arXiv:2606.25308v1 Announce Type: new Abstract: Behavioral detectors provide valuable insights into learner motivation and self-regulation. Among these, delayed start, a new session-level detector, has shown great promise as a valid behavioral measure that generalizes well across systems. In this paper, we examine cross-subject predictive validity of delayed start behavior. Using iReady data from 711 grade 7 students, we find delayed starts during Math practice are predictive of standardized test performance in both Math ($\beta$=.07 SD, p=.02) and English ($\beta$=.10 SD, p=<.001). Additionally, using mixture modeling and sensitivity analyses, we use a data-driven strategy to operationalize the identification of delayed starters in practice. We identify two underlying sub-groups of interest: "early starters" (<5 minute average delay, 20% of students) and "chronic delayers" (>13 minutes average delay, 20% of students). Relative to students in neither sub-group, early starters experienced greater growth (Math $\beta$=.11 SD, p=.07; ELA $\beta$=.15 SD, p=.02), while chronic delayers had the opposite trends (Math $\beta$=-.13 SD, p=0.05; ELA $\beta$=-.11 SD, p=0.11). Session-level measures provide a new opportunity for content-independent detectors, adding a behavioral component to the traditional usage and progress based on student engagement with content. This work aims to bridge education research with classroom practice by developing interpretable measures that align with behavioral cues teachers already use during classwork sessions to monitor and support students.
Science Journals
arXiv:2606.25728v1 Announce Type: new Abstract: Plans to establish a sustained human presence on Mars have moved from speculative ambition toward concrete engineering programmes, making the biological consequences of settlement an increasingly practical question. A Mars colony would place a small, closed population in an environment combining chronic radiation, altered immunity, constrained medical autonomy, and engineered indoor air. Latent infections are especially important because clinically silent carriers may become sources of transmissible disease when host control deteriorates. In this study, we develop a stochastic host-radiation-pathogen-habitat model of latent tuberculosis reactivation in a Mars colony. The model links galactic cosmic radiation to immune competence, immune competence to latent-tuberculosis reactivation, and reactivation to airborne transmission in a closed habitat. We also formulate countermeasure allocation as a partially observable sequential decision problem in which isolation and medication are selected by fixed baselines or by a proximal policy optimization policy trained on an agent-based simulator. Our simulations show that active tuberculosis can emerge endogenously despite no initial infectious cases, and that risk is most sensitive to latent reservoir size, radiation-immune coupling and reactivation sensitivity. Adaptive control reduced infectious burden and mortality while limiting unnecessary intervention. This framework supports mission-specific stress testing of screening, monitoring, shielding and treatment strategies before launch.
arXiv:2606.25092v1 Announce Type: new Abstract: Sparse Mixture-of-Experts (MoE) models route each token to a few of many experts, inviting the hypothesis that experts form functional modules tied to capabilities or languages. We test this causally on Command A+, a frontier open-weights MoE (218B total / 25B active; 128 experts, 8 active, +1 shared). We build a routing-mass atlas, pre-register six family-to-axis hypotheses before any intervention, and ablate each family at inference time against a size-matched random-expert null, measuring whether it selectively breaks its own axis (worst off-target effect at most one third of on-target). Crucially, we test the same families under four metrics and a held-out, independent-corpus run with bootstrap confidence intervals. Our finding is cautionary: robust functional modularity is rare and measurement-dependent. Of six pre-registered families, only one, the Arabic-language family, is a clean selective module that survives an independent corpus and a conservative statistical bar (1/6; a more permissive pre-registered point rule admits 3/6, but that count is threshold-sensitive). Every other family has a real causal effect yet fails selectivity, and its apparent modularity flips with the measurement: with the corpus, the metric, and the statistical bar. A positive control on Qwen3-30B-A3B recovers its published disjoint structure, confirming the method detects modularity when present. The verdict reproduces on the un-quantized BF16 model, ruling out a 4-bit quantization artifact. We conclude that ablation-based modularity verdicts are not safe unless the corpus, metric, and statistical bar are controlled. We release the atlas and ablation data.
arXiv:2606.25389v1 Announce Type: new Abstract: Extracting skills from multi-agent offline dataset improves learning efficiency via sharing task-invariant coordination skills among tasks. In settings where tasks occur sequentially and the space of skills grows exponentially, existing approaches that rely on heuristically designed and fixed-sized skill libraries struggle to resolve the problem of distributional shift and interference, facing catastrophic forgetting and plasticity loss. To address this problem and endow agents with the ability to continually discover and reuse coordination skills in open-environment, we propose COMAD, a principled framework for Continual Offline Multi-agent Skill Discovery via Skill Partition and Reuse. We first discover skills from mixed multi-agent behavior data with an auto-encoder to transform coordination knowledge into reusable coordination skills. Then we construct a skill-augmented policy learning objective with multi-head architectures, explicitly guiding the advantage function with reusable skills identified via a density-based reusability estimator. Theoretical analysis shows our method approximates the optimum of a continual skill discovery problem. Empirical results across diverse MARL benchmarks show that COMAD continually expands its skill library to mitigate interference, achieving superior forward and backward transfer for task streams compared to multiple baselines.
arXiv:2602.02219v2 Announce Type: replace Abstract: Large language models are widely employed as evaluators, a paradigm commonly referred to as LLM-as-a-judge. Prior research has predominantly examined point-wise or pair-wise evaluation protocols; in contrast, our focus is on rubric-based evaluation, which has been attracting increasing attention owing to its utility for training models in domains where verification is otherwise difficult. In this work, we show that rubric-based evaluation implicitly resembles a multiple-choice setting and therefore exhibits position bias: LLMs tend to prefer score options that appear at specific positions within the rubric list. Through controlled experiments across multiple models and datasets, we demonstrate that this position bias is consistent. Its direction, however, is model-specific: some judges favor the first option, while others favor the last. We further identify a second, orthogonal axis of bias: when a prompt scores several criteria simultaneously, the ordering of the criteria itself shifts the resulting scores. We additionally explore permuting the order of the rubric options as a means of mitigating position bias, and find that although the bias can be attenuated, improvements in the correlation between model judgments and human annotations are obtained primarily for models that exhibit strong bias. Our results recast rubric-based LLM-as-a-judge as a multiple-choice problem with measurable, model-specific position bias, and we further confirm that only a small number of random order permutations are sufficient to reduce the error introduced by this bias for the majority of models.
arXiv:2606.25417v1 Announce Type: new Abstract: The performance of two machine learning (ML) atmosphere models - GraphCast and FourCastNetV2 - is evaluated in the context of sub-seasonal prediction, including their ability to represent key climate drivers of variability, namely the Madden-Julian Oscillation and the Southern Annular Mode. Model skill is assessed over both a 38-year hindcast period and a 2.5-year hindcast period. The longer period overlaps with the training windows of the ML models but provides a larger sample for robust evaluation, while the shorter period is independent of the ML model training period. This dual evaluation illustrates a compromise approach to the problem of insufficient independent data for evaluation of the models for sub-seasonal prediction. The ML models are compared against the Bureau of Meteorology's physics-based seasonal prediction system, ACCESS-S2, for the 38-year period, and a more recent physics-based coupled model for the shorter hindcast period. Across the two evaluation periods, both ML models have surprisingly good skill for sub-seasonal timescales, given they were designed for forecasting on medium range timescales. In general, the ML models are as skilful as the physical model ensemble mean at shorter lead times and comparable to the physical model ensemble members at longer lead times.
arXiv:2409.12650v2 Announce Type: replace-cross Abstract: Stochastic effects significantly influence the dynamics of traffic flows. Many dynamic traffic assignment (DTA) models attempt to capture these effects by prescribing a specific ratio that determines how flow splits across different routes based on the routes' costs. Other models take a game-theoretic perspective and describe the equilibria resulting from the individual traffic participants' decisions instead of prescribing flow splits, however they usually neglect stochastic effects. In this paper, we propose a new unifying framework for DTA that incorporates the interplay between the routing decisions of each single traffic participant, the potentially stochastic nature of predicting the future state of the network, and the physical flow dynamics. Our framework consists of an edge loading operator modeling the physical flow propagation and a routing operator modeling the routing behavior of traffic participants. The routing operator is assumed to be set-valued and, thus, capable to model complex (deterministic) equilibrium conditions as well as stochastic equilibrium conditions assuming that measurements for predicting traffic are noisy. As our main results, we derive several quite general equilibrium existence and uniqueness results which not only subsume known results from the literature but also lead to new results. Specifically, for the new stochastic prediction equilibrium, we show existence and uniqueness under natural assumptions on the probability distribution over the predictions.
arXiv:2606.25434v1 Announce Type: new Abstract: Early and scalable detection of mild cognitive impairment (MCI) remains an unresolved clinical challenge. Existing EEG-based screening approaches are constrained by handcrafted feature pipelines that discard neurophysiologically meaningful domain structure and deep learning classifiers that sacrifice interpretability for performance. No existing work unifies physiologically organized concept encoders, cross-concept interaction modeling, and nonlinear tabular classification in a sleep EEG-based MCI detection framework. This study proposes Concept-guided Polynomial-transformed Tabular learning using Kolmogorov-Arnold Network (CPTabKAN), which maps heterogeneous EEG-derived features into domain-informed concept representations, expands them via degree-2 polynomial transformation to expose first- and second-order interactions, and applies a Fourier-parameterized TabKAN classifier to learn nonlinear decision boundaries. CPTabKAN was evaluated on the Study of Osteoporotic Fractures cohort (372 subjects, overnight polysomnography), using 1,379 features organized into ten physiologically motivated concept groups. Under 10-fold cross-validation, CPTabKAN-Second Order achieved a weighted F1-score of 0.9038 (SD 0.034), outperforming GradientBoosting by 5.65 percentage points (t(9)=1.934,p=0.043, one-sided paired test), with advantages persisting under SMOTE-based balancing. Ablation analysis confirmed independent contributions from each component. Concept importance analysis revealed that power spectral density, multi-scale entropy, and Hjorth parameters dominated first-order weights, while cross-concept interactions involving Lempel-Ziv-Welch complexity, statistics, demographics, and slow oscillations exceeded all first-order scores. These results demonstrate that concept-structured, interaction-aware tabular learning surfaces physiologically coherent reasoning, supporting clinical trust.
arXiv:2502.18065v2 Announce Type: replace-cross Abstract: We introduce merge-width, a family of graph parameters that unifies several structural graph measures, including treewidth, degeneracy, twin-width, clique-width, and generalized coloring numbers. Our parameters are based on new decompositions called construction sequences. These are sequences of ever coarser partitions of the vertex set, where each pair of parts has a specified default connection, and all vertex pairs of the graph that differ from the default are marked as resolved. The radius-$r$ merge-width is the maximum number of parts reached from a vertex by following a path of at most $r$ resolved edges. Graph classes of bounded merge-width -- for which the radius-$r$ merge-width parameter can be bounded by a constant, for each fixed $r=1,2,3,\ldots$ -- include all classes of bounded expansion or of bounded twin-width, thus unifying two central notions from the Sparsity and Twin-width frameworks. Furthermore, they are preserved under first-order transductions, which attests to their robustness. We conjecture that classes of bounded merge-width are equivalent to the previously introduced classes of bounded flip-width. As our main result, we show that the model checking problem for first-order logic is fixed-parameter tractable on graph classes of bounded merge-width, assuming the input includes a witnessing construction sequence. This unites and extends two previous model checking results: the result of Dvo\v{r}\'{a}k, Kr\'{a}l, and Thomas for classes of bounded expansion, and the result of Bonnet, Kim, Thomass\'e, and Watrigant for classes of bounded twin-width. Finally, we suggest future research directions that could impact the study of structural and algorithmic graph theory, in particular of monadically dependent graph classes, which we conjecture to coincide with classes of almost bounded merge-width.
arXiv:2503.19172v2 Announce Type: replace-cross Abstract: Quantum devices can process data in a fundamentally different way than classical computers. To leverage this potential, many algorithms require the aid of a quantum Random Access Memory (QRAM), i.e. a module capable of efficiently loading datasets onto the quantum processor. However, a realisation of this building block is still outstanding due to its formidable resource requirements, which become even more demanding in quantum error-correction schemes. Here we show that the challenge of implementing QRAM can be entirely reduced to a state-preparation problem: since such resource-state is independent on the memory, our approach allows one to prepare it offline, opening the door to new design strategies. As an example, we introduce a heralded 'QRAM factory' which enables improved fidelities with high acceptance rate. More broadly, our results introduce the concept of resource-state QRAM: we study its performance in noisy settings, showing that it preserves the noise-resilience of standard QRAM, and discuss how it can be efficiently combined with quantum error-correction. Finally, we propose an implementation with neutral-atom hardware, where our analysis suggests that high-fidelity and low-latency queries can be implemented.
arXiv:2606.24895v1 Announce Type: new Abstract: This project report presents a hybrid AI-assisted workflow for extracting and reintegrating archival metadata from League of Nations index cards. The project is situated in the broader context of the Total Digital Access to the League of Nations Archives project (LONTAD). Rather than attempting full OCR of the underlying archival collections, the workflow targets the index cards themselves as documentary access points to files, series, archival descriptions, and digital objects. The project evolved from a layout-aware pipeline combining YOLO, TrOCR, and local LLM post-correction to a hybrid architecture using a fine-tuned vision-language model for broad extraction while retaining specialized OCR for file and series identifiers.
arXiv:2606.25277v1 Announce Type: new Abstract: To address data overload and inefficient shape-level annotation in robotic visual inspection, this paper proposes a hardware-software integrated optoelectronic architecture. A non-imaging, low-data paradigm is established to minimize annotation dependency. First, a sensor-in-the-loop strategy reconfigures a Digital Micromirror Device (DMD) as a physical optical convolutional layer, enabling photonic-domain feature extraction that unifies sensing hardware and processing software. To suppress data volume at the source, a block-based compressed sensing strategy encodes spatial information into low-dimensional temporal signals, drastically reducing redundancy. Subsequently, to bypass laborious manual defect shape annotation, natural language descriptions guide the network to align with highly generalizable features from Contrastive Language-Image Pre-training (CLIP), steering the attention maps of the optoelectronic neural network toward defect shapes. Furthermore, a Localization Accuracy for Attention (LAA) metric is proposed to quantify shape-level defect localization performance. Experiments on transparent material defect detection validate the system's effectiveness. Parametric analysis reveals how measurement matrices, compression ratios, and block sizes affect accuracy. Results show that, compared to traditional imaging, the proposed architecture maintains equivalent accuracy while reducing data volume by 90% for Vision Transformers and computational workload by 60% for Convolutional Neural Networks. This low-data paradigm offers an efficient solution for industrial automation scenarios involving massive data streams, high acquisition costs, or constrained edge resources.
arXiv:2606.25473v1 Announce Type: new Abstract: Autoregressive video diffusion with causal diffusion transformers has emerged as a major paradigm for real-time streaming video generation and action-conditioned interactive world models. In this work, we extend rCM, an advanced diffusion distillation framework, to autoregressive video diffusion. The core philosophy of rCM lies in the complementarity between forward and reverse divergences, represented by consistency models (CMs) and distribution matching distillation (DMD), respectively, in diffusion distillation. This philosophy naturally carries over to the autoregressive setting, where teacher-forcing (TF) provides an offline, forward-divergence causal training paradigm, while self-forcing (SF) corresponds to an on-policy, reverse-divergence refinement. Our contributions are: (1) through extensive experiments, we show that teacher-forcing CM is currently the best complement to self-forcing DMD as an initialization strategy (2) we present the first implementation of teacher-forcing-based continuous-time CMs (e.g., sCM/MeanFlow) for autoregressive video diffusion, enabled by our custom-mask FlashAttention-2 JVP kernel, achieving 10$\times$ faster convergence compared to discrete-time CMs (dCMs) (3) we introduce Causal-rCM, a leading, unified, and scalable algorithm-infrastructure open recipe for diffusion distillation and causal training (4) we achieve state-of-the-art streaming video generation performance in both frame-wise and chunk-wise settings, using only synthetic data for training. Notably, our distilled 2-step causal Wan2.1-1.3B model achieves a VBench-T2V score of 84.63 with only 1 or 2 sampling steps. We further apply Causal-rCM to Cosmos 3, an advanced omnimodal world foundation model for physical AI with action-conditioned generation capability, enabling an interactive world model.
arXiv:2512.02968v2 Announce Type: replace-cross Abstract: Gravitational-wave data analysis relies on accurate and efficient methods to extract physical information from noisy detector signals, yet the increasing rate and complexity of observations represent a growing challenge. Deep learning provides a powerful alternative to traditional inference, but existing neural models typically lack the flexibility to handle variations in data analysis settings. Such variations accommodate imperfect observations or are required for specialized tests, and could include changes in detector configurations, overall frequency ranges, or localized cuts. We introduce a flexible transformer-based architecture paired with a training strategy that enables adaptation to diverse analysis settings at inference time. Applied to parameter estimation, we demonstrate that a single flexible model, called Dingo-T1, can (i) analyze 48 gravitational-wave events from the third LIGO-Virgo-KAGRA Observing Run under a wide range of analysis configurations, (ii) enable systematic studies of how detector and frequency configurations impact inferred posteriors, and (iii) perform inspiral-merger-ringdown consistency tests probing general relativity. Dingo-T1 also improves median sample efficiency on real events from a baseline of 1.4% to 4.2%. Our approach thus demonstrates flexible and scalable inference with a principled framework for handling missing or incomplete data, key capabilities for current and next-generation observatories.
arXiv:2606.24970v1 Announce Type: new Abstract: Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the network, producing smaller models that retain most of their accuracy. As attention layers are the most resource-intensive parts of LLMs, pruning them is a promising compression strategy. Prior work shows that up to 33% of attention layers can be pruned with minimal accuracy loss. Nevertheless, the impact of attention pruning on model interpretability, specifically faithfulness and confidence calibration, remains unstudied. To address this gap, we study how pruning attention layers affects explanation faithfulness and confidence calibration across five LLMs and eight datasets. While the pruned models often maintain high accuracy, we find that their faithfulness and calibration often degrade. Notably, faithfulness and calibration can fluctuate significantly, even when accuracy remains stable, highlighting a misalignment between model confidence, interpretability, and accuracy. Our findings suggest that layer pruning can affect LLMs' interpretability and reliability in ways not captured by accuracy and efficiency measures alone. We recommend including explainability and calibration metrics when evaluating pruned models.
arXiv:2604.14452v3 Announce Type: replace Abstract: Intensity-intensity correlation of a scalar optical beam carrying orbital angular momentum (OAM) across multiple modes is decomposed into intermodal contributions, thereby linking it, within the framework of the Hanbury Brown-Twiss effect, to the underlying modal coherence structure. Upon filtering the spiral phase dependence, the intensity-intensity correlations are governed by OAM-mode coherence and anisotropy reflecting classical entanglement between OAM and statistical properties of light. These results extend intensity interferometry to structured light fields and provide direct access to modal coherence properties without phase-sensitive measurements.
arXiv:2606.24715v2 Announce Type: replace-cross Abstract: We study the problem of model selection among probabilistic forecasting models evaluated on datasets of multiple time series. The performance of a model on a single time series is quantified by the average value (score) of a proper scoring rule over a test set, but extending model selection to data sets of time series requires aggregating these scores. Common approaches either rely on scaling scores and averaging them (mean scaled score) or avoid scaling by using alternative statistics such as mean ranks or win rates. However, these approaches can yield conflicting conclusions. We show that such discrepancies arise from the skewness of the distribution of the scores, which is particularly pronounced when test sets are short. The skewness can cause non-mean criteria (e.g., mean rank, median, win rate) to select misspecified models. In contrast, the mean score is immune from this problem. We further show that, as the size of the test sets increases, all aggregation criteria converge to the same model selection decision, mitigating these discrepancies. Our experiments on intermittent demand time series, including data from the M5 competition, highlight the importance of sufficiently large test sets; the mean scaled score appears to be the more reliable approach, also because empirically we found its decision to remain consistent when different scaling factors are adopted.
arXiv:2606.25334v1 Announce Type: new Abstract: Post-discharge clinical follow-up is critical for maintaining continuity of care and mitigating long-term health risks. However, traditional follow-up paradigms suffer from shortage of health workforce, fragmented patient histories, and information silos across clinical departments. While large language models have demonstrated potential in medical question-answering, their deployment in continuous care is hindered by hallucination risks and a fundamental inability to reason over longitudinal, patient-specific constraints. Here we present Healink, a memory-enhanced multi-agent framework to support AI-assisted post-discharge follow-up by generating prescription-grounded, traceable responses that improved completeness and perceived clinical utility in retrospective and physician-blinded evaluations. The architecture seamlessly integrates a triage routing mechanism, a unified memory enhancement module utilizing a robust relational database for optimal latency, and a strict constraint-based retrieval-augmented generation engine. By vectorizing historical clinical records and employing weighted similarity functions across diverse phenotypic and intervention dimensions, Healink ensures precise inter-patient and intra-patient case matching while actively preventing cross-departmental drug conflicts. We evaluated Healink on a dataset comprising 400 continuous and 85 highly complex real-world follow-up cases, alongside the webMedQA benchmark. In a rigorous single-blind evaluation conducted by clinical experts, the framework outperformed human physician baselines in both authoritativeness and clinical safety. By generating a traceable, white-box evidence chain, Healink provides a scalable, safe, and highly effective paradigm for intelligent patient management, ultimately enhancing societal healthcare outcomes.
arXiv:2606.25323v1 Announce Type: cross Abstract: Physics informed neural networks (PINNs) incorporate Maxwell's equations, initial conditions, boundary conditions, and measurement data directly into the learning process, transforming the solution of partial differential equations into a constrained optimization problem. As such, PINNs are attracting increased attention in computational electromagnetics as an alternative to traditional time-domain solvers. This paper presents a PINN formulation for the time-domain Maxwell's equations incorporating split-field perfectly matched layers (PMLs). A key advantage of the split-field PML formulation is that the same governing equations can be applied in both the physical and PML regions, simplifying the PINN formulation and loss construction. The proposed approach is validated using one-dimensional and two-dimensional Gaussian pulse problems with PML. The PINN solutions show good agreement with analytical and finite-difference time-domain (FDTD) reference solutions, demonstrating the feasibility of combining PINNs with split-field PMLs for open-domain time-domain electromagnetic simulations.
arXiv:2606.25550v1 Announce Type: new Abstract: Empirical research in Requirements Engineering is hampered by a lack of adequate datasets that pair source code with corresponding requirements. A tempting route to addressing this lack is the use of Large Language Models to synthesize requirements from existing code bases. We investigate this question by evaluating an LLM-based and RAG-supported agentic approach that generates requirements from source code, verifies their implementation status relying on a human-in-the-loop, and synthetically introduces requirements smells and non-implemented requirements. Our goal was to create datasets that mimic reality and foster empirical RE research. However, during the study, various problems arose, leading to this experience report. Contrary to our initial hypotheses, LLMs were unable to (i) generate non-implemented requirements reliably, (ii) generate high quality requirements, and (iii) reliably introduce synthetic requirements smells. Furthermore, neither an LLM nor a single human-in-the-loop suffices to detect requirements smells reliably. These findings suggest that the generation of code-to-requirements datasets using LLMs is not yet viable and requires human supervision, especially for quality assurance. We critically reflect on our lessons learned and draw relevant conclusions for both researchers and practitioners.
arXiv:2606.25561v1 Announce Type: new Abstract: Manual formal analysis of cryptographic schemes is labor-intensive and requires substantial expertise. While model-checking tools (e.g., Scyther and Tamarin) and computational-security tools (e.g., CryptoVerif and EasyCrypt) improve the automation of security proofs, they still rely on experts to abstract schemes and write tool-specific formal descriptions. Large language models (LLMs) are a promising alternative, but their effectiveness in this domain remains unexplored due to the absence of standardized evaluation methodologies. To fill this gap, we introduce CrypFormBench (C.F.B for short), a comprehensive benchmark jointly covering symbolic and computational security to evaluate five core LLM capabilities: interpretation, generation, completion, transformation, and correction. It comprises 700 instances spanning 677 schemes, 7 mainstream formal verifier languages, and 160 security properties. The evaluation of 9 state-of-the-art LLMs reveals that most of them perform well on interpretation and completion, given their code-awareness advantages, but struggle with generation, transformation, and correction. Overall, their performance remains limited, with Claude-3.5 achieving the highest score at 48.7 out of 100. We further provide practical guidance, e.g., few-shot prompting, Pass@K sampling, and lightweight fine-tuning, to mitigate the executability bottleneck and improve tool-usable outputs. Taken together, our benchmark and analyses offer a grounded view of current progress and concrete directions toward reliable LLM-assisted formal cryptographic analysis.
arXiv:2605.31220v2 Announce Type: replace Abstract: Confidence estimation (CE), i.e., quantifying the reliability of a model's prediction, has attracted great interest in the context of large language models (LLMs). However, most studies focus on English, ignoring the multilingual reality of LLM usage, while many CE methods degrade or require retraining across languages. To address this gap, we investigate whether multilingual LLMs encode shared, language-transferable confidence features in open-ended question answering. We use a lightweight linear probe that predicts answer correctness directly from intermediate representations. Trained monolingually, the probe generalizes zero-shot to unseen, typologically diverse languages without target-language supervision. Learned layer weights and multiple ablations reveal that confidence features concentrate in middle layers across languages, suggesting a shared confidence subspace. While zero-shot cross-lingual performance depends on similarity to the source language, the probe provides a strong baseline without any retraining and compares favorably to other popular confidence estimation methods.
arXiv:2606.25010v1 Announce Type: new Abstract: Neural scaling laws for transformer language models predict smooth improvements in pretraining loss with increasing parameters, but downstream capabilities such as in-context learning are known to emerge abruptly past a certain model scale. In this paper, we show that emergent capabilities arise stochastically throughout training, with larger models acquiring them earlier on average. We demonstrate that the emergence of capabilities such as pattern completion and indirect object identification corresponds to the abrupt learning of task-relevant attention patterns. To isolate this phenomenon, we train transformer models on synthetic linear map and cellular automata datasets, and we show that the difficulty of learning attention patterns depends on context length and pattern sparsity. Moreover, scaling the number of attention heads improves learning efficiency on our synthetic tasks, while increasing the head dimension yields diminishing returns past a minimum capacity. We additionally investigate architectures with alternative attention mechanisms, showing that MLP-Mixer outperforms a transformer on linear map tasks with complex attention patterns. Our findings provide a mechanistic insight into emergence, showing that downstream capabilities arise abruptly due to the intrinsic difficulty of learning sparse attention patterns in transformer models.
arXiv:2606.25013v1 Announce Type: new Abstract: Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88\%$ balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first $\sim20\%$ of thinking, despite giving the appearance of deliberation at the text level ($\sim74\%$ of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.
arXiv:2606.25076v1 Announce Type: new Abstract: Following the success of machine learning in producing weather predictions with competitive skill compared to complex traditional systems, this article shifts attention from forecast output to the working practices that make prediction systems possible. We argue that machine learning and recent digital technologies will reshape the forecasting value chain: how models are coded and developed, how observations and Earth-system data are exploited, how data and computing are managed, how systems are verified, and how information is created, evaluated and turned into services. We discuss six non-exhaustive areas in which agentic software engineering, open and compressed data, shared verification workflows, interactive computing and generative methods may make modelling, evaluation and service creation faster, more interactive and more widely accessible. These changes will require weather and climate centres to adapt their infrastructures, data stewardship, trust and quality-assurance frameworks, skills and service delivery while maintaining scientific understanding, operational reliability, human expertise and their public-service role.