Brief

Researchers have revisited Valiant's original 1984 learnability model, which differs from the more common PAC learning model by allowing learners to issue membership queries and requiring hypotheses with no false positives. They established a new characterization for learnability in Valiant's model, showing it is strictly between PAC learning and a variant without queries. The study also presents the first algorithm for learning $d$-dimensional halfspaces within Valiant's framework, demonstrating their learnability with queries. AI

Researchers have developed a new algorithm that tightens sample complexity bounds for identifying optimal policies in risk-sensitive reinforcement learning. The work addresses a gap between theoretical lower bounds and existing upper bounds, specifically for problems involving the entropic risk measure. By employing novel technical innovations, including sharper concentration bounds and a new stopping rule, the algorithm achieves a sample complexity that matches the established lower bound. AI

Researchers have introduced a new sampling method for Flow Language Models (FLMs) called marginal-conditioned bridges. This technique adapts continuous flow matching for token sequences, addressing limitations in standard diffusion model samplers. The proposed method samples endpoints from FLM token marginals and then uses an analytic Ornstein-Uhlenbeck bridge, offering improved quality-diversity tradeoffs and principled control over decoding. AI

Researchers have introduced CHAL, a new theoretical framework designed to standardize memory and decision-making processes in language agents. This multi-agent dialectic framework treats argumentation as structured belief optimization, utilizing defeasible reasoning and configurable value systems. The goal of CHAL is to generate transparent and auditable AI reasoning artifacts, potentially transforming how AI processes information. AI

Two articles discuss the nuances of fine-tuning AI models. One guide explores how to build specialized, smaller models that are efficient and outperform general-purpose ones. The other article questions the necessity of fine-tuning, suggesting that pre-trained models are often sufficient for many AI tasks. AI

Researchers have developed a new Python package called 'nonconform' to improve anomaly detection methods. This tool integrates with existing machine learning libraries to provide statistically calibrated p-values, moving beyond heuristic thresholding. The package aims to make conformal anomaly detection more accessible and reproducible for both experimental and production environments. AI

Researchers have introduced AntAngelMed, a 103 billion parameter open-source medical language model. It utilizes a Mixture-of-Experts (MoE) architecture, activating only 6.1 billion parameters per query for enhanced efficiency. This design allows it to match the performance of a 40 billion parameter dense model while achieving speeds over 200 tokens per second on H20 hardware. The model supports a 128K context length and has undergone a three-stage training process including pre-training on medical corpora, supervised fine-tuning, and reinforcement learning. AI

Microsoft researchers have identified a significant limitation in current AI models and agents: their inability to effectively manage long-running tasks. These systems struggle with tasks that require sustained operation or memory over extended periods. This deficiency impacts their potential for complex, multi-stage operations and highlights an area for future AI development. AI

Two new research papers explore efficient pre-training methods for large language models. The first paper compares dense and sparse Mixture-of-Experts (MoE) transformer architectures at a small scale, finding that MoE models improve validation loss when matching active parameters but do not surpass dense models at equal total parameter capacity. The second paper investigates various low-rank pre-training techniques, demonstrating that even when validation perplexity is similar, these methods converge to geometrically distinct solutions and do not fully replicate the generalization or internal representations of full-rank training. AI

Researchers have established a new theoretical sample complexity guarantee for off-policy actor-critic methods in reinforcement learning. The paper proves the first $\tilde{\mathcal{O}}(\epsilon^{-2})$ sample complexity for finding an $\epsilon$-optimal policy under minimal assumptions, specifically requiring only an irreducible Markov chain. This achievement contrasts with prior work that necessitated nested-loop updates or stronger, algorithm-dependent policy assumptions. AI

Researchers have introduced Rescaled Asynchronous SGD (ASGD), a novel method for optimizing distributed machine learning models under heterogeneous conditions. This approach addresses the bias in standard ASGD that arises when faster workers contribute more updates, by rescaling worker-specific stepsizes. The method theoretically guarantees convergence to the correct global objective and matches the known lower bound for time complexity in the non-convex setting. AI

Researchers have introduced Delight-gated exploration (DE), a novel algorithm designed to optimize decision-making in scenarios with vast action spaces. DE prioritizes exploratory actions based on their potential "delight," a metric combining expected improvement and surprisal, rather than broadly searching until uncertainty is resolved. This approach aims to be more efficient than traditional methods like ε-greedy, especially when exploration budgets are limited. The algorithm has demonstrated consistent performance across various bandit and MDP problems, showing reduced regret compared to Thompson Sampling and ε-greedy. AI

Researchers have developed a new framework called operator-adaptive calibration to streamline the selection of spectral preprocessing methods in near-infrared spectroscopy (NIRS). This approach integrates preprocessing selection directly into the calibration model, reducing the need for costly and time-consuming external pipeline searches. The new models offer faster, more robust, and auditable NIRS method development by producing traceable operator choices and retaining interpretable coefficients. AI

Researchers have developed a unified framework for analyzing the generalization capabilities of Physics-Informed Neural Networks (PINNs). This new approach relaxes previous restrictive assumptions and uses Taylor expansion to represent differential operators as linear operators in a high-dimensional space. The analysis reveals that while high-rank networks can generalize well, the nonlinearity of differential operators significantly impacts and potentially enlarges generalization bounds. AI

Researchers have developed a new adaptive algorithm for identifying multiple change points in data under bandit feedback. This algorithm aims to precisely locate discontinuities in a piecewise-constant function with minimal samples. The study establishes theoretical bounds on the algorithm's sample complexity, revealing that it depends not only on the magnitude of the jumps but also on the relative positions of these change points. AI

Researchers have developed a new likelihood-free transport filtering method that leverages couplings between state and observation variables. This approach reformulates the filtering analysis step as a minimization of the maximum mean discrepancy (MMD) between true and approximated joint measures. The method offers an analytic computation for the transport map, avoiding particle collapse and accurately approximating non-Gaussian filtering posteriors, with demonstrated superior performance in nonlinear, non-Gaussian scenarios. AI

Researchers have developed a new generative model for time series data that exhibits approximately periodic behavior. This model utilizes a Gaussian Process (GP) with a novel kernel to effectively capture both the common structure across repetitions and the subtle variations between them. The approach decouples intra-repetition dynamics from inter-repetition variability, enabling the generation of realistic synthetic trajectories. AI

Researchers have developed a theoretical framework to understand and quantify "hallucinations" in AI models used for inverse problems, such as medical imaging. The study shows that these realistic but incorrect details can stem from the inherent ill-posed nature of the problem itself, not just specific models. The new approach provides computable bounds on hallucination magnitudes and algorithms to assess reconstruction faithfulness, demonstrating broad applicability across various imaging tasks and modern generative models. AI

Researchers have developed a novel algorithm-agnostic approach for time series clustering using amortized neural inference. This method trains neural networks to approximate optimal partitioning rules from simulated data, reducing reliance on traditional clustering techniques. The framework leverages statistical features to learn a data-driven affinity structure, enabling automated determination of cluster numbers and achieving competitive or superior accuracy compared to existing methods, with a demonstrated application in financial time series analysis. AI

Researchers have developed new Determinantal Point Processes (DPPs) using wavelets to improve minibatch generation for machine learning tasks. These novel DPPs offer provably better accuracy guarantees and a general method to convert continuous DPPs into discrete kernels suitable for subsampling. This approach enhances variance reduction and computational efficiency, expanding the applicability of DPP-based methods to objective functions with low regularity. AI

Researchers have introduced a novel approach to density estimation in high-dimensional spaces by leveraging pre-training, a technique common in advanced AI. This method utilizes a pre-trained neural network to suggest suitable location-adaptive kernels for each data point, thereby improving efficiency and accuracy. The effectiveness of this strategy is demonstrated in numerical experiments, particularly when the target distribution aligns with the pre-training distribution, with options for fine-tuning to adapt to different distributions. AI

A new research paper introduces a "channel-transition" framework to explain why large language models struggle to maintain context and instructions over extended multi-turn conversations. The study proposes the Goal Accessibility Ratio (GAR) as a metric to quantify the degradation of attention to key instructions. Researchers found that while attention to instructions may close, relevant information can persist in residual representations, leading to varied failure modes across different model architectures. AI

Researchers have explored the effectiveness of generative meta-continual learning for spoken word classification across multiple languages. Their findings indicate that while multilingual models perform best, the performance differences between models trained on various language combinations are surprisingly small. The amount of unique training data appears to be a more significant factor in performance than the number of languages included. AI

Researchers have developed a new statistical method to determine when AI workflows should release their outputs, particularly for systems that use iterative generate-evaluate-revise loops. This "always-valid release wrapper" addresses the challenge of making release decisions with adaptively generated evaluator scores, where traditional calibration models are unavailable. The proposed wrapper creates a reference pool of failures to calibrate scores and uses an e-process for validity, aiming to control the probability of releasing on infeasible tasks while still allowing for releases on feasible ones. AI

Researchers have theoretically analyzed the mechanism of weak-to-strong generalization, a method for aligning advanced AI systems. Their work, focusing on reward-model learning with two-layer neural networks, demonstrates how a strong model can efficiently learn a new task by eliciting its pre-trained knowledge without catastrophic forgetting. This approach establishes that the strong model acquires target feature directions through this training process, preserving its general capabilities. AI

Researchers have published a paper detailing the use of digital twins as synthetic control arms in single-arm clinical trials. These advanced machine learning models can generate personalized predictions of disease progression, offering a more robust alternative to traditional methods. The paper discusses how these digital twins can overcome limitations of existing synthetic control approaches and provides guidance on their practical deployment, including considerations for FDA draft guidelines on AI in drug development. AI

Researchers have developed new strategies for training surrogate models by integrating data from multiple sources, including simulations and real-world measurements. One approach involves training separate models for each data type and then combining their predictions, while another trains a single model incorporating both data types. These hybrid methods aim to improve predictive accuracy and coverage, and to identify potential issues within existing simulation models, ultimately aiding in system understanding and future development. AI

Researchers have developed a new diagnostic tool to assess the reliability of confidence thresholding in pseudo-labeling pipelines for regression tasks. This method provides a way to predict the bias introduced by thresholding calibrated classifier scores, using the residual score variance on unlabelled data. The proposed $(V^{*}, \kappa)$ decision rule aims to help practitioners determine when confidence thresholding is a safe practice. AI

Researchers have introduced ISOMORPH, a novel digital twin designed for supply chain logistics, addressing a gap in existing time-series forecasting benchmarks. This simulator offers a configurable, multi-echelon network with interpretable parameters, allowing for realistic dataset generation and the study of phenomena like the bullwhip effect. Initial evaluations show that several foundation models, including Chronos and TimesFM, perform comparably to existing benchmarks when used with ISOMORPH, demonstrating its utility for both simulation and model evaluation. AI

Researchers have published a paper detailing a new method for extracting task-specific representations from generalist AI models. The work establishes theoretical guarantees for identifying and disentangling relevant latent information without requiring interventions or specific model structures. This approach aims to provide a provable foundation for moving from broad, generalist models to more specialized and efficient ones for downstream applications. AI

Researchers have developed FPILOT, a framework that enhances reinforcement learning agents for trading by incorporating price forecasts at inference time. This approach, inspired by Model Predictive Control, allows agents to optimize their trading strategies based on predicted future price trajectories without requiring retraining. Evaluations on the TradeMaster DJ30 benchmark demonstrated consistent improvements in total return and risk-adjusted metrics across various policy learning algorithms. AI

Researchers have established new theoretical bounds for training Kolmogorov-Arnold Networks (KANs), a structured alternative to standard MLPs. The work analyzes KANs trained with mini-batch stochastic gradient descent (SGD), including differentially private variants with correlated noise. These findings reveal a gap between non-private and private training regimes, suggesting that polylogarithmic network width is necessary for differential privacy. AI

Researchers have developed a new framework for robust sequential experimental design in A/B testing, specifically addressing challenges posed by model misspecification. This approach aims to improve sample efficiency by bounding the worst-case mean squared error of estimated treatment effects. The framework's effectiveness has been demonstrated through both synthetic data and real-world datasets from a major technology company. AI

Researchers have introduced Pion, a novel spectrum-preserving optimizer designed for training large language models. Unlike traditional additive optimizers like Adam, Pion utilizes orthogonal transformations to update weight matrices, maintaining their singular values and spectral norm. This approach offers a stable and competitive alternative for both LLM pretraining and finetuning, as demonstrated by empirical results. AI

Researchers have developed a new proximal gradient algorithm designed to sample from composite log-concave distributions. This algorithm assumes access to gradient evaluations for one part of the distribution and a restricted Gaussian oracle for the other. The proposed method achieves state-of-the-art iteration counts for sampling, matching previous results for simpler cases and extending to non-log-concave distributions and non-smooth functions. AI

Researchers have developed a new model-based bootstrap method for controlled Markov chains, particularly useful in offline reinforcement learning scenarios where the data-generating policy is unknown. This technique establishes distributional consistency for transition estimators and extends to policy evaluation and recovery, providing asymptotically valid confidence intervals for value and Q-functions. Experimental results on the RiverSwim problem demonstrate that the proposed confidence intervals offer improved calibration and coverage compared to existing methods, especially with limited data. AI

Researchers have introduced MedHopQA, a new benchmark designed to evaluate the multi-hop reasoning capabilities of large language models in the biomedical domain. This benchmark consists of 1,000 expert-curated question-answer pairs, each requiring information synthesis from two distinct Wikipedia articles, with answers provided in free text. The MedHopQA dataset was presented as a shared task at BioCreative IX, attracting 48 submissions from 13 teams, and highlighted the effectiveness of retrieval-augmented generation strategies for improved performance. AI

Two new research papers introduce advanced conformal prediction techniques to improve the accuracy and efficiency of prediction sets. The first paper, "Multi-Variable Conformal Prediction (MCP)," extends conformal prediction to handle vector-valued score functions, allowing for more flexible prediction set shapes without sacrificing coverage guarantees and eliminating the need for data splitting. The second paper, "Shape-Adaptive Conditional Calibration for Conformal Prediction via Minimax Optimization," presents the Minimax Optimization Predictive Inference (MOPI) framework, which optimizes over a flexible class of set-valued mappings to achieve superior shape adaptivity and more efficient prediction sets, even for complex conditional distributions. AI

Researchers have developed new reinforcement learning (RL) methods to address the qubit allocation problem in quantum computing compilation. Two distinct approaches, CO-MAP and QAP-Router, frame the problem as a combinatorial optimization or dynamic quadratic assignment task, respectively. Both methods leverage RL policies trained on real-world quantum circuit datasets, demonstrating significant reductions in SWAP gate overhead and CNOT gate counts compared to existing compilers. AI

Researchers have developed a new online algorithm for Learning-to-Defer (L2D) methods, designed to handle streaming data and dynamic expert availability. This algorithm is the first of its kind for multiclass classification with bandit feedback and a varying pool of experts. It offers theoretical regret guarantees and has demonstrated effectiveness in experiments on both synthetic and real-world datasets, extending L2D capabilities to more complex, dynamic environments. AI

Researchers are exploring the use of large language models (LLMs) for enhancing safety in air traffic control (ATC) and around non-towered airports. One study proposes a vision-language model approach to analyze radio communications, weather data, and flight trajectories for safety assessments, achieving high F1 scores with open-source models. Another paper introduces a safety-oriented evaluation framework that highlights the critical need for consequence-aware metrics, as standard accuracy measures can mask severe risks in ATC operations. AI

Researchers have developed a new framework for optimal policy learning that addresses combined budget and minimum coverage constraints. The study reveals a knapsack-type structure within the problem, allowing the optimal policy to be defined by an affine threshold rule. Two algorithms, Greedy-Lagrangian (GLC) and rank-and-cut (RC), are proposed to implement this approach, with GLC offering close approximation and RC showing near-optimality under specific conditions. AI

Researchers have developed a new method called Self-Supervised Laplace Approximation (SSLA) to directly approximate the posterior predictive distribution in Bayesian models. This approach draws inspiration from self-training techniques and quantifies predictive uncertainty by refitting the model on its own predictions. The SSLA method offers a deterministic, sampling-free approximation that outperforms classical Laplace approximations in predictive calibration for regression tasks, including Bayesian neural networks, while maintaining computational efficiency. AI

Researchers have developed a new method to improve the efficiency of training neural likelihood surrogates for stochastic process models. By augmenting the standard loss function with exact score information and adaptive weighting, the approach significantly reduces the computational cost associated with parameter inference. This technique demonstrates improved surrogate quality and can achieve performance comparable to a tenfold increase in training data with only a marginal increase in training time. AI

Frontier AI models are showing a rapid increase in their ability to handle complex tasks, with their reliability doubling every 4.7 months, a rate that has accelerated since late 2024. Recent models like Claude Mythos Preview and GPT-5.5 are outperforming these trends, though their exact capabilities are still being measured due to near-perfect success rates on current benchmarks. This rapid progress challenges existing testing methodologies, as models are pushing the limits of token capacity and agent scaffolding, making it difficult to accurately assess their performance and potential deterioration at scale. AI

Researchers have developed a novel physics-informed generative framework to model yield curve dynamics, addressing the conflict between deep learning's flexibility and fixed-income modeling's theoretical constraints. The proposed two-stage architecture, featuring a Student-t Conditional Variational Autoencoder with Dynamic Level Injection (CVAEsT+LS) and a Neural Stochastic Differential Equation penalized by a No-Arbitrage PDE, significantly reduces forecasting errors. This approach demonstrates superior performance in predicting term structures across various macroeconomic regimes and currencies, outperforming traditional models like HJM. AI

Researchers have developed a new theoretical framework for neural operators, a type of AI model used to learn solutions for complex systems like partial differential equations. This work specifically addresses the approximation analysis for nonlinear reaction-diffusion systems, which are crucial for modeling pattern formation. The study establishes explicit error bounds and demonstrates that their proposed Laplacian eigenfunction-based architecture can significantly reduce the parameter complexity required for accurate predictions. AI

Researchers have developed two new frameworks for improving 3D hand pose estimation from egocentric camera views. EgoForce utilizes a differentiable forearm representation and a unified transformer to achieve state-of-the-art accuracy across various camera types, reducing MPJPE by up to 28%. EgoEV-HandPose, on the other hand, employs stereo event cameras and a novel KeypointBEV fusion module to jointly estimate bimanual hand poses and recognize gestures, achieving an MPJPE of 30.54mm and 86.87% gesture recognition accuracy. Both methods aim to enhance applications in AR/VR and human-computer interaction by providing more robust and accurate hand tracking. AI

Researchers have introduced Random-Set Graph Neural Networks (RS-GNNs) to address uncertainty quantification in graph learning. This new framework models node-level epistemic uncertainty using a belief function formalism. Experiments on nine datasets, including autonomous driving benchmarks, show RS-GNNs offer improved uncertainty estimation capabilities. AI

Researchers have introduced Quantized Diffusion Schrödinger Bridges (QDSB), a novel method for learning generative models from unpaired data. QDSB addresses the computational challenges of traditional Schrödinger bridges by quantizing endpoint distributions and using cell-wise sampling to reconstruct the data plan. This approach significantly reduces training time while maintaining sample quality comparable to existing methods. AI