Brief

Researchers have developed two novel deep learning approaches for improving Positron Emission Tomography (PET) image denoising. UniPET utilizes domain generalization and region-aware learning to create a universal model capable of denoising images across various dose reduction factors, addressing issues of style misalignment and over-smoothing. U-TTT employs test-time training with dual-domain adaptation (spatial and frequency) to dynamically adjust model parameters during inference, enabling robust generalization even with unseen dose levels or scanner types. AI

A new benchmark called WorldOlympiad has been introduced to evaluate video-based world models. It assesses physical faithfulness, geometric consistency, and interaction fidelity, going beyond typical metrics like visual quality. The benchmark aims to reveal shortcomings in current models' ability to adhere to physical laws and maintain coherent 3D structures over extended periods. Experiments using WorldOlympiad on state-of-the-art models have exposed significant gaps in their reasoning and interaction capabilities. AI

Researchers have introduced "Next Forcing," a novel multi-chunk prediction framework designed to enhance autoregressive video generation. This method addresses limitations in current models by providing explicit signals about future dynamics, leading to faster training convergence and improved accuracy, particularly at high frame rates. The framework also accelerates inference and demonstrates better adherence to physical laws in generated videos. AI

Researchers have published a paper detailing a generalized universal approximation theorem for neural networks. This new theorem extends previous work by enabling the approximation of not only functions but also their derivatives. The findings are applicable to differentiable maps on infinite-dimensional manifolds and have implications for approximating non-anticipative functionals and path space functionals. AI

Researchers have introduced Topological Neural Operators (TNOs), a new framework for learning operators on cell complexes. TNOs extend existing neural operators by modeling interactions through Discrete Exterior Calculus, allowing for explicit cross-dimensional coupling. This approach respects the geometric properties of physical quantities and can improve accuracy on Partial Differential Equation benchmarks, especially for complex flow problems. AI

Researchers have developed a theoretical framework to understand when optimizing optical front-ends with neural network back-ends improves imaging classification performance. The study found that these gains are most significant under constrained detector readout, such as limited measurements or coarse sampling, by enhancing class separability. However, under full detector readout, conventional lenses perform comparably, and joint optimization offers no empirical advantage. The research also highlights that these optical-neural network co-designs are most effective with low detector noise and when discriminative content is concentrated at lower spatial frequencies. AI

Researchers have developed POTATR, a new lightweight image-to-graph model for extracting tables from documents. This 29 million parameter model significantly outperforms existing methods on the PubTables-v2 benchmark, achieving a GriTS_Con score of 0.964. POTATR is also considerably faster and more cost-effective than current large language models, with its output being spatially grounded for verification and further integration. AI

Researchers have developed a new automated architecture for time-series prediction in volatile cloud-edge environments. This system addresses the "cold start" problem for newly discovered nodes by merging sparse local telemetry data with a high-resolution public dataset called TimeTrack. A Neural Architecture Search engine then generates accurate baseline models, significantly improving forecasting accuracy and convergence speed. AI

Researchers have developed a new method called Intervention-Aware Variational Quantum Differentiable Predictive Control (IA-VQC-DPC) to better measure the safety contributions of AI policies versus their protective layers. This approach trains quantum circuit policies with a budget that penalizes over-reliance on safety filters. Evaluations on building control emulators demonstrated that IA-VQC-DPC significantly reduces pre-filter violations and reliance on safety layers, indicating improved policy-level safety. AI

Researchers have developed SemDINO, a new network designed for semantic change detection in remote sensing imagery. This model integrates a dual-branch encoder using CNNs and frozen DINOv3 features, along with a multi-scale temporal interaction module. SemDINO also incorporates modules for semantic purification and change enhancement to improve accuracy and robustness against pseudo-changes. AI

Researchers have developed Topo-Omni, a novel topographic multimodal model that integrates visual, auditory, and language processing onto a single in-silico sheet. This model, fine-tuned from a pretrained foundation model, demonstrates that a unified spatial principle can organize representations across different modalities and processing stages. The model's clusters align with human neuroimaging findings, and manipulating these clusters selectively impacts perception, offering testable hypotheses about cortical organization. AI

Researchers have developed a novel data synthesis method to create neural machine translation (NMT) models for low-resource Indigenous languages, specifically Q'eqchi' Mayan. By transforming dictionaries into a synthetic corpus and using Parameter-Efficient Fine-Tuning (PEFT) with LoRA adapters on an mT5-base model, they achieved strong structural acquisition. However, the resulting model showed a significant gap in lexical grounding compared to organic language, indicating that while synthetic data is effective for learning grammar, authentic data is crucial for semantic refinement. AI

Researchers have developed advanced machine learning models to predict Alzheimer's disease severity and progression. One approach uses multimodal data, including MRI scans and clinical information, with an ordinal regression framework to improve accuracy and interpretability in staging the disease. Another method introduces a personalized digital twin framework that leverages sparse longitudinal data to model disease transitions, enabling patient-specific trajectory analysis and uncertainty quantification. AI

Researchers have developed ATN3D, a new LiDAR-Radar framework designed for improved 3D object detection in sparse sensing conditions, crucial for autonomous vehicles. The system addresses challenges in long-range detection by employing density-aware early fusion and occupancy-gated aggregation to reduce noise and optimize detection of distant objects. ATN3D demonstrated significant performance gains on the VoD benchmark, particularly in foggy conditions and for objects over 30 meters away, indicating more reliable early detection in challenging environments. AI

A new academic paper analyzes the phenomenon of "AI minidramas" or "fruit dramas," short generative AI video series popular on social media. The research argues these seemingly cute videos perpetuate harmful gendered narratives, associating female characters with transgression and reproductive themes. Furthermore, the paper suggests these narratives can also encode racialization, with the videos' aesthetic serving to mask their ideological content. AI

Researchers have introduced "optical reasoning," a novel approach that utilizes images as the primary medium for AI reasoning, moving beyond traditional text-based methods. This technique involves two variants: typographic-based optical reasoning for compact rationale rendering and graphical-based optical reasoning for structured visual rationales. Experiments show that optical reasoning can match or surpass text-based reasoning in various benchmarks, significantly reducing reasoning tokens and improving token efficiency. AI

Researchers have developed a framework using Large Language Models (LLMs) to check if patient care aligns with clinical guidelines, even when those guidelines aren't in a machine-readable format. This system extracts patient data from unstructured texts and interprets rules from guideline documents to assess compliance. In a study of stroke care at Alessandria Hospital, the LLM-based approach found that over 86% of patient traces adhered to the guidelines. AI

Researchers have developed a new approach to Optical Music Recognition (OMR) specifically for real-world handwritten musical manuscripts. Existing OMR systems struggle with the diversity of manuscripts, which differ significantly from the digital formats they are typically trained on. By using synthetic data generated with fine-grained music notation graph annotations and a tool called Smashcima, the new method shows significant improvement in recognizing complex piano notation, even without in-domain symbol annotation. AI

This article provides a practical guide to fine-tuning large language models like Llama 3 using Parameter-Efficient Fine-Tuning (PEFT) methods, specifically LoRA and QLoRA. It explains that while base LLMs are general, fine-tuning can adapt them for specific tasks, tones, or knowledge. LoRA achieves this by training only a small set of adapter weights instead of the entire model, significantly reducing computational cost. QLoRA further optimizes this by incorporating 4-bit quantization, enabling fine-tuning of very large models on limited hardware. AI

Researchers have found that sleep EEG signal criticality, measured using Multifractal Detrended Fluctuation Analysis (MFDFA), can predict future cognitive decline in dementia. Analysis of longitudinal data showed that cognitively healthy individuals exhibited sleep dynamics closer to an optimally critical state compared to those who later developed dementia. These findings suggest that MFDFA measures could be integrated into automated sleep-based screening tools for earlier intervention. AI

Researchers have developed a new AI approach to automate the tracking of glacier calving fronts, a critical but labor-intensive task for monitoring climate change. By adapting a deep learning model with minimal new data, including hand-labeled images and geological maps, the system significantly reduces the error in identifying glacier boundaries. This advancement promises to enable more comprehensive global glacier monitoring, with initial applications already underway in Norway's Svalbard archipelago. AI

Researchers have introduced FOGO, a novel optimizer designed to combat forgetting during AI model training. FOGO addresses both short-term forgetting at each training step and long-term forgetting common in continual learning by detecting and resolving gradient interference. The optimizer uses spectral orthogonalization and a compact codebook memory to preserve past update directions, demonstrating improved convergence and knowledge retention across various tasks, including fine-tuning LLaVA-7B and pretraining GPT-2, outperforming existing optimizers like Adam and Muon. AI

Researchers have developed Flash-GMM, a new fused Triton kernel designed for efficient Gaussian Mixture Model (GMM) computations on GPUs. This kernel significantly reduces memory requirements by avoiding the materialization of the full responsibility matrix, leading to a 20x speedup and enabling the processing of datasets 100x larger than previously possible on a single device. Flash-GMM has been integrated into approximate nearest-neighbor search, offering a viable alternative to k-means clustering and improving recall rates. AI

Researchers have identified that Chain-of-Thought (CoT) fine-tuning, while improving reasoning, significantly degrades long-context recall in hybrid linear-attention models. This issue, termed "attention amnesia," causes performance drops on tasks like Needle-In-A-Haystack. A new training-free method called QK-Restore has been proposed to fix this by restoring specific query-key projection weights from a pre-fine-tuning checkpoint, successfully recovering long-context capabilities without sacrificing reasoning performance. AI

Kwai has released Keye-VL-2.0-30B-A3B, an open-source multimodal foundation model designed for long-video understanding and agentic intelligence. This model utilizes DeepSeek Sparse Attention to process up to 256K context, capturing essential frames and temporal dependencies in hour-long videos. It also incorporates Cross-Modal Multi-Teacher On-Policy Distillation to enhance multi-task alignment and agent collaboration across various scenarios. Evaluations show state-of-the-art performance on video understanding and temporal localization benchmarks. AI

A new benchmark called UOJ-Bench has been developed to evaluate Large Language Models (LLMs) on code generation, hacking, and repair tasks, moving beyond simple problem-solving. Initial tests show that even top-tier models struggle with identifying errors in human-written code, with success rates below 50% in one-shot evaluations. While test-time scaling improves performance significantly, it incurs substantial computational costs, limiting practical deployment. However, the best models can still identify errors in a small percentage of full-score submissions, suggesting potential for LLMs to offer complementary insights to existing judging systems. AI

Researchers have developed an extension for the hls4ml toolkit to enable the deployment of Spiking Neural Networks (SNNs) on Field-Programmable Gate Arrays (FPGAs). This new capability allows for clock-driven inference of SNNs trained in PyTorch, offering low-latency temporal processing. The system demonstrated inference times of approximately 34 microseconds on a quantized SNN, paving the way for streamlined optimization and deployment of SNN models for real-time applications. AI

Researchers have developed a new technique to enhance reinforcement learning (RL) policies by leveraging existing suboptimal baseline policies. This method gradually transfers control from the baseline to a trainable learning policy, improving training efficiency and ultimately producing a standalone policy that outperforms the original baseline. The approach is formalized with theoretical analysis and demonstrated through empirical results on continuous-control benchmarks, showing high goal-reaching rates throughout the training process. AI

Researchers have introduced iMaC (Image as Action Control), a new paradigm for embodied world models in robotics. This approach uses raw visual images as action representations, moving away from traditional low-dimensional vectors. iMaC aims to improve generalization, dynamic modeling, and control for diverse robotic agents by treating visual manipulation as image-based action tokens. AI

Researchers have developed a new method for representing complex appearance effects in 3D scene reconstruction, moving beyond traditional Spherical Harmonics (SH). Their work introduces the Normalized Anisotropic Spherical Gabor function, which efficiently models high-frequency details like specular reflections and glints. This new formulation offers improved reconstruction quality while being significantly more memory-efficient and faster to evaluate than existing approaches. AI

Researchers have introduced iOSWorld, a new benchmark designed to evaluate the personalization capabilities of AI agents on mobile devices. This benchmark features a simulated iOS environment with 26 interconnected apps that store user-specific data like messages and financial records. It includes 133 tasks, ranging from single-app operations to complex multi-app scenarios requiring memory and personalization inference. Initial evaluations show that even advanced models struggle with these tasks, with the best configuration achieving only 52% overall accuracy. AI

Researchers have developed a new imitation learning method called Difference-Aware Retrieval Policies (DARP). This approach improves generalization by using training data during inference, predicting actions based on k-nearest neighbors and their relative distances to query states. DARP achieves significant performance gains over standard behavior cloning in various domains, including robotics and continuous control. AI

Researchers have introduced Perturbative Contrastive Physical Learning (PCPL), a new framework where learning arises from contrasting how physical systems respond to slight variations. This approach unifies and extends existing methods like Equilibrium Propagation and Frequency Propagation. PCPL allows learning without centralized gradient computation, as the learning geometry emerges implicitly from the system's physical response. AI

Researchers have developed a new framework called Spatio-Temporal Bound Propagation (STBP) to improve the verification of neural networks used in safety-critical applications like autonomous driving and medical imaging. This method models adversarial perturbations with more realistic spatio-temporal constraints, leading to tighter approximations and better robustness guarantees than existing techniques. The framework also introduces ST-Bench, a new benchmark designed to systematically evaluate verifiable robustness in these domains. AI

Researchers have developed a new framework for understanding the training dynamics of feed-forward ReLU neural networks. Their work rewrites gradient descent not as a weight-space dynamic, but as a collective dynamic on the training-set space. For deeper networks, this reveals a hierarchical structure of weight-induced operators that manage information flow between layers. AI

Researchers have precisely defined the VC dimension for depth-L Transformers with W parameters, establishing an upper bound of O(LW log(TW)) and a nearly matching lower bound. The study also characterizes the sample complexity for chain-of-thought learning with these Transformers, showing teacher forcing achieves O(LW log((T+T')W)) complexity. Any learning rule utilizing chain-of-thought data requires at least \Omega(LW log((T+T')W/L)) examples. AI

Researchers have developed a novel unsupervised learning algorithm for neural disentanglement using holographic reduced representations (HRR). This approach treats disentangled representations as symbolic structures, moving away from continuous representations common in prior work. The HRR unbinding operation demonstrates an inductive bias for separating factors, achieving competitive results on disentanglement metrics and showing robustness to noise. AI

Researchers have introduced PRIME, a new capability that assesses task correctness and predicts proxy acceptance in AI models. This capability emerges before visible reward hacking occurs and can forecast the onset and severity of such issues. PRIME adapts to changing evaluators and can serve as an early warning signal for alignment risks in AI systems. AI

Researchers have developed a new diagnostic theory and benchmark to understand how well local score models can extrapolate across different system sizes. They found that architectural locality alone is insufficient for stable size extrapolation, which is instead governed by the quasi-locality of the Gaussian-smoothed score. The study introduces the Finite-Depth Local Flow (FDLF) benchmark to empirically validate these findings, demonstrating that stable extrapolation depends on the interplay between spatial mixing, score quasi-locality, and model receptive fields. AI

Researchers have developed AdvGRPO, a novel co-training framework designed to enhance the adaptive red teaming of language models. This method addresses the instability of GRPO in attacker-defender optimization by employing dense multi-channel rewards and decoupled advantage normalization. The training process follows a curriculum, starting with single-turn attacks and progressing to multi-turn scenarios before initiating co-training, ultimately producing more effective attacks and robust defenders. AI

Researchers have developed a new method called Human-Perceptible Adversarial Attacks (HPAA) that exploits the difference between human and large language model (LLM) perception of harmful content. By using typographic manipulations like spacing and visual emphasis, these attacks can make harmful text easily recognizable to humans while remaining undetected by LLM-based moderation systems. In tests, HPAA achieved over 86% human recognition with less than 1% detection by moderation systems, revealing a significant vulnerability in current content moderation. AI

Researchers have developed Cranio-Diff, a novel diffusion-based framework for reconstructing faces from 2D X-ray skull images. This method addresses limitations in existing generative models by integrating skull-conditioned structural guidance and biometric text conditioning to ensure semantic and structural alignment between the skull and the generated face. The framework was evaluated on a unique dataset of 120 subjects, generating synthesized faces across different age groups and BMI variations, and demonstrated superior performance in image quality and retrieval tasks compared to existing approaches, suggesting its utility in forensic investigations. AI

A new paper introduces a comprehensive catalog of 84 numeric formats used in machine learning hardware, addressing the challenge of silent divergences when porting models across different accelerators. The catalog includes bit-exact conformance packs for various formats like FP8, BF16, and MXFP4, serving as a vendor-neutral reference. This work aims to provide a shared standard for engineers to diagnose and resolve discrepancies, ensuring greater consistency in model performance across diverse hardware. AI

Researchers have developed GenEyePose, a novel pipeline for generating synthetic eye movement data to train AI models for neurophysiologic biomarker development. This approach addresses the scarcity of real-world clinical data and privacy concerns associated with eye-tracking studies. A deep learning classifier trained on this synthetic data demonstrated promising performance in distinguishing normal from abnormal saccadic eye movements, showing potential for clinical applications in screening and localization of brain abnormalities. AI

Researchers have developed an enhanced system for the SoccerNet 2026 Player-Centric Ball-Action Spotting Challenge. Their approach builds upon existing FOOTPASS baselines by incorporating gradient checkpointing for efficient fine-tuning, fusing graph neural network (GNN) outputs with visual features, and applying square-root frequency class weighting to balance imbalanced training data. The system achieved a Macro F1 score of 0.548 on the test set and 0.446 on the challenge set. AI

Researchers have introduced MeCo, a novel one-step generative corrector for multi-channel speech separation. This method uses a MeanFlow-based approach to map estimated audio directly to clean speech, aiming to improve human listening quality beyond traditional discriminative models. MeCo incorporates Data-Space Optimization with an $\mathbf{x}_r$-loss and an Endpoint SI-SDR loss to enhance both signal fidelity and subjective listening experience. AI

A new research paper investigates how "thinking" mechanisms in large language models affect instruction following. The study found that while overall performance changes are minor, the "thinking" process alters error patterns, improving some instructions while worsening others. Specifically, "Planning" constraints benefit from thinking, whereas "Precision" constraints consistently degrade. Analysis of model traces revealed differing correlations between trace relevance and final answer compliance across these constraint types. AI

Researchers have developed FMplex, a novel system designed to optimize the serving of foundation models (FMs) by treating them as a virtualization substrate. This approach allows multiple downstream tasks to share a single physical FM instance, reducing memory waste and amortizing costs associated with batching and loading. FMplex enables task-specific extensions and isolation while improving efficiency, demonstrated by significant reductions in latency and increased task hosting capacity. AI

Researchers have developed a new method called auto-targeting to optimize e-commerce marketing campaigns by jointly selecting users and products. The approach uses constrained spectral biclustering, greedy local search, and a multi-armed bandit framework to create disjoint campaigns with strong user-item affinities. Evaluations on synthetic and real-world datasets indicate that biclustering yields high-quality, fair campaigns, while bandit methods offer scalability for larger datasets. AI

Researchers have developed new methods to improve supervised fine-tuning (SFT) for large language models. One approach, FisherAdapTune, uses the Fisher information geometry to dynamically select parameter groups for adaptation, enhancing in-distribution performance and zero-shot transfer. Another set of methods, including Target-SFT and PriFT, reinterprets SFT as target distribution design. These techniques aim to create more stable and effective training objectives by better aligning the fine-tuning process with the model's pretrained knowledge, leading to state-of-the-art results on various reasoning and code generation tasks. AI