

arXiv:2605.27676v1 Announce Type: cross Abstract: Fine-tuning a pretrained language model on a curated dataset can produce spurious correlations between the fine-tuning task and unintended latent factors -- such as misaligned personas or political slant -- that the curation procedure has entangled with the task. The model can latch onto these spurious correlations, leading to bias and reduced out-of-distribution generalisation. We prove that under reasonable assumptions on task complexity and the spurious correlation, such latent factors can be identified, without supervision, from the weights of a naive LoRA fine-tune. Existing approaches to removing bias, such as activation steering, remove identified factors from residual-stream activations, either at inference or during training. We argue, however, that the goal should be to remove the spurious correlation, not the latent factor itself, as the pretrained model may rely on it for genuine task signal. To enable this, we propose GRASP

arXiv:2605.27690v1 Announce Type: cross Abstract: LLM agents increasingly operate through multi-turn tool use and environment interaction, where safety risks often emerge from intermediate steps long before they surface in the final outcome. Reactive auditing is therefore insufficient: post-hoc diagnosis frequently misses the chance to flag risks while they are unfolding. We propose TRACES, a representation-based proactive auditor that learns prefix-level trajectory risk states from the hidden representations of an observer LLM. TRACES induces latent mechanism features from step representations and models their temporal evolution to estimate whether a partial trajectory is drifting toward unsafe behavior. To sidestep the cost and ambiguity of step-level risk annotation, TRACES is trained with weak trajectory-level supervision while still producing dense prefix-level risk estimates. Across multiple agent safety benchmarks, TRACES improves both full-trajectory safety prediction and proac

arXiv:2605.27725v1 Announce Type: cross Abstract: High-fidelity computational fluid dynamics (CFD) is widely used for thermal-fluid design, but repeated CFD solves remain expensive for design optimization, uncertainty analysis, and digital-twin workflows. Recently, our team has demonstrated that a proper orthogonal decomposition and neural-network (POD-NN) surrogate can predict two-dimensional thermal fields in an electronics-cooling cold plate with large inference speedups while preserving physically interpretable modal structure. Reproducing and extending such workflows, however, typically requires custom scripts for parameter sampling, Fluent automation, data extraction, reduced-order model construction, neural-network training, validation, and prediction. This paper introduces CFDTwin, an open-source Python package and optional desktop graphical user interface (GUI) that packages these steps into a reusable workflow for ANSYS Fluent simulations. CFDTwin allows users to define simul

arXiv:2605.27747v1 Announce Type: cross Abstract: Existing training approaches for large language models learn a single set of parameters, based on large volumes of data, which is typically heterogeneous, conflicting and often outright contradictory. As a result, the model is forced to compress conflicting goals, and inherent uncertainties into a single, averaged pattern of behaviour. We propose an $\alpha$-R\'{e}nyi variational framework for learning distributions over post-training parameters, offering an uncertainty-aware alternative to deep ensemble approaches. The resulting variational objective interpolates between classical variational Bayes and predictively oriented posterior learning, balancing between globally plausible individual models against systems of complementary specialists. We identify local stability criteria, demonstrating how model misspecification can make non-degenerate posterior spread locally favourable, manifesting contradictory or conflicting data as epistem

arXiv:2605.27756v1 Announce Type: cross Abstract: High-performance computing enables simulation of high-dimensional physical systems, but downstream analyses such as inverse problems and control remain computationally expensive, motivating model order reduction (MOR) to construct efficient low-dimensional surrogates. Proper Orthogonal Decomposition (POD), a widely adopted data-driven MOR method, projects dynamics onto linear subspaces spanned by the most energetic modes. However, POD struggles for problems with slowly decaying Kolmogorov \(n\)-widths, such as advection-dominated and turbulent flows, requiring many modes for accurate reconstruction. Moreover, energy-based selection can discard crucial low-energy modes needed to capture small-scale features. Recent nonlinear manifold methods using polynomial mappings with alternating or greedy mode selection achieve better reconstruction with fewer modes. However, these methods fix the nonlinear mapping form a priori, limiting expressivi

arXiv:2605.27769v1 Announce Type: cross Abstract: We study the query complexity of sampling from high-dimensional Gaussian distributions using gradient information. In the standard oracle model, exact gradients expose only matrix-vector products with the precision matrix, leading to polynomial approximation barriers and a characteristic \(\sqrt{\kappa}\) dependence on the condition number. We show that this barrier disappears when the sampler is allowed to query \emph{smoothed scores}, namely gradients of the logarithms of the Gaussian-convolved densities. For a Gaussian target with precision matrix \(\Lambda\), a smoothed-score query at noise level \(\tau\) gives access to the resolvent \((\Lambda+\tau^{-1}I)^{-1}\). Combining geometrically spaced noise levels with sinc-quadrature rational approximation, we obtain a sampler with $q=O\!\left(\bigl(\log\kappa+\log(e\sqrt d/\delta_{\rm TV})\bigr)\log(e\sqrt d/\delta_{\rm TV})\right)$ smoothed-score queries for total variation error \(\de

arXiv:2605.27772v1 Announce Type: cross Abstract: Audio large language models (Audio LLMs) demonstrate strong performance on speech understanding tasks, yet their ability to understand paralinguistic information remains limited. To systematically quantify this issue, we introduce VoxParadox, an adversarial benchmark with 2,000 verified examples, spanning 10 paralinguistic tasks, created with controlled speech synthesis to intentionally mismatch transcript claims and speaking style, enabling direct measurement of speech paralinguistic understanding. Evaluation of a diverse set of Audio LLMs reveals consistently low accuracy on acoustic ground truth and a strong tendency to follow language-implied (incorrect) answers. To understand the cause of this gap, we perform layer-wise probing and find that (i) paralinguistic cues can degrade in deeper encoder layers and at the encoder--LLM interface, and (ii) even when such cues are available in audio tokens, the language model frequently ignores

arXiv:2605.27825v1 Announce Type: cross Abstract: Membership inference attacks (MIAs) test whether a target data record belongs to a system's private data, and have become a standard tool to measure privacy leakage in machine learning systems. Prior work has primarily focused on training corpora or retrieval databases. However, MIAs against agent memory have received less attention, even though such memory can contain sensitive user-agent interactions, retrieved facts, and user preferences. Therefore, in this work, we focus on chat agent memory MIAs, where an adversary infers whether a candidate memory unit belongs to the chat agent's memory store. We propose Multi-Recall Memory MIA (MRMMIA), a unified attack that utilizes multiple recall probes to the agent to extract the membership signal across black-box, gray-box, and white-box settings. Our experiments demonstrate that MRMMIA consistently outperforms baselines. Our results expose the privacy risk in agents and provide an initial e

arXiv:2605.27912v1 Announce Type: cross Abstract: We study efficient differentially private algorithms for estimating monotone statistics, i.e., statistics that are monotone under the addition of new observations. The starting point for our investigation is subsample-and-aggregate: a classical paradigm that partitions the dataset into blocks, estimates the statistic on each block, and then privately aggregates the estimates.While practical and generically applicable, this approach is quite data-hungry. We improve upon this framework for the class of monotone statistics -- compared to subsample-and-aggregate, our algorithms save a factor of $t$ in sample complexity and pay a factor of $e^t$ in running time, where $t>0$ is a tunable parameter. We complement our results with a query-complexity lower bound, showing that our algorithms are essentially optimal for this task. As an application, we obtain improved results for private eigenvalue estimation, private loss estimation, and privatel

arXiv:2605.27929v1 Announce Type: cross Abstract: Active sensing links behavior and learning through an action-perception loop: actions determine the observations used to update internal predictive models of perception, which subsequently guide the next actions. Predictive-coding frameworks provide a natural way to model this process, since internal representations are continuously updated to predict future observations. Here, we ask how exploratory and exploitative behavioral strategies shape these internal predictive representations. We build an online learning agent in a tree-like maze with a controllable parameter regulating the balance between exploratory and exploitative regimes. The agent updates a predictive-coding-based perception model from experience generated by its own behavior. The model predicts both future maze states and reward probability, allowing the agent to select actions either by expected information gain during exploration or by predicted reward during exploita

arXiv:2605.27937v1 Announce Type: cross Abstract: Scintillation detectors with excellent timing resolution enable more precise localization of radiation sources in positron emission tomography, leading to substantial improvements in diagnostic capability for diseases such as cancer and dementia. At the extreme timing precision required for such applications at the picosecond scale, detector performance is governed by the microscopic dynamics of scintillation photons generated within the detector and their subsequent detection processes. However, detector signals have conventionally been treated only as collective responses of many photons due to structural constraints inherent to photodetectors. In this study, we overcome this fundamental limitation using deep learning, enabling direct access to the timing information of individual photons. The proposed method estimates photon-by-photon arrival times directly from detector waveforms without requiring any modification to the detector st

arXiv:2605.27942v1 Announce Type: cross Abstract: Principal component analysis (PCA) is traditionally implemented through a covariance or kernel matrix, leading-eigenvector extraction, and hard rank-$k$ projection. These steps can be computationally costly in high-dimensional and quantum-data settings, sensitive to small eigengaps, and unnecessary when downstream tasks only require principal-subspace scores. Such score-based objectives are important in applications such as anomaly detection, spectral-energy profiling, and other postselection tasks. To address these needs, we introduce a measurement-based soft PCA framework replacing the hard top-$k$ projector with an entropy-regularized Fermi--Dirac filter. This filter is the unique optimizer of an entropy-regularized variational formulation of PCA and converges to the classical PCA projector in the zero-temperature limit. This filter has a direct interpretation as a quantum measurement, which naturally suggests a quantum approach. For

arXiv:2605.27991v1 Announce Type: cross Abstract: Deep neural networks (DNNs) have achieved remarkable empirical success, yet their training dynamics remain understood mainly from optimization rather than statistical principles. Here we develop a statistical framework for DNN training in the over-parameterized regime by showing that the prediction induced by continuous-time neural tangent kernel (NTK) gradient flow is exactly equivalent to that from a classical random-effects model. In this framework, training time acts as a variance component, or equivalently an empirical Bayes covariance hyperparameter, governing the allocation of variation from noise to structured signal. This equivalence reveals an optimization-inference duality: the gradient-flow path is both an optimization trajectory and an empirical Bayes random-effects inference path. Conditional on training time, the network output is the posterior mean of the latent signal, and estimating training time by restricted maximum

arXiv:2605.28014v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) improves the reasoning performance of large language models (LLMs) by providing dense token-level supervision for on-policy rollouts. However, existing OPSD methods often yield limited gains on in-domain reasoning and generalize poorly to out-of-domain problems. We identify two key causes: conditioning the self-teacher on a verified solution encourages imitation of training-domain reference trajectories rather than error-specific correction, and applying distillation to the full response can overwrite valid reasoning prefixes and reinforce overfitting. We propose Reflective On-policy Self-Distillation (ROSD), a framework that turns reference-solution imitation into targeted reasoning correction through reflection-guided, error-localized distillation. For each rollout, ROSD uses a self-reflector to extract a corrective idea and locate the first erroneous span. The corrective idea guides the self-teacher

arXiv:2605.28153v1 Announce Type: cross Abstract: Accurate and timely weather forecasts are critical for high-impact decisions in modern society. Machine-learning-based weather prediction is emerging as an alternative for producing initial conditions, forecasts, and even both in end-to-end systems. These methods deliver predictions faster and often with higher skill than traditional numerical weather prediction (NWP). However, even end-to-end models typically rely on NWP-generated reanalyses for supervision, thereby inheriting the biases and resolution limitations of those NWPs, and limiting adaptation to settings where suitable reanalysis products are unavailable, infrequently updated, or expensive to produce. Here we introduce ObsCast, a regional system that generates both analysis and predictions, without using any NWP-derived data in either training or inference, while still achieving state-of-the-art performance in short-term high-resolution regional modeling. Over the contiguous