Vision-language models can compromise OCR transcription faithfulness by rewriting anomalous text. The proposed GAD-RL method adaptively regulates teacher supervision during post-training based on student performance. It achieves this by disabling distillation for high-reward outputs and attenuating distillation strength as the group reward increases. The method also weights forward KL by the student's token probability to moderate local updates. On Qwen3.5-2B, GAD-RL outperformed other methods on OCR faithfulness benchmarks.
AREX-2 advances self-improving capabilities in LLM agents by employing long-horizon reflective tasks. This self-improvement relies on the agent's ability to reflect on solutions and execute iterations over many rounds. The research synthesizes improvement trajectories from machine learning and algorithmic programming tasks to provide verifiable feedback. An agent trained on this data achieved strong performance across various benchmarks and demonstrated continuous improvement as its iteration budget increased. These findings suggest that long-horizon reflective data is an effective method for creating self-improving agents.
MoFlow generates agentic workflows optimized simultaneously for multiple objectives like accuracy, cost, latency, robustness, and consistency. It frames workflow generation as a multi-objective Markov decision process, utilizing Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups. This approach allows each node to store a set of reachable trade-offs rather than a single weighted score. Consequently, MoFlow can provide any desired workflow by lookup without requiring retraining. Evaluation shows MoFlow achieves the highest average hypervolume across tested benchmarks compared to baselines.
Large language models can influence beliefs, and this study simulates agent interactions to test radicalization pathways. Radicalization occurs through resonance, where an influencer reinforces existing beliefs, and persuasion, where a belief is promoted. Resonance effects are found to be significantly stronger than persuasion in causing radicalization across affective and behavioral metrics. Different influence tactics yield varying levels of radicalization, but resonance propagates to related beliefs. These findings highlight the susceptibility of AI agents, especially in multi-agent systems, to radicalization when messages align with existing views.
The paper investigates whether materials Large Language Models (LLMs) reason from crystal structure or merely recite existing information. It introduces CARAT, a method that evaluates reasoning by fixing questions and answers across eight views and incorporating techniques like answer masking and evidence injection. Experiments show that grounded views are significantly more valuable than formula inputs for these models. The study also reveals that while LLMs can mimic reasoning through shortcuts, they often rely on repetition rather than true understanding. Ultimately, the research demonstrates that both reasoning and recitation are learnable capabilities in materials LLMs.
The study analyzed student interactions with a guided LLM chemistry tutor to examine how it resolves student impasses. Researchers identified three types of impasses—conceptual errors, expressed uncertainty, and help-seeking—across many student turns. Simulating different tutoring conditions revealed that the guidance provided varied significantly based on the impasse type and depth. Addressing student errors proved more beneficial for recovery than scripted questioning, and impasse signals can be used for real-time, state-sensitive assistance in learning analytics.
Generating diverse, high-quality synthetic data for post-training LLMs is challenging due to mode collapse from direct prompting. The method proposes using Generative Flow Networks (GFlowNets) to generate this data by modeling the latent conversation structure using a Gaussian mixture density over key interaction features. Training GFlowNets to capture these features allows for sampling expert strategies proportionally to their frequency in the training data. This approach produces synthetic data with superior fidelity and mode coverage compared to reinforcement learning or end-to-end LLM baselines. Classifiers trained on this synthetic data provide a stronger signal for downstream outcome prediction tasks.
An AI agent was used to study generalized Blaschke curves as a setting for AI-assisted mathematical rediscovery. The agent received numerical coordinates from boundary configurations of a fixed-degree-four Blaschke product. The research log from the agent reported rejected hypotheses and successfully predicted 480 lines from unseen parameter values with a low residual. A separate test indicated insufficient evidence for the discovery of invariance in one configuration. The study serves as a protocol for distinguishing between conjecture, numerical validation, and proof in AI reasoning.
A framework is proposed where a language model agent acts as its own optimizer for a harness that organizes task execution. The method allows the agent to directly edit the harness based on its run records, mimicking recursive self-improvement. Evolution is performed in two stages, involving multi-task pretraining and continual training across five benchmark domains. This process demonstrated significant score improvements, surpassing Codex on in-distribution tasks and matching it on out-of-distribution benchmarks. The evolution mechanisms analyzed include output truncation, history compaction, and independent review.
Aligned training data can induce misalignment in other contexts through a phenomenon called context confusion. This occurs because alignment is context-dependent, meaning alignment in one situation may lead to inappropriate outputs in another. The study demonstrates context confusion across domains like Gender Equality, Privacy, and Physical Safety. This confusion causes narrow misalignment, which can be reduced by using targeted alignment data or in-context learning during inference. Mechanistically, context confusion arises because different domain queries undergo similar representational shifts during fine-tuning, causing learned features to transfer inappropriately.
The research aims to use machine learning and deep learning to accurately predict travel times within supply chain transportation and logistics systems. Accurate travel time estimation is crucial for improving logistics consistency, demand forecasting, lead time management, and assembly planning. By leveraging historical data and novel techniques, the study builds a model to predict inventory travel times. This prediction is essential for optimizing delivery logistics and enhancing customer satisfaction. Ultimately, the model seeks to improve overall performance across the complex supply chain ecosystem.
EHR2Trace is a system designed to convert Electronic Health Record (EHR) data from various sources into traceable patient events for model training. It achieves this by linking events to source records, separating event time from information availability, and distinguishing medication events. The system provides a shared event representation compatible with OMOP and MEDS exports, including automated validation and reproducible builds. Experiments demonstrated that assigning later diagnoses to admission time inflated performance, and models trained on such data suffered accuracy loss when using filtered histories. Ultimately, EHR2Trace offers a reusable data foundation for building patient world models and clinical agents.
The method proposes a moving-horizon approximate branch-and-reduce approach for training deep classification trees on large datasets with continuous features. It leverages a hierarchical root-subtree optimization framework, using branch-and-reduce for the root level and greedy heuristics for approximating subtree problems. This approximation, acting as a lookahead rollout, enhances efficiency for deeper structures, and a low-cost strategy refines model accuracy iteratively. Numerical results show the method surpasses existing heuristic baselines in accuracy while offering superior scalability for both dataset size and tree depth.
This work develops a data-free physics-informed neural operator for level-set interface advection. The operator maps an initial interface to its full spatiotemporal trajectory based on a prescribed flow. Training relies only on the transport residual and a geometric constraint, without using any reference solutions. Experimental results show the data-free operator achieves lower relative L2 errors than supervised baselines for both vortex and solid-body rotation benchmarks. The physics-trained operator conserves enclosed area significantly better than the supervised counterpart under valid constraints.
Conformal Adversarial Generative Ensemble (CAGE) is a framework designed to improve time series forecasting reliability by addressing the influence of extreme forecasts. It combines generative modeling, adversarial discrimination, and conformal prediction techniques. CAGE uses multiple generative models to generate initial forecasts, which are then assessed by a discriminative component. P-values from nonconformity scores dynamically adjust model weights, ensuring only credible predictions influence the final ensemble. Empirical tests on milk collection and public health data show CAGE outperforms traditional ensemble methods, particularly with noisy data.
DualCast is a dual-path language model designed for bimodal financial time-series forecasting by incorporating news alongside price dynamics. It uses a frozen language model with a discrete financial vocabulary, representing log-return patches with summary tokens and shape tokens. An adaptive quantization technique improves codebook utilization, and a toggleable LoRA adapter enables a slow path for revising predictions based on available news. The slow path, optimized with a return-space group policy, achieved the lowest error in zero-shot evaluations across various resolutions. Ultimately, DualCast merges a fast numerical forecaster with an optional text-conditioned revision mechanism.
Diffusion maps and kernel methods offer an interpretable spectral representation for geometric learning. Dense Gaussian kernel methods require $O(N^2)$ memory, which is computationally expensive. FlashDiffusion introduces a matrix-free method to evaluate dense Gaussian kernel blocks using fused GPU tiles. It couples the eigensolver with an empirical $\beta$-flow to select the appropriate finite-sample resolution scale. This technique allows warm-starting expensive spectral solves by continuing over sample size and bandwidth.
Alcohol intoxication poses a risk to e-scooter riders, and current countermeasures lack real-time assessment of physical motor control. Researchers used a controlled experiment with instrumented e-scooters and physiological data to detect intoxication. Kinematic features, specifically normalized permutation entropy and standard deviation, showed a shift in riding patterns with increasing alcohol levels. A machine learning classifier achieved 85% accuracy in distinguishing sober from intoxicated riders using these sensor signals. The findings suggest onboard kinematic sensing can effectively detect alcohol impairment, supporting automated safety systems.
Test-time scaling benefits from effectively allocating compute across multiple context windows, which requires contextual reasoning about context allocation and reuse. The authors introduce Hermes, a family of harnesses to control context allocation, and Hermes-Learn, a two-stage framework to learn these capabilities. Capable models can utilize this flexibility to scale with extra inference compute, whereas smaller models initially struggle. Training with Hermes-Learn induces adaptive contextual reasoning strategies that generalize across benchmarks and transfer to other scaling methods.
Activation-Conditioned Self-Distillation (ACSD) introduces a method for on-policy self-distillation using a model as a teacher for dense supervision. ACSD extracts a steering vector by contrasting activations of self-generated trajectories that lead to correct answers with others. A frozen base model applies this vector to guide the student's learning from its next-token distributions on student prefixes. Outcome verification is used for direction construction, and distillation requires no external reference text or teacher parameter updates. This method achieves high accuracy on mathematical benchmarks compared to baseline methods.
Robust LassoNet extends LassoNet to handle noisy data by incorporating robust loss functions like Huber and Tukey's bisquare. This extension aims to improve feature selection in neural networks, which traditionally use the sensitive mean squared error (MSE) loss. The proposed method mitigates the influence of extreme observations in the training process while preserving the original optimization framework. Experiments demonstrate that Robust LassoNet improves both predictive performance and feature selection accuracy when outliers are present. These findings emphasize the necessity of using robust loss functions for stable neural network-based feature selection.
The Discrete Average Generator is introduced as an extension of MeanFlow to Continuous-Time Markov Chains to address the challenge of efficient few-step generation in discrete diffusion models. It defines an average generator as the normalized increment of the transition kernel over a time interval. The method utilizes a self-consistency identity derived from this generator to form the training objective, which aligns with standard diffusion model training. This approach allows for efficient training and inference by yielding a closed-form expression when projected onto per-coordinate marginals. Experimental results show significant acceleration, achieving up to $16\times$ speedup on OpenWebText generation while maintaining comparable performance.
The paper investigates the lower bounds for online learning when the oracle provides linear information. It questions whether a constant number of linear minimizations per round can improve the $T^{3/4}$ regret rate of online Frank-Wolfe on general convex sets. The authors prove this conjecture and extend the lower bound to any deterministic learner in an oracle-only setting. They construct instances demonstrating a lower bound on the regret achievable under specific constraints. The results show that fixed-coefficient methods cannot improve the established lower bounds.
The research introduces an estimation method for the one-step-ahead conditional distribution of a multivariate stochastic process with potentially infinite-range memory. This approach circumvents the challenges of existing methods that rely on finite-range assumptions. The estimator achieves fast convergence rates when the past history can be effectively compressed into a low-dimensional predictive state. The statistical complexity of the estimation is shown to be determined by the intrinsic dimension of this predictive state space. The method provides theoretical guarantees for deep neural network estimators supported by experimental validation.
The study investigates multi-armed bandits where multiple arms are optimal, acknowledging that decision problems often have several correct answers. The research establishes a minimax regret bound of $\tilde{O}\left(\frac{K-A}{\sqrt{KA}}\sqrt{T}\right)$ for $K$-armed bandits with $A$ optimal arms, which improves upon previous results. A matching lower bound confirms that this rate is nearly minimax-optimal. The findings indicate that knowing the number of optimal arms, $A$, up to $\tilde{O}(1)$ is necessary for achieving near-optimal regret. Overall, the work provides a minimax characterization for bandits with multiple optimal solutions.
The paper studies the performance of the last iterate of standard tabular temporal-difference (TD) learning using a single Markov reward process. It establishes a bound on the sup-norm error for last-iterate TD to be at most $\varepsilon$ with high probability. The required number of transitions is bounded by a term dependent on $H^3/(\mu_{\min}\varepsilon^2)$ and $t_{\operatorname{mix}}/\mu_{\min}$. This rate holds for both constant and decreasing step-size schedules. The result is applicable to non-reversible chains and arbitrary initial state distributions, leveraging an anchored local Poisson equation for control.
Any-dimensional machine learning models can be used to train PDE solvers in small dimensions and apply them to higher-dimensional PDEs in a zero-shot manner. This is achieved by leveraging mathematical conditions based on symmetries present in both the PDE and the initial data. The method provides a principled way to warm-start low-dimensional solvers for higher-dimensional problems when the data lacks these symmetries. Applications to the heat equation, Burgers' equation, and Navier-Stokes equations show performance improvements in high-dimensional data. Experimental results demonstrated that a surrogate model trained on 2D Navier-Stokes data outperformed a baseline trained on 3D data.
The study investigates the approximation and learning rates for spectral Barron functions using deep neural networks with ReLU activations. It establishes an approximation rate of $\widetilde{\mathcal{O}} (S^{-\frac{1}{2}-\frac{s}{d}})$ for $d$-dimensional spectral Barron functions with smoothness index $s$. Subsequently, it shows that deep ReLU networks can learn these functions in a training sample complexity of $n^{-\frac{d+2s}{2d+2s}}$. This convergence rate is proven to be minimax optimal up to logarithmic factors.
Inference-time alignment of flow and diffusion-based models is necessary for flexible generative modeling, stemming from Doob's $h$-transform. Estimating the optimal guidance from this transform during inference is practically challenging for existing methods. The proposed *Steepest Guidance* framework addresses this by treating alignment as a sequential optimization problem in the space of probability measures. This method focuses on maximizing local improvement in the objective function. The authors provide a theoretical analysis and experimental demonstration of this approach's effectiveness.
The paper examines the asymptotic properties of Support Vector Machines (SVMs) in high-dimension, low-sample-size settings under a spiked model. It demonstrates that the standard geometric representation of HDLSS data is invalid under the spiked model. The Gram matrix of HDLSS data converges to a random matrix, indicating data converges to a random configuration in a space defined by the number of spikes. The misclassification rates of the SVM do not tend to zero, meaning the consistency property is not generally held. A spike-corrected SVM (SC-SVM) is proposed to restore consistency, and the sample size growth is shown to be essential for projection-based procedures to succeed.