Papers

Autonomous large language model (LLM) agents increasingly face reliability, context consumption, and execution stability bottlenecks when deployed on complex, long-horizon tasks. While monolithic prompt engineering and stateless tool-calling paradigms struggle to scale, the field is rapidly converging toward agentic skills: modular procedural abstractions that externalize execution knowledge into reusable, executable, and portable artifacts. This paper establishes a unified systems foundation and reference architecture for the agentic skills ecosystem. We formalize skills as externalized procedural knowledge bridging high-level cognitive planning with deterministic execution environments, and systematically delineate the architecture across a nine-stage lifecycle: autonomous discovery, authoring and representation formats, memory storage, dynamic retrieval and routing, composition and orchestration, execution and repair, lifelong adaptation, empirical evaluation, and security governance. We further examine marketplace dynamics, public registries, and emerging adversarial threat vectors, alongside runtime verification and defense mechanisms. Finally, we categorize system implementations across software engineering, operating system navigation, embodied robotics, and scientific discovery, while highlighting critical open challenges in continual learning and benchmark realism. This work establishes agentic skills as a foundational paradigm for building scalable, robust, and verifiable autonomous language agents.

Chaotic time series forecasting is a challenging task due to its sensitivity to initial conditions and long-term unpredictability. Traditional methods typically rely on sufficient temporal trajectories to learn long-term dynamics, which limits their applicability when only short-term observations are available. While recent Large Language Models (LLMs) have shown great potential for time series forecasting, their temporal representations are not explicitly tailored to the phase-space structure and nonlinear evolution of chaotic systems. To address these issues, we propose PAC-LLM, a phase-space-aware adaptive fusion framework for long-term chaotic time series forecasting powered by LLMs. PAC-LLM leverages learned phase-space features and textual information to fully enable LLM's time series forecasting capacity. In particular, we design an auxiliary feature module and a gated weighting mechanism for multivariate coupling information fusion and selection. Extensive experiments on representative chaotic systems demonstrate that our method outperforms existing fine-tuned and zero-shot baselines in both short-term and long-term predictions. Our ablation study further confirms the effectiveness of each key component in PAC-LLM.

Online learning-based control is a promising approach to control uncertain systems, where unknown components are identified during operation to improve control performance. However, resource-intensive online learning algorithms introduce non-negligible computational delays, especially when executed on systems with limited local computational resources. To mitigate this, an in-network online learning-based control structure is employed by deploying the learning-based controller on a remote computation node and connecting it via a communication channel. In this paper, control performance guarantee is first established by deriving tracking error bound for the in-network control architecture, while accounting for computational delays. The derived tracking error bound allows for diverse communication and computation strategies under a specific condition, including time-/event-triggered mechanisms. Additionally, the trade-off between communication and computation performances is shown for a given desired control performance. Furthermore, to enhance the efficiency in both communication and computation, an efficient control framework with an asynchronous event-triggered mechanism in both control and online learning is devised under the existence of computational delay. The proposed event-triggered strategy is proven to achieve the same control performance as time-triggered scenario while excluding Zeno behavior. Finally, we derive an explicit expression of the proposed event-trigger condition for exponentially stabilizable systems, and demonstrate its effectiveness through simulations.

Gradient-based jailbreak suffix optimization methods typically update the suffix by retaining the candidate with the lowest current loss. We show that this seemingly natural design is fundamentally myopic: candidates that look better under the current-step proxy often fail to produce better jailbreak outcomes later in the search, revealing a form of selection-stage reward hacking. This suggests that candidate selection, rather than candidate generation alone, is a hidden bottleneck in suffix optimization. To address this issue, we propose TACS, a trajectory-aware candidate selection framework for jailbreak suffix optimization. Instead of selecting candidates solely by their immediate loss, TACS augments per-step evaluation with a trajectory-aware proxy and stabilizes selection with reference-policy regularization and a discriminator-estimated chi-squared correction, encouraging choices that remain effective beyond the current step. Experiments on HarmBench show that TACS consistently outperforms strong baselines under the same search budget, substantially improving attack success rates while exhibiting more stable optimization behavior throughout the search. Our findings highlight that mitigating selection-stage reward hacking caused by myopic candidate selection is critical for improving jailbreak suffix optimization.

School coaches prepare for opponents with game film and intuition. The analytics tools of professional teams stay out of reach. We ask how far public data can close this gap. Professional basketball is our case study, chosen for its data rather than the league. We fuse five public sources into one per-shot dataset of 4.23M shots over 21 seasons. The sources are shot locations, two play-by-play feeds, official matchup tracking, and player biometrics. Alignment across them is 99.5% to 100%. We also report two data pitfalls that are easy to miss. We then model a half-court possession as a sequential game. Shot values come from ShotNet, an embedding multilayer perceptron (MLP). On a held-out season it beats a zone-rate baseline and a logistic baseline, and its probabilities are well calibrated. A depth-limited expectimax search then solves the offensive decision tree, with branch-and-bound pruning to keep it real time. All training runs offline, so the online system stays light. A scouting planner and a playable simulator both run in a single browser page.

Ensuring the safe operation of multi-agent systems (MASs) under uncertain environments is crucial for cooperative robotic, where external disturbances and inaccurate dynamic models can significantly compromise performance and reliability. To address this challenge, calibrated machine learning models, particularly Gaussian process (GP) regression, are extensively employed due to their interpretable performance quantification. As the interconnected communication of MASs facilitates cooperative learning, agents are able to enhance learning performance by exchanging local GP inferences with their neighbors and aggregating the received information via distributed GP strategies. However, variations in computational power and prediction tasks among agents inevitably lead to heterogeneous computational delays and differences in query points, which are often overlooked in existing aggregation methods. To overcome these limitations, this work proposes an asynchronous cooperative learning strategy that explicitly accounts for prediction accuracy, query point variations and delay effects. Additionally, a distributed control law based on an adjoint MAS is developed to ensure the desired control performance. Simulations on unmanned surface vehicles validate the effectiveness of the proposed approach, demonstrating substantial improvements in both learning and control performance compared to the state-of-the-art approaches.

A company with a fixed artificial intelligence (AI) budget must decide which large language model (LLM) handles each recurring workload. What it lacks is the quality table, how well each model performs on each workload. Given that table, the decision is a multiple-choice knapsack problem and is routine to solve, so estimating it is the difficulty, and that estimation fails in two ways. Models are rarely compared on the same work, and the recorded score is usually a proxy rather than the outcome the company values. Causal and off-policy methods repair the first but condition on the second, while evaluator-validation methods estimate the second but stop short of the decision. Worse, buying more re-evaluation cannot settle the second: randomization governs which requests are scored, not how a score is produced, so the table stays uncertain however much evaluation is purchased. Yet the deployment decision may still be determined even when the table is not. We therefore ask whether one assignment stays optimal across every quality table consistent with the evidence. For the fixed-budget problem, this admits an exact two-solve certificate: solve once at the estimated table and once at a least-favourable table. Agreement certifies the assignment; disagreement identifies the model-workload pairs where further evidence can matter. We propose CASE (causal active sequential experimentation), which targets evaluation to those pairs and repeats the test as evidence accumulates. On a production log, the measurement failure is the larger of the two: correcting assignment exactly still leaves most of the loss, and randomized re-evaluation does not remove it. In our experiments, the available evidence often does not determine the assignment. On paid software tasks, better information about model quality yields more savings than further optimization of the assignment on the same estimates.

Geospatial foundation models such as the AlphaEarth Foundation produce compact and globally consistent representations of the Earth's surface that transfer effectively to a wide range of downstream tasks. However, because these models are trained primarily on Earth-observation imagery, their embeddings mainly capture physical and spectral characteristics while encoding human activity and urban function only weakly. To address this limitation, we propose BEACON, a tri-modal contrastive learning framework that aligns three complementary views of urban space: physical representations from AE embeddings, semantic representations from point-of-interest (POI) text, and human behavioral representations from hourly POI visitation, while keeping the deployed representation image-only. Using the Houston Metropolitan Area as a case study area, we evaluated the performance of the BEACON framework on nine downstream tasks, including seven regression and two classification tasks against six baselines (raw coordinates, Space2Vec, SatCLIP, TESSERA, Clay and AlphaEarth), using frozen linear and MLP probes over five seeds. Under a linear probe, BEACON improves relative R^2 over AlphaEarth by up to 43% for obesity prevalence, 34% for poor mental health, and 22% for median household income, while remaining competitive in the prediction of physical and environmental variables. These findings highlight the value of augmenting geospatial foundation models with semantic and behavioral signals, extending their applicability from physical Earth observation to human-centered urban analytics.

As context lengths scale, attention increasingly becomes a primary computational bottleneck in large language models. Standard Transformers remain powerful but computationally inefficient, as they allocate the same attention budget to every token regardless of its contextual demand. Existing local-global hybrids provide a more efficient alternative by mixing restricted- and full-context attention, but they typically allocate span statically across layers or heads. To address these limitations, we propose LoGo, a token-level dynamic local-global attention mechanism that uses attention span as a direct proxy for attention budget allocation. Each LoGo layer contains coupled local and global branches: all tokens receive efficient local attention over a restricted context window, while a learned gate activates global attention with full-context access only for tokens requiring long-range information. A threshold-based budget controller maintains a target global ratio without auxiliary losses, and a progressive masking schedule stabilizes training before sparse routing takes effect. We further implement query-sparse Triton kernels that convert reduced global-attention computation into practical speedups. Extensive experiments validate LoGo's effectiveness, showing that it preserves the scaling behavior of full-attention Transformers across model sizes. In controlled comparisons, LoGo improves over the full-attention Transformer and matched-budget static local-global hybrids, with clear gains on long-range retrieval. Analysis further shows that LoGo learns interpretable span allocation patterns. These results suggest that learned token-level span allocation is an effective and scalable way to improve the long-context performance-compute trade-off.

Longitudinal clinical agents must maintain an evolving patient state from evidence distributed across visits, time points, and specialties. However, how agent memory should be designed for this setting remains unclear. We introduce a benchmark of multi-visit, multi-specialty patient records that evaluates long-context evidence retrieval, cross-time evidence aggregation, and cross-specialty clinical reasoning. Using this benchmark, we systematically study four memory design choices: curation, organization, retrieval, and memory-augmented reasoning. We find that temporal validity is more important than simply retaining more history; specialty-factorized memory reduces context but can hide shared evidence; and multiple agents help when specialists must reason together, not merely when evidence comes from multiple memories. Guided by these findings, we propose MedCache, a hybrid framework that constructs temporally valid patient memory, organizes evidence into overlapping specialty views, routes each query to relevant memories, and adaptively invokes one or multiple specialists. Experiments show that MedCache improves reasoning accuracy and memory efficiency over strong single-agent and multi-agent baselines, while generalizing across model backbones and external datasets.

Machine unlearning enables deep neural networks to selectively remove the influence of specific data in response to privacy and regulatory requirements. While prior work largely studies single-shot unlearning, real-world systems must accommodate continual unlearning, where multiple unlearning requests occur sequentially over time. In this work, we identify a fundamental limitation of this setting: plasticity collapse, a progressive breakdown in a model's ability to effectively forget. Through theoretical analysis of continual unlearning dynamics, we show that continual unlearning operations accumulate geometric constraints in parameter space, leading to saturated subspaces that restrict future updates. This structural effect induces two distinct failure modes: (1) Forward failure – diminishing forgetting quality for subsequent tasks, and (2) Backward failure – spontaneous re-memorization of previously forgotten information. Extensive experiments across multiple architectures, datasets, and methods in image classification confirm that plasticity collapse is not an artifact of specific implementations, but a pervasive phenomenon inherent to continual unlearning. Our findings reveal a critical barrier to the long-term reliability of machine unlearning systems and motivate the development of plasticity-preserving unlearning algorithms. Our code is available at https://github.com/TIML-Group/Continual-Machine-Unlearning-Plasticity-Collapse

Aerial object detection is increasingly deployed in real-world applications, but models remain vulnerable to physical, universal adversarial patches that cause them to miss objects. Furthermore, defenders face the practical constraint of training data scarcity: aerial imagery is costly to collect and label, so a deployment site typically yields hundreds of images rather than the tens of thousands that adversarial robustness benchmarks assume. To tackle model vulnerability and training data scarcity, we propose Adversarial Robustness with Manifold-Oriented Training (ARMOR), a novel defense that realizes the core insights of on-manifold adversarial training (OMAT) in low-data regimes. ARMOR builds on the insight of OMAT to model the data manifold - the compact structure capturing the data's relevant features - to learn and robustify these features during training. While OMAT relies on the data-intensive operations of training large generative models and adversarial training to achieve this, ARMOR adopts a data-efficient approach that reuses labels the detection task already supplies: ARMOR (i) masks image backgrounds to retain object-relevant features, and (ii) injects randomized patches on objects to improve feature robustness. Our low-data experiments with physically-realizable adversarial patches evaluate both query-free transfer attacks and defense-aware attacks. ARMOR maintains strong clean performance of over 0.90 model confidence, while improving adversarial robustness by up to 0.32 in model confidence over state-of-the-art defenses. Physical experiments with printed patches confirm that these gains survive deployment. Overall, ARMOR translates insights from manifold-based training to defend object detectors amidst training data scarcity.

Diffusion models are increasingly used not only for sampling from learned data distributions, but also for generating samples that optimize task-specific objectives. A common approach is to guide the reverse diffusion process using gradients of an external objective. However, when the data distribution is supported on a structured feasible set, such as a manifold or a constraint set, gradient guidance can move samples away from the learned data geometry. In this paper, we study a simple projected-gradient-guided diffusion update based on the observation that the Stein denoising operator can act as an approximate projection onto the data geometry. The proposed update incorporates the objective gradient inside the denoising step, yielding an inference-time method that uses only a pretrained denoiser and gradient evaluations. We analyze this update as an inexact projected-gradient method for constrained optimization over learned feasible geometries. Our theory covers three settings: linear manifolds, compact convex feasible sets, and compact Riemannian submanifolds. In all these settings, we prove descent and finite-time convergence guarantees. Numerical experiments support the theoretical interpretation and illustrate how the proposed update balances objective descent with preservation of the learned geometry.

Worst-case online classification is governed by sequential complexity, such as Littlestone dimension, and can be impossible even for statistically simple classes, such as thresholds of VC dimension one. We study a preview model in which an oblivious adversary fixes an entire labeled sequence of length $T$, a uniformly random subset of size $pT$ is revealed before prediction begins, and the remaining $(1-p)T$ examples are then presented in their original adversarial order. Against the best full-sequence hypothesis evaluated on the unrevealed examples, we characterize the dependence on the preview rate $p$: for binary classes of VC dimension $d$, the optimal excess loss is $\Theta(d/p+\sqrt{dT})$, up to the trivial cap at $T$; for multiclass classes we obtain the corresponding $\widetilde O(d_{\rm DS}/p+\sqrt{d_{\rm Nat}T})$ bound with no dependence on the number of labels. Thus a random preview can replace worst-case sequential complexity by classical statistical dimensions without randomizing the online order. To achieve the sharp binary bound, our ChainedPrediction algorithm uses an online analogue of chaining, implemented as a multiscale aggregation algorithm rather than only as an analytic argument.

Accurate estimation of body-composition outcomes, including body fat percentage (BFP), bone mineral density (BMD), and appendicular lean mass (ALM), is important for evaluating metabolic, skeletal, and muscular health. Direct assessment using dual-energy X-ray absorptiometry (DXA), however, requires specialized equipment and involves ionizing radiation. We propose a target-aware, state-adaptive $p$-Dirichlet energy-flow graph neural regression ($p$SADE-GNR) framework for estimating these outcomes from non-invasive anthropometric measurements. A neural encoder maps participant representations to hidden states that are propagated over an outcome-specific participant-similarity graph by a state-adaptive forward-Euler discretization of the graph $p$-Dirichlet energy flow. Graph distances weight each original or latent coordinate by its normalized absolute training-fold correlation with the outcome. Using clinical data from the Pennington Biomedical Research Center and five-fold cross-validation, the correlation-weighted model using the original standardized measurements achieved the lowest root mean squared error in all nine primary outcome-cohort combinations and outperformed previously reported support vector regression or least-squares support vector regression reference values in eight of nine comparisons. Autoencoder, variational-autoencoder, and Gaussian-mixture variational-autoencoder representations generally did not improve primary-outcome prediction or reduce computational cost. In an exploratory age-prediction analysis including ALM, BMD, and BFP as predictors, the correlation-weighted GMVAE model achieved the lowest mean error in all three cohorts. These results support target-aware, state-adaptive $p$-Dirichlet graph neural regression for non-invasive body-composition estimation.

Much of human health and function unfolds beyond the clinic, through the movements of everyday life. Wrist-worn accelerometers capture these movements continuously, yet their rich signals are often reduced to a small set of predefined behavioural summary measures. Here, we present Sensori, a self-supervised foundation model that learns general-purpose health representations directly from 24 hours of raw tri-axial wrist movement. We developed and evaluated the model across four population-based cohorts from the United Kingdom, China and the United States, comprising 122,640 participants contributing 683,617 person-days of free-living recordings. Sensori condensed each day of movement into a representation that captured diverse movement behaviours, demographic characteristics, health axes and physical function. Evaluation in independent cohorts showed that these representations generalised across populations and measurement settings without retraining. When added to common clinical covariates, Sensori significantly improved prevalent disease classification for 52 of 102 eligible conditions (median delta AUROC, 0.060; range, 0.012-0.242) and incident disease risk prediction for 26 of 87 eligible conditions (median delta Uno's C-index, 0.064; range, 0.025-0.172), with the largest gains for neurological and psychiatric disorders. These findings establish 24-hour wrist movement as a rich and scalable source of health information, with the potential to support passive health monitoring and disease prediction at population scale.

Tabular-to-image encoding methods enable the application of models based on both convolutional neural networks and vision transformers to tabular data, transforming feature vectors into images. Existing methods employ linear and nonlinear dimensionality reduction techniques (e.g., Principal Component Analysis (PCA), t-SNE, and UMAP) to determine pixel positions, resulting in images whose spatial layout do not inherently reflect feature relationships. This paper introduces Visibility Graphs for Tabular-to-Image Encoding (VG-TIE), a novel method that encodes the structure of feature values using Natural Visibility Graph (NVG) and Horizontal Visibility Graph (HVG) into a two-dimensional space obtained through PCA. The resulting images are model-agnostic and intrinsically interpretable. Each pixel corresponds to an input feature, its intensity reflects the magnitude and direction of deviation from the population mean, and edges represent formally defined visibility relationships between features. VG-TIE provides two interpretability methods: (i) feature ranking from node degree distributions; and (ii) local and global feature importance from pixel intensity combined with Grad-CAM. Experiments on six public tabular datasets show that VG-TIE is competitive with other tabular-to-image methods while providing interpretability on feature importance and ranking similar to intrinsic interpretable methods. The results highlight the potential of the proposed image-based transformation to provide an effective framework that expands the use of deep learning across tabular data domains.

Video object detection on edge devices runs computationally expensive detectors over long frame streams, causing high energy consumption and sustained GPU utilization. Although consecutive frames are highly redundant, naive frame skipping is content-blind: it skips during critical moments such as object entry, occlusion recovery, and abrupt motion, degrading detection quality. We present Albireo, a detector-agnostic, codec-free adaptive inference framework that wraps off-the-shelf detectors and decides when detector invocation can be safely skipped based on scene content and per-object temporal state, requiring no detector modification or retraining. Albireo maintains a 10-dimensional Kalman filter (KF) per active object and invokes the detector only when prediction uncertainty exceeds a threshold; on skipped frames, boxes are predicted from the KF state at near-zero GPU cost. A KF-based rescue mechanism preserves confirmed objects through brief detector misses to prevent output fragmentation, while a lightweight empty-scene screen avoids detector calls on objectless frames. We evaluate Albireo on the BDD100K MOT validation split with three architecturally distinct detectors (YOLO11x, YOLO26x, RF-DETR-Large) on two NVIDIA Jetson platforms (AGX Thor, AGX Orin). Across all configurations, Albireo keeps AP@50 within +/-1.2 pp of per-frame inference while reducing total energy by 12.1-17.6%. On YOLO26x, it improves AP@50 by +0.8 pp while reducing energy by 17.6% (Thor) and 14.4% (Orin) and per-frame energy-delay product by 24.9% and 26.1%, respectively. Thus, the default operating point improves accuracy, energy, and latency together. In contrast, FixedSkip-2, a fixed-interval baseline with a 50% skip rate, loses 8.6 pp AP@50. Source code, evaluation pipeline, and per-clip results are available at https://github.com/amirtaherin/albireo

Recent topic models leverage pretrained embeddings, but neural architectures produce latent representations without grounding in specific texts, and clustering-based pipelines assign representative documents only post hoc, relying on absolute distances distorted by hubness and anisotropy in high-dimensional spaces. We introduce MARETopic, a training-free framework that casts topic discovery as rank-based prototype selection. After projecting embeddings onto a low-dimensional manifold, MARETopic builds ranked lists encoding ordinal neighborhood structure. A greedy algorithm selects exactly K exemplar documents, real corpus texts, whose neighborhoods cover the corpus. Two variants share this criterion. MARETopic$_\text{Corr}$ scores candidates with a query performance predictor and a rank correlation measure, leading Purity and NMI on the two benchmarks with the most categories, ahead of both neural and clustering-based topic models. MARETopic$_\text{Diff}$ scores them with a rank-based diffusion matrix, needs neither measure, and runs 1.7 to 1.9 times faster. Without a single gradient update, MARETopic leads topic coherence on two of three datasets. A novel inter-topic Maximal Marginal Relevance step raises vocabulary diversity at little cost in coherence. Our code is available at https://github.com/thcastilho/maretopic.

Worst-case Optimal Recovery (OR) and Bayesian learning describe the same Gaussian-quadratic-Hilbert problems in two vocabularies. We sharpen the correspondence - the radius of information equals a nugget-optimized GP posterior variance and is attained by the posterior mean at a closed-form balance nugget - and measure, inside three published Bayesian systems, where the worst-case side pays. The ledger is two-sided: the losses instruct as much as the wins. Morozov calibration tracks a test-access oracle within $1.00$-$1.19\times$ where $\sigma$-blind rules fail, is $4.9$-$6.3\times$ more reproducible across noise draws ($p=0.002$-$0.004$), and is the only deployable rule whose selection survives a change of backend ($1.36\times$ against $12$-$30\times$ for the released weight, ML-II and GCV); tight certificates cover at the information-theoretic floor with no numerical slack. But on exchangeable data split-conformal beats the OR head on interval score, a water-filling prior adds nothing without an oracle noise hint, and under covariate shift the OR band keeps coverage on every dataset yet loses interval score to split-conformal, and to a feature-free constant band, on most cells; what pays is not shift but shift on a learnable target, which a training-free audit statistic predicts before any model is fitted. In Bayesian optimization the certified width is a validity floor whose scalar inflation we prove inert under a checkable margin condition and check at every step. Inertness is graded, not binary, and in the size of the inflation as much as in the objective: $\kappa{=}2$ is inert wherever $\kappa{=}5$ is and on more cells besides, while $\kappa{=}5$ moves half the Ackley seeds and every Griewank seed. Exploration is a shape problem, not a scale one. The design rule: match the guarantee tool to the data regime, and audit the regime first.

Enterprise analytics agents solve long-horizon tool-use problems over distributed business data, requiring retrieval, reasoning, API calls, code execution, and adaptation to intermediate observations. Supervised fine-tuning (SFT) calibrates tool syntax and teacher-supported behavior, whereas reinforcement learning (RL) can explore reward-supported behaviors beyond demonstrations; applied uniformly, however, RL can perturb already-calibrated skills. We study how to balance SFT and RL under production-mirroring beta APIs. We observe that, in our controlled experiment, checkpoint trajectories retrospectively separated into three regimes: Imitation, where SFT captured reliable teacher behavior; Lift, where both stages helped; and Discovery, where useful reward-observable behavior lay outside reliable teacher support. We leverage this prospectively, using teacher support and reward-observable headroom to route features to SFT only, SFT then RL, increased RL allocation, or further environment development. Across 18 subsequent feature-specific experiments, the diagnostic predicted 15/18 observed trajectories. On GPT-OSS 120B, targeted SFT then RL produced positive point estimates on 7/8 advertiser skills relative to a frontier Control; five positive gains had paired 95% confidence intervals excluding zero, while one skill had a confidence-supported regression. The largest gain was non-disclosure (+11.27 points; 95% CI [+9.72, +12.82]). A separate SME audit surfaced that targeted RL reduces standard leakage from 11.8% to 2.9% and adversarial leakage from 22.9% to 6.8% relative to SFT while preserving actionability (86.2% to 85.7%). In a matched uniform-versus-targeted comparison with shared rewards and optimization, targeted RL improved the seven-skill mean delta from +1.62 to +3.57 while using 43% less incremental RL compute.

Heterogeneous multivariate time-series systems exhibit segment-specific nonlinear dynamics that challenge monolithic forecasting architectures. We propose SegTSim, a big-data-driven segmented temporal simulation framework that integrates segment-specific elasticity modeling with adaptive min-gating, dynamic production relocation optimization, multi-factor data fusion with exchange-rate propagation, and a deep ensemble validation pipeline. The framework is validated on US–Japan automotive trade data from USITC repositories spanning 2015 to 2025, comprising approximately 13000 annual records. Under a 25% perturbation scenario, Japanese import volume declines by 20.4% to 0.93 billion USD, while all output variables maintain coefficients of variation below 3.5% across 1000 ensemble inference runs.

Machine learning is being increasingly used to predict where active wildfires will burn the following day, helping inform evacuation boundaries and containment lines. Most models are evaluated using Average Precision (AP), which summarizes performance across all decision thresholds, although acting on a forecast requires choosing one. We benchmarked five discriminative architectures and one generative model on WildfireSpreadTS using a shared evaluation pipeline and two input configurations. We found that model rankings varied depending on whether performance was measured by AP or by threshold-dependent metrics like F1 and IoU. At the default 0.5 threshold, the highest-AP model flagged 4 to 5 times the observed active-fire area and ranked fifth of six on F1 and IoU, and the most recall-heavy model flagged 16 to 23 times. The three models whose flagged area stayed close to the observed area had AP scores 24 to 37% lower. Across models, we identified three distinct prediction profiles: over-predicting, balanced, and under-predicting, which AP alone could not distinguish. Expanding the input from 7 to 23 channels changed AP by 0.03 on average, against a 0.21 to 0.24 spread across models. Ultimately, these results show that AP alone can favor models whose predictions are poorly suited for operational wildfire forecasting.

The creation of telescope bibliographies is a crucial part of assessing the scientific impact of observatories and ensuring reproducibility in astronomy. This task involves identifying, categorizing, and linking scientific publications that reference or use specific telescopes. However, this process remains largely manual and resource intensive. In this work, we present an efficient SciBERT-based approach for automatic classification of scientific papers into four categories - science, instrumentation, mention, and not telescope. Despite strict context-length constraints (maximum 512 tokens) and limited compute resources, our approach achieved a macro F1 score of 0.89, ranking at the top of the WASP-2025 leaderboard. We analyze the effect of truncation and show that even with half the samples exceeding the token limit, SciBERT's domain alignment enables robust classification. We discuss trade-offs between truncation, chunking, and long-context models, providing insights into the efficiency frontier for scientific text curation.

Knowledge distillation trains a small student model to reproduce the outputs of a large teacher model, and its progress is typically monitored through the teacher–student discrepancy. The quantity of ultimate interest, however, is the student's error with respect to the true task. We study the relation between these two objectives in a minimal three-party model, a true teacher (generative model), a teacher, and a student, all soft committee machines, in which the true teacher contains a shared latent factor that the teacher cannot represent, with mismatch strength controlled by a single scalar $\dmiss$. Within an order-parameter description of online distillation, and exploiting closed-form (arcsine-type) expressions for all errors under error-function activations, we prove that the learning dynamics and the distillation error $\Ets$ are exactly invariant to $\dmiss$, whereas the true error $\Etzs$ and the gap $\Delta=\Etzs-\Ets$ are strictly increasing in $\dmiss$, with a rate that is amplified linearly by the complexity $M_0$ of the true teacher. Numerical phase diagrams over the plane spanned by true-teacher complexity and student capacity confirm the predicted deformation: the contours of $\Ets$ do not move while the landscape of $\Etzs$ rises systematically, and a teacher-miss regime, where mimicry succeeds but the task fails, expands with $\dmiss$. The results give a quantitative warning against evaluating distillation solely through teacher-mimicry metrics and identify the gap $\Delta$ as a minimal diagnostic for distinguishing teacher-miss from capacity-limited failure.

Deployed decisions are often optimized once and retained because updates impose operational, regulatory, or switching costs. As operating conditions change, when should such decisions be re-optimized? We study this question for stochastic optimization when the objective's functional form is known but the decision maker's trade-offs are encoded by an unknown preference parameter. Standard distribution-shift tests are poorly aligned with this goal: they can flag detectable yet decision-irrelevant changes without determining whether the incumbent decision has become materially suboptimal. We propose RADAR (Regret-based Assessment of Decision Adequacy and Risk), a decision-focused framework that uses inverse optimization to infer latent preferences and tests the deployed decision's optimality gap under the current distribution. By targeting regret, RADAR ignores decision-irrelevant shifts while detecting changes that warrant re-optimization. We develop two-sample and sequential changepoint procedures and establish asymptotic guarantees for Type-I error and power. Across synthetic optimization problems, a semi-synthetic capacity allocation task, and police-zone planning, RADAR more reliably distinguishes harmful from harmless shifts than decision-agnostic alternatives.

A benchmark score is a joint property of the model, the evaluation harness, the elicitation budget, the sampled population, and contamination status. Leaderboards publish the model and the score, so capability and leakage stay observationally equivalent. Existing taxonomies classify contamination for automated detection, not the question a reporter faces at publication: given the mitigations already applied, which validity threats remain open? We introduce a taxonomy organized by the mitigation each type defeats – direct, derivative, temporal, distributional, and acquired – spanning training-time and evaluation-time leakage. Holding out a private test set closes the first alone. The fifth is acquired during the evaluation itself; because it is a property of one run, it must be recorded with the reported score rather than with the benchmark release. We operationalize it as a four-field disclosure protocol in which "unknown" is a valid entry, released under CC BY 4.0 with a JSON Schema, a validator, and worked examples. Two coders external to the design team applied a pre-registered instrument to 41 documents. Per-variable linear-weighted $\kappa$ runs from 0.00 to 0.35 (median 0.21) over 29 main-pass documents against a single-coder test-retest ceiling of 0.84, collapsing under the class skew the registration anticipated; pooling raises it to 0.46 through chance correction rather than better agreement. Two variables fall below the prevalence-robust threshold registered in advance: strata reporting and the acquired type introduced here. Disagreement concentrates on when a variable applies rather than on what a document states. Elicitation budgets are reported in 13% of documents, and no document addresses all five types. The contribution is the taxonomy, the score-side artifact that follows from it, and a pre-registered measurement of instrument reliability and current disclosure.

Sandbagging models strategically underperform on evaluations while retaining the capabilities being measured. The evaluations that guide frontier-model deployment and governance then understate what these models can do. To understand the mechanism, we propose a causal model of how sandbagging is carried in the residual stream. Early layers write the sandbagging intent onto a single axis of the stream, and a later layer reads that axis and commits the answer. We study three instruction-tuned models (Qwen2.5-7B, Llama-3-8B, and Mistral-7B) and four ways of installing a sandbagging lock (prompting, fine-tuning, reinforcement learning, and circuit breaking). Each lock decides from the prompt whether to sandbag, and the fine-tuned and circuit-broken locks answer honestly whenever a password appears. The causal model predicts a window of layers, after the last sandbagging write and before the answer commit, in which a single-layer reference graft of the sandbagging axis to its honest value restores the full capability. The single-layer graft recovers the capability in 28 of the 33 runs of the prompted, fine-tuned, and RL-trained locks, with a median held-out recovery of 96%. The circuit-broken lock rewrites the whole state through a band of layers, and the single-layer graft fails at every layer. We therefore introduce a second intervention, context grafting, which replays the password's cached key/value activations so that every layer's attention reads them as additional context. Context grafting provably and empirically restores the full capability on all three models, and the recovery is surprisingly insensitive to the exact password content. More broadly, an auditor can use this causal model to design interventional auditing techniques for sandbagging models.

Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on. The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. The direction tracks how the sandbagging was induced rather than what is withheld – fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation.

LLMs are increasingly used as human surrogates, often on the premise that richer persona data could make them substitutes or exploratory tools for specific individuals. We test this premise across four datasets covering more than 400,000 participants and more than 6,000 survey items and experimental outcomes. LLMs perform well at the aggregate level: their average responses closely align with average human responses to the same items. But this success largely reflects predicting each item's average human response. Once each item's human mean is removed, LLM predictions explain only 3.05% of the remaining respondent-specific variation, far below the 53.6% human test-retest benchmark. Richer personas, model variants, and fine-tuning do not close this gap. In variance analyses, once item means are removed, the reliable remaining signal is person-by-item. It captures how a respondent departs from the mean on a particular item and is about 8.9x larger than the stable person effect. Persona data encode the respondent, but not this item-specific deviation. LLM responses also compress human response distributions, using less spread, fewer response categories, and distorted distributional shapes. We call this pattern item-mean surrogacy. Current LLM surrogates can approximate item averages, but not the distributions or respondent-specific deviations needed to replace individual humans. We propose four empirical tests for LLM-based human-surrogate claims.