Brief

Researchers have developed GridProbe, a novel method to improve the efficiency of long-video Visual Language Models (VLMs). This technique adaptively selects relevant frames during inference, reducing the computational cost associated with processing thousands of frames. GridProbe achieves this by probing frame importance in the answer space, allowing for a dynamic adjustment of the number of frames processed based on question difficulty without sacrificing accuracy. AI

Researchers have introduced RadThinking, a new dataset designed to train AI systems in longitudinal clinical reasoning for radiology. The dataset includes visual question-answering pairs across three difficulty levels, focusing on atomic perception, single-step reasoning, and multi-step compositional reasoning. RadThinking aims to enable AI to not just detect cancer but also reason about it, using over 20,000 CT scans and incorporating clinical reporting standards. AI

Researchers have developed TINS, a novel method for Out-of-Distribution (OOD) detection in vision-language models. TINS addresses limitations of static negative labels by learning dynamic negative semantics during test-time inference. It employs image-to-text modality inversion and an ID-prototype-separated regularization to prevent contamination from in-distribution concepts. Experiments show significant improvements, such as reducing FPR95 from 14.04% to 6.72% on the Four-OOD benchmark. AI

A new paper proposes "Agent Cybernetics" as a theoretical framework for understanding and developing advanced AI agents. The authors argue that while foundation agents are increasingly used for complex, long-horizon tasks, their design is largely empirical. By mapping principles from classical cybernetics to agent design, the paper introduces a framework aimed at ensuring reliability, continuous operation, and safe self-improvement for these agents. The proposed approach offers concrete engineering recommendations for domains like code generation and automated research. AI

Researchers have developed a new framework called FedMITR to improve one-shot federated learning, particularly in scenarios with highly non-independent and identically distributed (non-IID) data. This method addresses the issue of low-quality synthetic data generated by existing approaches by employing sparse model inversion to focus on meaningful image patches and avoid background noise. Additionally, FedMITR uses a token relabeling strategy for Vision Transformers (ViTs) to enhance prediction robustness by distinguishing between high and low information density patches. AI

Researchers have developed a new multimodal large language model (MLLM) framework for analyzing remote sensing data, specifically focusing on construction sites. This framework utilizes the Sentinel-2 satellite imagery dataset and transforms existing annotations into natural language question-answer pairs for spatiotemporal analysis. The system, trained on a dataset of over 21,000 image chips and millions of temporal comparison examples, aims to enable reasoning about ongoing construction processes and their evolution over time. AI

Researchers have developed iPay, a new framework for recognizing payment actions in transit surveillance footage. This system utilizes a multimodal mixture-of-experts architecture, combining RGB and skeleton data streams with a dual-attention fusion mechanism. An additional Spatial Difference Discriminator explicitly models hand-to-anchor motion to enhance discriminability. iPay achieved 83.45% recognition accuracy on a dataset of over 500 payment clips collected from real onboard transit surveillance, demonstrating its suitability for edge deployment. AI

Alibaba's Qwen-Image-2.0 is a new foundation model designed for both high-fidelity image generation and precise editing within a single framework. It addresses limitations in existing models concerning ultra-long text rendering, multilingual typography, photorealism, and instruction following. The model utilizes Qwen3-VL as a condition encoder and a Multimodal Diffusion Transformer, trained on extensive data, to achieve improved multimodal understanding and flexible generation capabilities. AI

Researchers have introduced AllocMV, a novel framework designed to optimize the generation of music videos by treating it as a Multiple-Choice Knapsack Problem. This approach uses a structured persistent state, including character entities and scene priors, to maintain consistency across different shots. AllocMV estimates segment saliency and employs dynamic programming to optimally allocate computational resources, balancing quality with expenditure under budget and rhythmic constraints. AI

Researchers have developed a new strategy to enhance Graph Neural Networks (GNNs) for drug discovery tasks like Quantitative Structure-Activity Relationship (QSAR) studies. This method involves pre-training GNNs to predict Extended-Connectivity Fingerprints (ECFPs), a classical molecular featurization approach. The pre-trained GNNs demonstrated statistically significant improvements in performance across several benchmarks, particularly for out-of-distribution splits. However, the effectiveness varied with dataset heterogeneity and endpoint complexity, with some instances showing underperformance in out-of-distribution settings. AI

Researchers have developed AURORA, a new framework designed to diagnose and mitigate "grey failures" in computing systems. This uncertainty-aware resilience micro-agent uses parallel agents that integrate causal inference and state-graphs to perform root-cause analysis. AURORA's dual-gated mechanism ensures interventions only occur when causal confidence is high and uncertainty is low, otherwise escalating the issue. Experiments show AURORA achieves 62.0% repair accuracy with a 0% destructive action rate and a 3ms mean time to repair. AI

Researchers have developed new post-training objectives for large language models that optimize for the best-of-N performance, rather than just the average reward. This is crucial because current deployment strategies involve sampling multiple responses and selecting the best one, a process that standard training objectives do not adequately address. The proposed Tail-Extrapolated (TEA) estimators and Prefix-TEA can approximate the best-of-N objective using significantly fewer per-prompt rollouts during training than would be required for deployment, showing improved performance on instruction-following tasks. AI

Researchers have established conditions for successful sparse recovery using data from sources of varying quality. Their work introduces the concept of the 'Price of Quality,' which quantifies the trade-off between high-quality and low-quality samples needed for recovery. The study reveals that algorithmic recovery methods like LASSO demonstrate robustness to data heterogeneity, matching homogeneous-noise thresholds. AI

Researchers have introduced NASH, a new framework for data selection in machine learning that aims to improve the effectiveness of methods like Data Shapley. NASH decomposes utility functions into simpler, Shapley-informative components and aggregates them non-linearly to select high-quality data subsets. The framework is designed to boost performance with only a minimal increase in runtime cost. AI

Researchers have introduced a novel approach to machine unlearning that focuses on the underlying data distributions rather than just model parameter updates. This method aims to infer these distributions precisely to distill an exact unlearning signal. Theoretical analysis and experimental validation on three forgetting scenarios demonstrate that this framework achieves a classifier closer to an ideal retrained model than existing methods. AI

Researchers have developed a new method called Gated Cropped Attention-Delta steering (GCAD) to improve the reliability of controlling language model behavior. Standard activation steering can degrade performance in long conversations due to issues with the KV-cache. GCAD addresses this by extracting steering signals from self-attention mechanisms and applying them with token-level gating, significantly enhancing long-horizon coherence and trait expression in multi-turn dialogues. AI

Researchers have developed bViT, a novel Vision Transformer architecture that utilizes a single transformer block applied repeatedly for image recognition. This recurrent approach achieves accuracy comparable to standard ViTs on ImageNet-1K with significantly fewer parameters. The study suggests that a substantial portion of a ViT's depth can be achieved through recurrent computation, especially when the representation space is wide, enabling parameter-efficient fine-tuning for downstream tasks. AI

Researchers have developed BCJR-QAT, a novel method for quantizing large language models to 2 bits per weight, a significant advancement beyond current post-training quantization techniques. This new approach uses a differentiable relaxation of the Viterbi algorithm, enabling quantization-aware training and achieving better perplexity scores on benchmarks like WikiText-2. The method has been demonstrated to improve performance on models such as Llama-3.2-1B, outperforming existing methods by a notable margin. AI

Researchers have developed a new Gaussian Process-based acquisition function called AB-SID-iVAR for active learning problems. This method addresses the challenge of learning an unknown function under a self-induced Boltzmann distribution, which is common in computational chemistry but difficult due to the unknown and intractable nature of the target distribution. The proposed approach approximates the Bayesian target distribution without needing to estimate the partition function, making it applicable to both discrete and continuous domains. Experimental results show improvements over existing methods on synthetic benchmarks and real-world tasks in PES modeling and drug discovery. AI

Researchers have developed a new criterion for initializing weights in gated recurrent neural networks, crucial for the performance of reservoir computing models. This criterion, derived from random-matrix theory, helps identify an effective critical point that separates ordered and chaotic phases in randomly initialized models. The method closely tracks the optimal gain for gated RNNs on forecasting tasks and could inform future initialization strategies. AI

Researchers have developed diffGHOST, a new conditional diffusion model designed to generate synthetic mobility trajectories while preserving user privacy. Unlike previous methods that made assumptions about implicit privacy, diffGHOST aims to provide explicit privacy guarantees. The model achieves this by identifying and mitigating the memorization of sensitive data through the use of conditional segments within its learned latent space. AI

Researchers have identified that the internal representation of personality in Large Language Models (LLMs) can act as a defense against emergent misalignment. By mapping LLM personalities using psychometric profiles, they found that specific vectors related to social valence, like 'evil' or a newly introduced 'Semantic Valence Vector', function as intrinsic guardrails. Ablating these vectors significantly increased misalignment rates, while amplifying them suppressed harmful behaviors. This suggests that even after fine-tuning on benign data, the core personality representations remain stable and can be leveraged to regulate emergent misalignment across different model distributions. AI

Researchers have developed a new evaluation framework for coreference resolution that goes beyond aggregate statistical metrics. This semantically-enhanced approach uses Concept and Named Entity Recognition to assign semantic labels to mentions and clusters, allowing for evaluation stratified by semantic class like people, locations, or events. Experiments on datasets such as OntoNotes show this method uncovers systematic weaknesses not visible with traditional metrics and can inform targeted data augmentation for improved out-of-domain performance. AI

Researchers have proposed a new framework for evaluating fairness in automatic speech recognition (ASR) systems. The proposed methodology emphasizes the importance of clearly defining the fairness hypothesis and tailoring metrics accordingly. It also highlights the need for fine-grained analysis of demographic intersections within datasets to avoid misidentifying mistreated speaker groups. AI

Researchers have proposed a new set of features to better categorize linguistic expressions involving light verbs and aspectual variants. These expressions, such as 'take on debt' versus 'have debt,' are often difficult to classify as verbal idioms, light verb constructions, or compositional phrases. The proposed features aim to establish clearer boundaries between these categories, leading to more accurate linguistic analysis. AI

Researchers have identified a pretraining failure mode in language models where upper layers prematurely specialize their attention patterns before lower layers have stabilized. This "premature upper-layer attention specialization" can be mitigated by temporarily slowing the Q/K projections in these upper layers during early training. This intervention improves final perplexity and downstream accuracy without changing other model parameters, suggesting a critical interaction between decoder architecture and optimization. AI

A recent analysis delves into the evolution of Convolutional Neural Network (CNN) architectures, specifically examining ResNet, EfficientNet, and ConvNeXt. The author investigates whether advancements in state-of-the-art CNNs are primarily due to architectural innovations or improvements in scaling and training strategies. The findings suggest that both factors play a significant role and are difficult to disentangle, with ResNet enabling greater depth, EfficientNet introducing principled scaling, and ConvNeXt adopting transformer-like training recipes. AI

This paper introduces a PAC-Bayes framework designed to learn controllers for unknown stochastic linear discrete-time systems. The research provides a data-dependent bound on controller performance and proposes new learning algorithms with theoretical guarantees. These algorithms are applicable to both finite and infinite controller spaces and offer performance comparable to LQG controllers in specific scenarios. AI

Researchers have developed a new computer vision framework to automatically quantify fish communities and their biomass from underwater video. This method uses deep learning for fish identification, tracking, and 3D reconstruction to provide species-level abundance and biomass estimates. Applied over 20 days with hourly observations, the system revealed dynamic fluctuations in fish populations, offering a scalable solution for continuous, non-invasive ecological monitoring. AI

Researchers have developed a new theoretical framework to understand how neural networks learn features, particularly in large-width networks. Their work reveals that feature learning occurs through a series of sharp, discontinuous transitions as more data becomes available. This understanding leads to precise "neural scaling laws" that dictate the Bayes-optimal generalization error based on the effective number of learnable features and the data budget. AI

Researchers have investigated the impact of temporal sampling frequency on end-to-end autonomous driving trajectory prediction models. They found that while dense frame sampling is often assumed to improve performance, this is not always the case. Smaller models often perform best with lower or intermediate sampling frequencies, suggesting that dense sampling can introduce redundant information and noise that burdens models with limited capacity. Larger, vision-language-model-style architectures, however, continued to improve performance even at the highest tested sampling frequencies. AI

Researchers have developed a new kernel learning approach using cokriging to solve nonlinear partial differential equations (PDEs). This method leverages empirical information from multifidelity simulations to fit a differentiable non-stationary kernel to low-fidelity data. The approach then derives a high-fidelity kernel and mean, which are integrated into a Gaussian process framework for solving PDEs, demonstrating effectiveness on the Burgers' equation. AI

A new paper published on arXiv details a taxonomy for understanding and quantifying uncertainty in machine learning models used within physics. The research clarifies the distinction between predictive and inference uncertainties, offering a unified framework for both frequentist and Bayesian approaches. It also introduces and demonstrates validation tools such as coverage, calibration, and bias tests, crucial for scientific discovery relying on probabilistic statements. AI

Researchers have introduced SleepWalk, a new benchmark designed to rigorously test instruction-guided vision-language navigation capabilities of AI models. This benchmark focuses on localized, interaction-centric embodied reasoning within 3D environments, evaluating a model's ability to predict a trajectory that aligns with natural language instructions while respecting scene geometry and avoiding collisions. SleepWalk categorizes tasks into three difficulty tiers to allow for detailed analysis of how models handle increasing spatial and temporal complexity, revealing significant failures in grounded spatial reasoning, particularly with multi-step instructions and occlusion. AI

Researchers have developed a novel approach to enhance mobile GUI agents by training world models across four modalities: delta text, full text, diffusion-based images, and renderable code. These models achieved state-of-the-art performance on relevant benchmarks, demonstrating the utility of different representations for predicting action consequences. The study found that while renderable code offers high fidelity for data construction, text-based feedback is more robust for online execution, and generated trajectories can improve agent performance despite distribution shifts. AI

Researchers have developed a new machine learning model called the Wildfire Ignition Set Predictor (WISP) to forecast active fires at a high resolution. Unlike previous methods that predict danger on a regional scale, WISP reformulates the problem as predicting a set of localized fire cluster centers. The model utilizes 48 hours of meteorological and satellite data to predict fire locations on a 375m grid, achieving significant accuracy in localization and coverage on a global test set. AI

Researchers have introduced LeapTS, a new framework that reframes time series forecasting as an adaptive scheduling problem. This approach moves away from fixed mappings to a dynamic process where a hierarchical controller selects optimal prediction scales and advancement lengths at each step. The system utilizes neural controlled differential equations to manage temporal dynamics and scheduling feedback, leading to improved forecasting accuracy and significantly faster inference speeds compared to existing Transformer-based models. AI

Researchers have introduced BROS, a novel method for memory-efficient single-loop bilevel optimization. This approach addresses the significant memory demands of existing methods when dealing with large neural networks in deep learning tasks. BROS utilizes randomized subspaces and a bias-correction technique to achieve convergence rates comparable to exact methods while reducing peak memory usage by up to 44.9%. The method has demonstrated effectiveness in various applications, including hyperparameter learning and sample reweighting for Vision Transformers. AI

Researchers have developed a new Gaussian process framework that uses neural feature maps to create more expressive kernels. This method allows for efficient and accurate Gaussian process inference, applicable to both regression and classification tasks across various data types like images and tabular data. The approach demonstrates superior accuracy and efficiency compared to existing methods on benchmark datasets. AI

Researchers have developed a novel cold diffusion framework to address the challenge of dereverberating percussive audio signals, such as drums, which have been largely overlooked in favor of speech processing. This new method models reverberation as a progressive degradation and employs two reverse-process parameterizations with UNet and diffusion Transformer backbones. Experiments show the framework significantly outperforms existing score-based and conditional diffusion baselines on both in-domain and out-of-domain datasets, using specialized metrics for percussive audio. AI

Researchers have introduced MARGIN, a new framework designed to improve the detection of software vulnerabilities, particularly in datasets with imbalanced frequencies and difficulties. MARGIN addresses these challenges by analyzing the geometric distortions in hyperspherical representation space. The framework employs adaptive margin metric learning and hyperspherical prototype modeling to create more discriminative vulnerability representations and stable decision boundaries. Experiments show MARGIN outperforms existing methods, enhancing classification, detection, robustness, interpretability, and generalization. AI

A technical blog post details a method for detecting drift in Retrieval-Augmented Generation (RAG) systems when switching between large language models. The author proposes using the `ragvitals` library to monitor five independent drift dimensions: QueryDistribution, EmbeddingDrift, RetrievalRelevance, ResponseQuality, and JudgeDrift. By carefully separating live traffic from reference probes, the system can accurately identify that only ResponseQuality changed when the generator was swapped from Claude Sonnet to Gemma 4 9B, avoiding false alarms on other dimensions. AI

Researchers have developed FERA, a novel framework for improving large language model reasoning in a federated setting. This approach allows a central server to enhance reasoning by collaborating with multiple clients that hold private demonstration data, without needing to share raw data. FERA uses iterative co-refinement where clients provide reasoning traces with uncertainty estimates, which the server synthesizes to improve future reasoning rounds. The system incorporates Uncertainty-Aware Self-Critique Aggregation (UA-SCA) to revise flawed reasoning steps and improve trust-based weighting, leading to consistent performance gains over existing federated methods. AI

Researchers have developed PHAGE, a novel graph encoder designed to better represent patent documents. Unlike previous methods that linearize claims and lose hierarchical information, PHAGE explicitly encodes the dependency structure between patent claims. It distinguishes between different types of claim relationships and integrates this topological information into a token-level attention mechanism. This approach significantly improves performance on patent classification, retrieval, and clustering tasks. AI

Researchers have developed PixelFlowCast, a novel two-stage framework for precipitation nowcasting that enhances both prediction accuracy and inference speed. This method avoids latent space compression, which is common in diffusion-based models and often degrades fine-grained details. PixelFlowCast first generates coarse forecasts and then uses a KANCondNet to extract spatiotemporal features for conditional guidance, enabling a latent-free predictor to generate high-quality, fast predictions. Experiments on the SEVIR dataset show PixelFlowCast outperforms existing methods, particularly for longer forecast sequences. AI

Researchers have introduced Lakestream, a new data plane designed for large foundation model training that operates directly on object stores without a broker. It offers transactional global batches with ACID semantics extended for training consistency, including atomic visibility and exactly-once recovery. Evaluations show Lakestream surpasses colocated dataloader throughput and Apache Kafka in ingestion speed and consumer latency. AI

Researchers have developed new theoretical frameworks for training and calibrating language models in distributed settings with limited bandwidth. The Federated Probe-Logit Distillation (FPLD) protocol offers a statistical consistency rate that depends on factors like node count, sample size, and quantization budget, with bandwidth entering through a vanishing quantization term. Additionally, the Federated Conformal RAG (FC-RAG) protocol provides a distribution-free marginal-coverage bound where retrieval bandwidth is a key parameter, showing improvement with more nodes. AI

Researchers have developed a new hypothesis suggesting that human learning of reusable programming abstractions is prospective, meaning it anticipates future task demands rather than solely relying on past data. This prospective compression approach was tested using the Pattern Builder Task, where participants created geometric patterns. The study found that human abstraction behavior adapts to evolving task-generating processes, outperforming existing retrospective compression algorithms and current large language model-based program synthesis methods. AI

Researchers have developed Geometric 4D Stitching, a new framework designed to improve the geometric consistency of 4D scene generation. This method efficiently identifies and fills in missing geometric areas, constructing complete 4D scene representations in under 10 minutes on a single GPU. The framework also enables interactive expansion and editing of 4D meshes, offering a more robust approach to 4D content creation. AI

Researchers have developed Yeti, a novel protein structure tokenizer designed for multimodal AI models. Unlike previous methods that prioritize reconstruction, Yeti uses a lookup-free quantization approach trained with a flow matching objective, enabling both accurate reconstruction and effective generation of protein sequences and structures. This compact tokenizer, with significantly fewer parameters than existing models, facilitates the training of efficient multimodal models capable of co-generating plausible protein designs. AI