Closed-loop autonomous experimentation is explored using coding-agent infrastructure for synchrotron X-ray experiments on nanoparticle superlattices. The research demonstrates the use of human-agent and agent-agent transcripts to audit decision provenance and cognitive labor division. Analysis of three experiments revealed that 61-67% of messages involved interpretation or judgment, and methodological contributions increased over time. Error corrections frequently involved multiple participants, and decisions often incorporated input from other agents. These findings suggest design principles for multi-agent scientific systems and autonomous experimentation.
EarthRecord establishes a benchmark for evaluating multimodal agents on land-change analysis using four decades of raw satellite imagery. It incorporates data from Landsat, MODIS, VIIRS, and Sentinel-2, requiring agents to handle cloud masking and cross-sensor harmonization. The benchmark poses 238 land-change questions across ten tasks, assessed against global references and method agreement. Evaluations indicate that while agents complete most analyses, they struggle with mapping from imagery to real-world change. Ultimately, EarthRecord offers a framework to measure and advance agents from pixel-level data to actual land change.
A framework called TopicScope uses sparse autoencoders to train causal topic classifiers on Sparse Autoencoder (SAE) features assigned to every token. This allows for topic labeling to change throughout a model's generation process. Monitors trained on features from prefill hidden states were shown to generalize to decoding, even when detecting topics across different multi-topic conversations. These classifiers proved robust to perturbations like inserting off-topic asides or reordering sections. The research demonstrates that compact topic monitors can be trained efficiently in prefill and deployed during generation.
LandslideFM is a unified foundation model designed to improve landslide perception and understanding by connecting visual perception tasks with contextual interpretation. It utilizes a Source-Agreement Graduated Trust semi-supervised learning framework to adapt to multiple visual tasks using limited labeled and abundant unlabeled data. The model incorporates predicted landslide regions to guide a Vision-Language Model through parameter-efficient fine-tuning for comprehensive reasoning. LandslideVL is presented as a new benchmark combining visual imagery with textual prompts for landslide analysis. Experiments show LandslideFM achieves superior performance in specialized landslide reasoning compared to general-purpose Vision-Language Models.
The study investigates whether large language models internally represent psychological constructs or merely generate outputs resembling psychometric structures. Using responses from 272 university students, the models' predictions across six psychological scales were analyzed. Larger model variants demonstrated an amplification of human cross-scale correlations at the behavioral level. Contrastive direction analysis showed geometric alignment with the human matrix in larger instruct variants, but less alignment in base variants. The findings suggest a measurable gap between internal structure and output behavior, termed the probe-behavior gap readout attenuation.
Multimodal input, specifically text-plus-graph visualization, improves knowledge graph question answering performance when using large vision-language models. Adding a graph image resulted in an average F1 gain of 1.85 points, with gains being greatest for low-complexity questions. These multimodal benefits depend heavily on graph layout, as hierarchical Graphviz layouts perform better than less organized renderings. Optimally, visual prompting should be adaptive, selectively applying diagrams to small, well-structured subgraphs to maintain accuracy and efficiency.
Reinforcement Learning is applied to control cancer chemotherapy treatments by estimating optimal dynamic treatment regimens from clinical data. The research explores state-of-the-art reinforcement learning techniques, including online and offline approaches, for sequential treatment generation. A simulation using traditional Q-learning was presented, extending to offline learning using a model-based architecture with conservative Q-learning. The proposed Chemo RL model effectively addresses cancer chemotherapy scheduling and supports junior oncologists in treatment decisions. This solution offers a data-driven approach based on real human colon cancer clinical data.
The paper proposes a competence circuit (IKC + KC + C) to decompose an agent's impact into intelligence-mediated discovery, procedural know-how, and direct control. It formalizes conditions showing that fast catastrophic supremacy relies more on rapid knowledge acquisition or control attainment than solely on intelligence. Simulations indicate that intelligence-driven speedups are limited by the ratio of competing intelligence levels. The research argues that controlling and setting knowledge boundaries are more critical safety levers than merely limiting intelligence. Ultimately, safety in agentic systems depends on restricting control and knowledge, especially when institutional delays exist.
Tabular foundation models (TFMs) are evaluated against tuned gradient boosting trees for predicting telecom customer churn, considering profit and calibration. Experiments compare models like TabPFN-3.5 and TabICL v2 with optimized GBDTs across five public datasets. The TFM approach demonstrated a significant expected maximum profit (EMP) advantage over the best tuned GBDT on the Cell2Cell and Iranian datasets. However, performance differences varied across datasets, and real-world applicability depends on the licensing of the foundation model weights. Overall, the results suggest TFMs offer a compelling performance gain, though caveats regarding cost and production use exist.
SetPrint is a training-free method for black-box LLM fingerprinting by analyzing response changes across perturbed queries for two models. It captures both the relations among response variations within each set and the correspondence between response changes between the two models. The method achieves this by expressing response changes relative to the mean change of their respective sets. SetPrint consistently outperforms existing training-free fingerprinting methods on the LeaFBench dataset. The performance gains are attributed to the captured within-set relations and between-set correspondence.
Assisted history matching for the Brugge field is addressed using an ensemble of graph generative models. The approach models each stratigraphic zone with a dedicated generative model possessing a joint latent space for optimization. Graph-based generators handle the irregular geometry of the grid directly, enabling optimization in the 160-dimensional latent space via Covariance Matrix Adaptation Evolution Strategy. The objective function incorporates well data mismatch, production mismatch, and a geological realism penalty based on the latent metric. The resulting matched models preserve geological relations and reproduce production results comparable to prior studies.
The research introduces ConceptTime, a method using predictive surprise as a self-grounding signal for interpretable time series analysis. A frozen probabilistic forecaster predicts future observations, and the mismatch between its prediction and the actual signal is termed predictive surprise. This surprise locates segment boundaries and characterizes segments into "concepts," which are dynamical regimes. A lightweight head classifies based on these concepts, and the distance to the nearest concept provides a measure of reliability. ConceptTime demonstrates competitive accuracy and faithfulness across various benchmarks without requiring external annotations.
ADIndex is an Adaptive and Distillation-driven Indexing method designed for token-level sparse attention to reduce latency in long-context inference. It incorporates a layer-wise calibration strategy and a group-level distillation scheme to efficiently train adaptive group indexers from query signals. A low-latency inference strategy further speeds up processing by sharing token bank information within query groups. Experiments show ADIndex preserves performance while achieving significant speedups, up to 6.22x and 3.40x for prefill and decoding, respectively.
Reward heterogeneity in personalized federated reinforcement learning is framed as an asset for exploration. The paper introduces PF-UCBVI to achieve linear speedup by separating shared dynamics from personalized objectives. Personalized Federated Exploration-Free Value Iteration (PF-EFVI) is proposed as a greedy algorithm utilizing reward diversity for state-action coverage. PF-EFVI is shown to achieve logarithmic regret without explicit exploration bonuses under sufficient reward diversity. Agent disagreement shifts exploration from temporal complexity to spatial diversity, allowing efficient collective learning.
StatPrompt-LLM is a lightweight statistical refinement method developed to improve foundation-style time-series forecasting. It augments the Time-LLM prompting mechanism by incorporating dataset-level statistical information. The method also introduces a zero-initialised autoregressive bypass as an alternative linear forecasting path. Evaluation on electricity and meteorological benchmarks showed StatPrompt-LLM achieved lower prediction errors than Time-LLM with reduced-prefix configurations. Further analysis indicated that performance gains were dataset-dependent and computationally incurred some overhead.
MADSNet is a Multi-scale Axial-attentive Dense Spiking Network designed for spike camera image reconstruction. It utilizes a hierarchical spiking encoder-decoder structure incorporating a multi-scale spike stem, axial spike attention, and a dense spiking bottleneck. This network reconstructs continuous intensity images from the spike-domain representations. Experiments show MADSNet surpasses other methods in reconstruction fidelity while achieving significantly lower latency and energy consumption.
The research investigates whether the utility of a self-update persists after subsequent native updates in recursive self-improvement. The study evaluates five 3-8B backbones on math and code tasks using 64 fresh runs and candidate/continuation evaluations. Results show that immediate utility is often near zero or negative, with continuation slightly reducing utility for some models. The study suggests that persistent self-updates should be evaluated as adaptation transitions rather than immediate performance. Reliable mitigation mechanisms or single mechanisms for retention changes were not identified.
Multimodal attributed graphs (MAGs) integrate text, images, and other modalities with relational edges, applicable across various domains. Learning on MAGs involves graph-centric methods focusing on modality fusion and foundation-model-centric approaches using pre-trained encoders. The survey organizes methods into five integration levels and identifies cross-modal sensitivity as a key property. It reviews benchmarks, defines model-independent descriptors, and maps evaluation protocols. Frontiers include connecting multimodal graph learning to AI agents and graph engineering for workflow representation.
Linear probes are used to study geometric structures in Large Language Models (LLMs) by analyzing their activations. The study involves a four-stage Bayesian pipeline to test claims of decodability, shape, ownership, and causal use of these shapes. Applying this pipeline to two-digit arithmetic in models like GPT-J and Llama 3.1 revealed that probe results alone overstate what a model computes. Results indicate that 19% of shapes are owned and 81% are inherited. The authors propose reporting probe results based on all four claims instead of a single accuracy measure.
The paper introduces the One Table Benchmark to evaluate generalist tabular models across generation, imputation, and prediction tasks. It compares performance across 56 datasets and 46 methods using aligned data splits and explicit task contracts. Analysis reveals that generation has broad positive auxiliary effects, while prediction can improve imputation and synthetic data utility, with results varying by backbone. Fixed models show trade-offs between utility and fidelity, but frozen foundation models demonstrate strong performance across all three tasks. The findings suggest a need for unified tabular models that integrate useful conditional information.
The study investigates the sample complexity for agnostic PAC learning under deterministic labeling, where labels are provided by an arbitrary deterministic function. The research shows that deterministic labels can improve the sample complexity bound from the classical agnostic guarantee. The main result establishes a sample complexity of $\widetilde{O}\left( \frac{K^{1/3} \, d^{2/3}}{\varepsilon} \right)$ for any class with VC dimension $d$ and finite diameter $K$. The authors construct classes where this bound is tight and also analyze the online learning analogue. This performance is achieved using memorization-based learning algorithms that leverage the deterministic label structure.
Causal mediation analysis decomposes treatment effects into indirect and direct pathways through mediators. Tabular foundation models can estimate conditional means and propensity scores needed for this analysis via in-context learning. MediPFN introduces an amortized inference framework that adapts a pretrained tabular model for nuisance prediction. It combines these predictions using a cross-fitted influence-function estimator to quantify uncertainty. Experiments show MediPFN achieves coverage closer to the nominal level and is significantly faster at uncertainty quantification than bootstrap methods.
The Continuous Horizon (CH) strategy generalizes existing multi-horizon forecasting methods by treating forecasts as a continuous operator and the horizon as a query variable. The paper establishes theoretical foundations, including a Universal Operator Approximation Theorem and a continuous-horizon bias-variance decomposition, to prove estimation variance improvements. Empirical validation on 16 M-series benchmark datasets shows CH achieves competitive Root Mean Square Error (RMSE) accuracy. This approach uniquely enables parallel inference, targeted horizon querying, and interpolation/extrapolation across unseen horizons.
A novel Gaussian process framework is proposed for time series change point detection. A change is defined as an alteration in the hyperparameters of the Gaussian process prior. The method is formulated within statistical hypothesis testing for both offline and online detection of multiple change points. This approach avoids the need for detailed prior specification of the time series before and after a change. Performance is evaluated on synthetic and real data, demonstrating efficiency.
The paper computes the VC-dimension for metric ball classifiers across various Riemannian geometries, including Euclidean, spherical, and hyperbolic spaces. It establishes upper and lower bounds for the VC-dimension in contexts such as general normed vector spaces and manifolds. The work addresses situations where the VC-dimension is large or infinite by using covering arguments under additional data distribution assumptions. The findings are applied to coverage-based algorithms in active learning, considering data on low-dimensional subsets. This research connects geometric machine learning theory with generalization bounds and active learning methods.
Methods are developed for inferring scores and rankings from human pairwise comparisons while preserving individual privacy using edge differential privacy within the Bradley-Terry model. New leave-one-out analysis provides finite-sample bounds on estimation errors for objective-perturbed estimators. The work combines plug-in ridge-bias correction with Gaussian-Laplace calibration to improve finite-sample performance. A bootstrap procedure is also established for simultaneous rank inference, offering confidence sets for ranks and top-$K$ membership. The methods are evaluated through simulations applied to ranking large language models for creative writing.
The study focuses on simultaneously identifying near-optimal regions in continuous simulation optimization, considering multiple tolerances. It investigates how to find these regions with a specified estimation error while balancing complexity and sampling. The authors establish a lower bound on the number of required simulation evaluations for this region identification. They propose SASORI, a method that estimates both an optimal point and the regions for all tolerances from the same evaluations. This method achieves the optimal trade-off between error and regret, reducing to standard optimization at zero tolerance.
The study investigates the effect of outliers on the hard-margin linear Support Vector Machine when the feature dimension increases while sample sizes are held constant. For one-sided contamination, the centered Gram matrix shows a structure of inlier blocks and an outlier-diagonal component, leading to a closed-form dual problem. These effects allow for the explicit determination of thresholds related to loss of support, margin shrinkage, and misclassification. Contamination in both classes results in at most one class collapsing or being misclassified, while symmetric contamination preserves the limiting discriminant function despite margin reduction. Numerical experiments confirm these findings regarding dual allocation and phase transitions.
The paper addresses the problem of estimating the power spectral density of a Vector-valued Autoregressive Moving-Average (VARMA) process. It proposes a new moment problem based on transportation entropy for learning the spectral density. This method allows for estimating the spectral density without restricting the moving-average (MA) component to be fixed or scalar. Existing methods are limited because they assume a fixed or scalar MA component. The proposed approach extends the class of ARMA models for spectral estimation.
The paper constructs a random forest framework designed for compositional responses, accommodating representations on the simplex, log-ratio coordinates, and the unit sphere. It utilizes log-likelihood gain as the split criterion, employing transformations like the square-root and log-ratio transformations to facilitate flexible criteria. The work introduces a tangent-normal (TN) distribution for the sphere and a multivariate logistic-normal distribution for Euclidean space. Simulations indicated that the TN forest was the most accurate across the different compositional designs.