Brief

Researchers have developed a neuro-symbolic framework called NSCR to address stereotype-prone reasoning in AI systems designed for educational settings. This framework aims to distinguish between observable evidence and culturally biased interpretations, treating unsupported claims as safety risks. NSCR processes multimodal data, including video, audio, and text, to generate typed facts with provenance and cultural context, enabling executable reasoning and policy enforcement. The paper also proposes a benchmark agenda and metrics to evaluate stereotype leakage, evidence faithfulness, and cultural calibration in classroom AI. AI

A new position paper proposes a shift from low-level, click-based interactions to typed actions for web agents. This approach, termed 'web verbs,' would expose web operations as typed functions with structured inputs and outputs, enhancing reliability and auditability for long-horizon tasks. The authors argue that this semantic layer is crucial for building trustworthy and scalable agentic web systems. AI

Researchers have developed an AI-integrated Learning Management System (LMS) designed for middle school students to provide timely and targeted support. This system aims to offer formative feedback, recommend practice based on mastery, and alert teachers to persistent struggles, addressing the common issue of students receiving help too late. The platform prioritizes privacy with a data minimization approach and auditable logs, and its effectiveness will be studied longitudinally through high school and beyond to assess its impact on learning trajectories. AI

A new tutorial paper explores the integration of blockchain infrastructure with embodied AI systems, focusing on post-quantum security and trustworthy data economies. It highlights the need for crypto-agile architectures to protect data provenance and governance as quantum computing advances threaten current cryptographic primitives. The paper proposes blockchain as a foundational layer for decentralized intelligent environments, offering open-source frameworks for quantum-resistant, interoperable, and data-trustworthy systems. AI

Researchers have developed a new framework called Deep Active Re-Labeling (DAR) to improve the efficiency of active learning in machine learning. This method addresses the issue of human annotation errors, which can significantly degrade active learning performance. DAR strategically re-annotates a portion of already labeled data to identify and correct noisy labels, leading to more data-efficient training and a cleaner final annotation dataset. AI

Researchers have introduced PAFO, a new framework designed to address personalized reward bias in large language models. This bias occurs when reward models, trained on diverse user preferences, disproportionately favor users with more common preferences. PAFO formulates fairness as a Pareto optimization problem, aiming to enhance the experience for under-served users without negatively impacting others. The framework trains specialized models for different user groups and then distills their knowledge into a single model, improving accuracy and fairness across the board. AI

Researchers have developed two novel sampling-free frameworks, PC-SNGP and PC-SNER, designed to enhance the reliability and physical interpretability of probabilistic models for industrial prognostics. These frameworks improve performance by maintaining distance-preserving representations and increasing uncertainty estimates as input data deviates from the training manifold. The methods were validated on rolling-element-bearing prognostics datasets, demonstrating superior prediction accuracy and well-calibrated uncertainty compared to existing approaches, even under adversarial conditions. AI

Researchers have introduced a new framework called Pro-SF to address strategic classification problems where agents deviate from pure rationality due to psychological biases. This framework, grounded in prospect theory, models agents' strategic manipulations by incorporating mechanisms like asymmetric benefit/cost perception, subjective reference points, and probability distortion. Experiments on synthetic and real-world data demonstrate Pro-SF's effectiveness in bridging machine learning and behavioral economics for more reliable real-world applications. AI

Researchers have developed FormalASR, a novel end-to-end system designed to convert spoken Chinese directly into formal written text. This approach bypasses the need for a separate post-editing step by an LLM, reducing latency and computational costs. The system utilizes two models, 0.6B and 1.7B parameters, fine-tuned from Qwen3-ASR, and is trained on newly created large-scale datasets, WenetSpeech-Formal and Speechio-Formal. AI

Researchers have developed DecepGPT, a new system designed to detect deception in multimodal data by analyzing audiovisual cues. The system aims to provide auditable reports by incorporating structured reasoning chains and cue-level descriptions. DecepGPT also introduces a large, multicultural dataset called T4-Deception, featuring over 1600 samples from four countries, to improve generalization across different cultural contexts and prevent shortcut learning. AI

A new research paper argues that studies on anthropomorphic AI misalignment require more rigorous evidence. The paper highlights issues like conceptual ambiguity and weak experimental designs that can lead to overinterpretation of AI behaviors. It proposes a framework of evidence levels and a diagnostic checklist to improve methodological standards in this critical area of AI safety research. AI

Researchers have developed a new framework called VATS to exploit vulnerabilities in how AI models handle tool errors. This method systematically mutates error messages to inject malicious instructions, bypassing standard safety measures. In tests with leading models like Gemini 3.1 Pro and GPT-5.5, this error-path injection technique significantly increased the success rate of prompt injection attacks, reaching up to 100% in some evaluations. While current production safeguards can offer some protection, the underlying susceptibility in the models themselves presents a risk to custom AI agent workflows. AI

Researchers have developed a new method called error-based Predictive Coding (ePC) that significantly speeds up neural network training on digital hardware. Traditional Predictive Coding (PC) methods suffer from signal decay in simulations, hindering their effectiveness with deeper networks. ePC reformulates PC to eliminate this decay, allowing it to achieve performance comparable to backpropagation even on complex models, while running orders of magnitude faster. AI

Researchers have developed a novel deep learning framework called OnlyDense to model complex Lagrangian simulations, which are often computationally intensive. This method represents the system's state as a function evolving in Hilbert space, using learned neural basis functions to create a linear subspace. This approach unifies classical reduced-order modeling with deep learning, allowing for accurate prediction of dynamics even with a reduced number of basis functions, as demonstrated in large-scale simulations. AI

Researchers have identified a key issue in scaling up AI model training data mixtures, termed "repetition mismatch." This occurs when the optimal data mixture changes as training budgets increase due to the varying repetition rates of high-quality, limited datasets. A new subsampling procedure that matches the target repetition rate can accurately predict optimal mixtures from significantly smaller experiments, improving efficiency and accuracy. AI

Researchers have developed new theoretical findings regarding the discrete geometry of ReLU networks, focusing on their connectivity graphs. These graphs, where nodes represent linear regions and edges connect regions sharing a face, demonstrate an average degree upper-bounded by twice the input dimension, irrespective of network depth or width. Furthermore, the graph's diameter has an upper bound independent of input dimension, even as the number of regions grows exponentially. These theoretical results were validated through experiments on networks trained with both synthetic and real-world data. AI

Researchers have developed a new framework called Zero-Expert Self-Distillation Adaptation (ZEDA) to make Mixture-of-Experts (MoE) language models more efficient. ZEDA allows post-trained static MoE models to dynamically skip over half of their experts during inference with minimal accuracy loss. This method was tested on Qwen3-30B-A3B and GLM-4.7-Flash, showing significant reductions in computation and an inference speedup of approximately 1.20x. AI

Researchers have developed a transformer-based recommender system called MedicalRec to help select optimal machine learning models for medical image classification tasks. This system aims to reduce the energy consumption and waste associated with the trial-and-error process of model selection. MedicalRec was evaluated on a new dataset, MedicalRec-Bench, which contains over 5,000 records of models tested across various medical imaging categories, achieving a HitRate@100 of 75.5%. The dataset and code are publicly available. AI

Researchers have developed a new deep learning framework for 3D intraoral reconstruction, aiming to improve vertex distribution in predicted point clouds. While the previous model achieved 77.49% accuracy, it suffered from vertex clustering. The updated model introduces Hungarian matching and Repulsion Loss to create a more uniform vertex distribution, though this resulted in a lower accuracy of 68.02%. Despite the numerical decrease, the new approach significantly alleviates the vertex clustering issue, leading to more evenly spread vertices across the reconstructed surface. AI

Researchers have developed CARE, a novel post-hoc safety layer for medical summarization using large language models. This model-agnostic system overlays calibrated flags for omissions and hallucinations without requiring model retraining. CARE provides formal guarantees on error rates, aiming to balance safety with the burden on clinicians reviewing summaries. AI

Researchers have developed a new framework called Mobility-Embedded POIs (ME-POIs) to improve geospatial representations of locations. This framework integrates human mobility data with language model embeddings to better understand how places are used, going beyond static textual descriptions. ME-POIs encodes individual visits and aligns them with learnable POI representations through contrastive learning, effectively capturing usage patterns. The system also includes a mechanism to address data sparsity by propagating temporal visit patterns from frequently visited nearby locations. AI

Researchers have introduced Q-RACL, a novel framework for constraint learning that prioritizes repair over immediate veto of infeasible candidates. This approach accepts a candidate if a repair plan can restore feasibility and value, otherwise providing structured rejection credit. The framework specifically targets scenarios where the repair-feasibility inference is hidden, such as in discrete logarithm problems, making it accessible to quantum agents through quantum feature access. AI

Researchers have developed a highly accurate image classification system for historical documents, capable of distinguishing between text, tables, and graphics. Fine-tuned deep learning models, specifically RegNetY-16GF and ViT-large, achieved over 99% accuracy on a dataset of over 48,000 scanned pages. This system is designed to facilitate content-specific processing in large-scale digitization projects, with the models, dataset, and software made publicly available under open-source licenses. AI

A new paper argues that the current dominant method for training large language models (LLMs), which involves extensive post-training stages like supervised fine-tuning (SFT) and reinforcement learning (RL), is essentially a return to older "pre-train then fine-tune" approaches. The authors demonstrate that models trained from scratch on modern reasoning datasets can achieve significant performance on competitive benchmarks, suggesting that current post-training primarily serves to fit models to specific distributions rather than fostering general capabilities. They propose a shift towards training procedures that emphasize "learning how to learn" to develop more generally capable models. AI

Researchers have developed SmartMixed, a new two-phase training strategy that enables neural networks to learn optimal activation functions for individual neurons. The first phase uses a differentiable mixture mechanism for neurons to select from a pool of candidate functions, while the second phase fixes these selections for computational efficiency. Experiments on the MNIST dataset with feedforward networks showed that neurons in different layers develop distinct activation function preferences, outperforming models with a single fixed activation function. AI

Researchers have developed a new verification protocol called TAO (Tolerance-Aware Optimistic Verification) designed to ensure the integrity of floating-point neural network computations, particularly in cloud-based ML services. TAO addresses the challenge of nondeterministic floating-point execution across different hardware by accepting outputs within principled acceptance regions rather than demanding bitwise equality. The system combines theoretical worst-case bounds with empirical percentile profiles and uses a dispute game to recursively narrow down discrepancies to individual operators, making verification scalable and practical for real-world ML models. AI

Researchers have developed a novel deep neural network approach for intelligent character recognition in handwritten forms. This method integrates character detection and classification into a single task, outperforming traditional two-task methods. The system achieved an 88.28 percent recognition rate on real exam data, though limitations with the EMNIST dataset were noted. AI

Researchers have developed scCBGM, a novel framework for interpretable single-cell counterfactual editing using concept bottleneck generative models. This approach adapts concept bottleneck architectures for single-cell data, incorporating decoder skip connections and a cross-covariance penalty to enhance disentanglement. The framework has been extended to flow matching models, allowing for concept-guided editing in both encoding-decoding and generation scenarios, and includes a new synthetic benchmark for evaluation. AI

Researchers have developed a new Test-Time Adaptive (TTA) composition framework designed to improve the effectiveness of Machine Learning as a Service (MLaaS) in dynamic Internet of Things (IoT) environments. This framework addresses challenges with existing adaptive methods by introducing a TTA-aware composability model to ensure service compatibility and a service-level adaptation model to adjust individual services during inference. Experiments show this approach significantly reduces computational time compared to traditional methods. AI

Researchers have developed SNR-ST-Mix, a novel data augmentation framework for spatial transcriptomics imputation using deep neural networks. This method addresses limitations in current augmentation strategies by ensuring that mixed samples preserve local biological structure and spatial smoothness. Experiments show that SNR-ST-Mix outperforms existing methods without increasing computational complexity, leading to improved prediction performance. AI

Researchers have developed a new framework called Item Response Scaling Laws (IRSL) that integrates Item Response Theory with language model scaling laws. This approach aims to make the estimation of scaling laws more efficient and generalizable by disentangling model ability from question characteristics, reducing the complexity from O(M x N) to O(M + N). IRSL uses empirical response probabilities from LMs, such as token probabilities or pass rates, to derive more reliable scaling estimates with significantly fewer questions, enabling accurate performance forecasting across different benchmarks. AI

Researchers have developed NeuroAlign, a novel hierarchical framework designed to fuse dynamic and structural neuroimaging data for the analysis of Mild Cognitive Impairment (MCI). The system employs dual-modal hierarchical alignment to model multi-scale connectivity and align functional-structural embeddings, alongside dual-domain hierarchical interaction for fine-grained feature modulation. NeuroAlign also includes Synergistic Activation Mapping, a gradient-free attribution method for inspecting model-derived brain patterns, and has demonstrated competitive performance on multiple datasets. AI

Researchers have developed an agent-guided multi-fidelity framework to improve the accuracy of simulating electronic and optical properties in nanomaterials. This new approach addresses computational challenges like numerical instabilities and convergence failures inherent in demanding calculations. By assigning confidence weights and using high-accuracy reference points, the framework corrects artifacts and enhances agreement with experimental data, proving transferable to various optoelectronic nanomaterials. AI

Researchers have developed a new anomaly detection scoring scheme called PAI, designed to address the limitation of amplitude-agnostic embeddings in existing representation-based methods. PAI incorporates a diagnostic module to assess amplitude information capture and a score augmentation function that fuses representation scores with median deviation and local mean-shift scores. This approach significantly improves performance on datasets like TSB-AD-U-Eva and TAB UV, with one combination outperforming the state-of-the-art by 15%. The findings highlight the importance of retaining amplitude information in time-series anomaly detection. AI

Researchers have developed SafeECGMatch, a novel semi-supervised learning framework designed for electrocardiogram (ECG) classification. This method addresses the challenge of limited labeled data in clinical settings by effectively handling unlabeled data that may contain out-of-distribution anomalies. SafeECGMatch utilizes a dual-branch architecture to extract time-frequency representations and incorporates adaptive calibration techniques to ensure reliable OOD rejection and accurate pseudo-labeling. AI

Researchers have developed OSMGraphCLIP, a novel model that learns global location representations using OpenStreetMap data. This model encodes geographic environments as graphs, capturing topological and semantic relationships between features like roads and buildings. OSMGraphCLIP demonstrates strong performance across various downstream tasks, including climate, ecology, and public health, often matching or surpassing satellite-based methods, particularly for socioeconomic and health-related predictions. AI

Researchers have identified a flaw in self-supervised learning methods like JEPA, where contrastive objectives can mistakenly encode slowly varying noise instead of the actual dynamics of a system. This leads to representations dominated by trajectory-specific noise, hindering downstream performance. The study proposes a solution: sampling negative examples within a single trajectory rather than across trajectories, which forces the model to learn relevant dynamics and improves representation quality even with strong noise. AI

Researchers have developed SurfDesign, a new framework for protein design that focuses on molecular surface geometry and physicochemical properties. This method integrates continuous geometric manifold modeling of surfaces with protein language models. SurfDesign reportedly outperforms existing surface-conditioned and backbone-only approaches in designing novel binders and enzymes, and also shows strong performance in inverse-folding tasks. AI

Researchers have developed a novel topological framework to analyze and compare trained Graph Neural Networks (GNNs). This method maps the induced Stochastic Block Models onto the unit n-sphere, creating a low-dimensional "fingerprint" of the GNN. This fingerprint can be used for visual inspection, nearest-neighbor searches, and identifying candidates for transfer learning without requiring retraining. AI

Researchers have developed a new framework for anti-money laundering (AML) transaction monitoring that leverages large language models (LLMs) for improved explainability and accuracy. This system treats triage as an evidence-constrained decision process, combining retrieval-augmented evidence bundling with LLMs that provide structured outputs and explicit citations. The framework also incorporates counterfactual checks to validate decisions and rationales against plausible perturbations, aiming to reduce hallucinations and enhance auditability in regulated workflows. AI

Researchers have introduced HA-VLN 2.0, a new benchmark designed to evaluate how well AI agents can navigate in environments with dynamic human interactions. This benchmark includes a standardized task with metrics for goal accuracy and personal-space adherence, along with a dataset and simulators that model multi-human scenarios. Initial tests show that current leading agents struggle significantly in these complex, socially aware situations, highlighting the need for explicit social modeling in navigation systems. AI

Researchers have developed Larch, a new framework designed to optimize the execution of semantic filters within AI SQL queries. Larch addresses the high inference costs and latencies associated with semantic operators, which treat AI-generated filters as black boxes, hindering traditional optimization. The framework utilizes embedding-augmented neural networks and supervised learning models to predict filter selectivities and determine optimal evaluation orders, significantly reducing token usage. AI

Researchers have developed MEnvAgent, a framework designed to automate the creation of executable software engineering environments across multiple programming languages. This system addresses the scarcity of verifiable datasets for training AI agents by employing a Planning-Execution-Verification architecture and an environment reuse mechanism to reduce computational costs. Evaluations on the MEnvBench benchmark showed MEnvAgent improved task completion rates by 8.6% and reduced time costs by 43%, also enabling the creation of the largest open-source polyglot dataset for verifiable Docker environments. AI

Researchers have introduced DIVERGE, a new retrieval-augmented generation (RAG) framework designed to enhance diversity in responses for open-ended information-seeking tasks. Unlike traditional RAG systems that assume single correct answers, DIVERGE iteratively explores diverse viewpoints and uses diversity-aware retrieval to improve the quality-diversity trade-off. Experiments show DIVERGE can double response diversity without sacrificing quality, addressing a key limitation in current RAG systems. AI

Researchers have developed a new method for improving automatic speech recognition (ASR) in noisy environments. This technique, called intelligibility-guided observation addition (OA), fuses noisy speech with enhanced speech to boost recognition accuracy. Unlike previous methods that required training, this new approach is training-free, deriving fusion weights directly from the ASR system's intelligibility estimates, which enhances its generalization capabilities. AI

Researchers have introduced STAR, a novel approach to Mixture-of-Experts (MoE) routing that treats routing as a structure-aware subspace learning problem. Unlike traditional MoE methods that use limited linear projections, STAR incorporates an evolving principal subspace to track dominant input structures, enhancing routing stability and expert specialization. This method has demonstrated improved performance on language and vision tasks, with potential for further robustness through optional test-time subspace updates. AI

Researchers have developed a new framework called RLVR to improve long-horizon maritime trajectory and destination forecasting using large language models. This approach converts vessel trajectories into semantic textual representations, enabling reinforcement learning to align LLMs with forecasting objectives. Experiments show that LLMs trained with RLVR significantly outperform existing deep learning methods, particularly in predicting destinations accurately, with 4B LLMs demonstrating optimal performance. AI

Researchers have introduced SRT, a new framework for generating high-resolution time series data from lower-resolution inputs. SRT disentangles time series into trend and seasonal components, aligning them with target resolutions using neural representations and cross-resolution attention. A larger version, SRT-large, demonstrates strong zero-shot capabilities, outperforming existing methods across nine datasets. AI

Researchers have developed MMR-GRPO, a novel method to accelerate training for mathematical reasoning models. This approach reweights rewards based on the diversity of model completions, recognizing that redundant outputs offer limited learning value. By prioritizing unique solutions, MMR-GRPO significantly reduces the number of training steps and wall-clock time needed to achieve peak performance, as demonstrated across various model sizes and benchmarks. AI

A new research paper explores how shared AI models used by competing sellers could inadvertently lead to supracompetitive pricing. The study, which uses a duopoly model, suggests that configuring AI for robustness and reproducibility might cause a phase transition to higher prices above a certain output-fidelity threshold. While competitive pricing is stable below this threshold, above it, the AI model can exhibit bistability, leading to either competitive or supracompetitive outcomes depending on the AI's initial propensity. AI