Brief

Researchers have developed TractFM, a novel foundation model designed to learn representations directly from diffusion MRI tractograms. This model uniquely combines a local streamline encoder with a permutation-equivariant tractogram encoder, enabling it to process all streamlines from a subject simultaneously. By pretraining on anatomical parcellation, TractFM generates reusable embeddings for both individual streamlines and compact subject-level descriptors. The model demonstrates strong generalization capabilities, achieving accurate tract parcellation and predicting subject phenotypes like age and sex across different tractography algorithms and datasets. AI

Researchers have developed a new method to improve the accuracy of attribution patching, a technique used to understand how different parts of a language model contribute to its behavior. The current method, a first-order approximation, can be unreliable due to network non-linearities. The new approach introduces a second-order correction using Hessian-vector products, which significantly enhances the fidelity of circuit recovery. This method is computationally feasible for larger models and offers practical tools for detecting untrustworthy estimates and quantifying errors. AI

Researchers have developed a new method called Two-Stage Latent Feature Optimization (TS-LFO) to bypass copyright protection in diffusion-based image customization. This technique addresses existing defenses by restoring the mapping between input images and their latent representations. TS-LFO uses a two-stage optimization process to suppress noise and refine latent features, demonstrating effectiveness against state-of-the-art copyright protection methods. AI

Researchers have introduced a new theoretical framework for compressing bioelectrical signals, moving beyond traditional waveform preservation methods. This "Bioelectrical Information Theory" considers physiological structure, model capacity, and task requirements to determine compression limits. The approach involves reducing noise, creating structured representations, and discarding task-irrelevant information, ultimately reframing compression as a model- and task-conditioned quantity. AI

Researchers have introduced Sigma-Branch (SigmaB), a novel framework designed to optimize deep neural networks for memory-constrained edge devices. SigmaB restructures dense networks into a hierarchical tree with shared backbones, routers, and specialized leaves, enabling dynamic inference. This approach significantly reduces the number of active parameters per inference by executing only a single root-to-leaf path, thereby minimizing off-chip weight transfers without sacrificing overall model capacity. AI

Researchers have introduced the "strategic confinement problem," which addresses how to prevent programs processing confidential data from leaking it when interacting with strategic agents. This problem arises because these agents can concentrate residual communication capacity on specific, high-impact data predicates, allowing for significant harm even with negligible information leakage. The paper argues that AI systems, due to their unpredictable learned conventions and potential for covert communication, naturally instantiate this challenge, shifting the focus from information flow to the strategic outcomes achievable by agents. AI

Researchers have found that metrics used to curate training data for AI policies do not necessarily improve the performance of those policies. In experiments on a pick-and-place benchmark, a metric that was highly effective at detecting defects actually resulted in the worst-performing policy. Conversely, a metric with lower defect detection accuracy produced a policy that was nearly as good as one trained on perfect data. The study also revealed that many metrics incorrectly use episode length as a proxy for defects, inflating their apparent accuracy. AI

Researchers have developed a new framework called GeoABC to improve the accuracy of neural operators in aerodynamic simulations. This method explicitly models the anisotropic nature of flow near boundaries, where behavior differs along the wall versus perpendicular to it. By incorporating boundary geometry as a directional prior, GeoABC enhances predictions and significantly reduces errors, making neural operators more suitable for high-fidelity aerodynamic simulations. AI

Researchers have developed a novel divide-and-conquer modeling strategy specifically for the CTF-4-Science Lorenz benchmark. This approach tailors different model classes to distinct prediction tasks within the benchmark, rather than using a single model for all scenarios. The system achieved a final public score of 79.63 by employing techniques like smoothing-based reconstruction for denoising, NG-RC/NVAR models for long-time forecasting, and a fitted Lorenz transition correction for short-time prediction, demonstrating the effectiveness of scenario-specific updates. AI

Researchers have developed a new method for detecting AI-generated text by learning style representations without needing authorship labels. This approach uses a style encoder to reconstruct human text from its machine-generated paraphrase, effectively capturing non-semantic stylistic features. The learned representations perform competitively in both few-shot and zero-shot detection scenarios, even generalizing to unseen language models and tasks like authorship verification. AI

A new research paper explores the emotional characteristics of translations produced by Large Language Models (LLMs). The study compares LLM translations of Margaret Atwood's "Oryx and Crake" with human translations and post-edited versions. Findings indicate that LLMs imprint distinct emotional patterns on their translations, which can obscure the original author's voice and are only partially corrected by human post-editing. AI

Researchers have developed a Pareto-guided teacher alignment framework to address fairness issues in personalized text generation. This framework aims to reduce demographic disparities while maintaining personalization fidelity by combining several techniques, including candidate generation, feasibility gating, and Pareto-style selection. Evaluations on persuasion tasks revealed that different alignment strategies occupy distinct regions of a fairness-personalization Pareto frontier, highlighting the objective-dependent nature of fairness mitigation and the need for multi-audit model selection. AI

Researchers have developed MMClima, a new framework designed to advance AI capabilities in climate science. This framework includes a large dataset of over 104,000 expert-validated question-answer pairs that integrate text, video transcripts, and scientific figures across five key climate domains. MMClima also provides an evaluation pipeline and a domain-adapted baseline model, mmclima-70b-txt, to facilitate standardized multimodal AI assessment in climate research. AI

Researchers have developed a novel probabilistic foundation model capable of predicting detailed spectral data for galaxies using only broadband images. This model, trained on extensive data from the Dark Energy Spectroscopic Instrument (DESI), bypasses the need for traditional Integral Field Unit (IFU) spectroscopy, which is observationally expensive. The system achieves IFU-like spectral resolution and spatial mapping capabilities by leveraging a masked autoencoder framework and incorporating fiber and redshift-aware encodings, demonstrating performance comparable to supervised methods trained directly on IFU data from the MaNGA survey. AI

Researchers have developed an improved Generative Adversarial Network (GAN) specifically for restoring partially missing micro-resistivity imaging logs. The proposed method incorporates a Feature Pyramid Network (FPN) as the generative backbone, enhanced with depth-separable convolutional residual blocks and Inception modules to better capture pixel and semantic information across multiple scales. Experimental results show an average structural similarity measure of 0.903 on test data, outperforming other methods by approximately 0.3 and improving semantic structure coherence and texture details for subsequent interpretation. AI

A new perspective paper explores the integration of foundation models into robots for patient and elderly care. While these models offer potential for personalized assistance and flexible communication, current systems face challenges with reliability, such as hallucinations and conversational breakdowns. The paper highlights that while usability and engagement benefits are reported, evidence for significant clinical or care-related outcomes remains limited, suggesting a need for care-specific evaluation standards and better integration into existing workflows. AI

Researchers have developed a novel framework for detecting advanced persistent threats (APTs) across different operating systems without requiring any labeled data from the target system. The approach uses natural language processing to describe process behavior, embeds these descriptions using pre-trained language models, and then applies optimal transport methods to quantify deviations from normal behavior learned from a source operating system. Evaluations on multiple APT scenarios and operating systems demonstrated improved detection accuracy compared to existing source-only methods. AI

Researchers have developed a new dual-branch gated fusion framework to improve the tracing of audio deepfake sources. This system combines XLSR-53 with a novel 66-dimensional descriptor called CORES, which captures a wider range of synthesis artifacts than previous methods. An input-conditioned gate adaptively weights these two branches to overcome representational imbalance and enhance performance on out-of-domain datasets. AI

Researchers have developed a systematic study to improve the correlation between simulation and real-world robot evaluations for vision-language-action (VLA) policies. The study analyzes how well simulation platforms preserve real-world conclusions regarding policy ranking and performance. It also offers guidance on leveraging simulation for policy improvement, including when simulator-based fine-tuning is beneficial and how data volume impacts alignment. AI

Researchers have developed a novel attack method called Test-time Adversarial Takeover (TAKO) that allows real-time hijacking of robotic systems controlled by diffusion-based policies. This attack manipulates the visual conditioning input to the robot, enabling an attacker to steer the robot's actions and achieve custom objectives. TAKO utilizes universal patches learned through diffusion inference, proving effective across various robotic tasks, visual encoders, and generative inference models, with human operators achieving 100% takeover success in all tested scenarios. AI

Researchers have developed a new method called the Distributed Detectability Band to counter AI sabotage attacks that are designed to evade detection. These attacks distribute harmful actions across many seemingly benign steps, making them difficult for standard AI monitors to identify. The proposed technique encodes harm within the temporal correlation structure of actions, rather than relying on individual action scores, to effectively detect these sophisticated, sub-threshold sabotage attempts. AI

Researchers have developed Visual-TCAV, a new framework for explaining image classification models. This method combines local saliency maps with concept-based attribution, addressing limitations of existing techniques. Visual-TCAV can pinpoint where a specific concept is recognized within an image and quantify its contribution to a prediction, demonstrating improved faithfulness over prior methods. AI

Researchers have developed a method to assess personality dimensions in dyadic role-play scenarios, moving beyond static trait assumptions. Their findings indicate that perceived personalities significantly vary across different situations, such as neutral interviews versus stressful client interactions. Acoustic and non-verbal features, rather than speaker embeddings, proved more effective in predicting these perceived traits, with stress specifically correlating with neuroticism. AI

Researchers have developed a computational agent designed to simplify the design and maintenance of building-integrated photobioreactors (PBRs). This agent uses a Constraint Satisfaction Problem (CSP) solver to translate user sketches into fabrication-ready configurations, successfully handling complex layouts up to 15x15 grids. Additionally, a computer vision system has been created to monitor algae health from photographs, enabling autonomous maintenance and supporting the goal of scalable carbon capture systems. AI

Researchers have developed SPARX, a framework for accelerating Convolutional Neural Networks (CNNs) on edge devices. This system integrates approximate computing with security and privacy features within a RISC-V System-on-Chip. SPARX utilizes a custom RISC-V instruction extension and an approximate logarithmic CNN accelerator, enhanced by a differential-noise privacy engine and authentication mechanisms. Evaluations show significant reductions in area and power, alongside improved throughput, with a minimal impact on accuracy for specific CNN models. AI

Researchers have developed an automated pipeline to generate labels for low-light pedestrian detection using infrared and RGB cameras. This method involves detecting pedestrians in infrared images and then transferring those labels to corresponding RGB images. Models trained with these generated labels outperformed those trained on ground truth in several key metrics, suggesting a scalable approach for creating large low-light pedestrian datasets. AI

Researchers have developed GeoLoom, a new framework designed to generate high-quality geometric diagrams from textual descriptions. This system translates natural language into a formal language called GeoLingua, which is then used by a coordinate solver to produce precise diagrams. To support this, a new dataset called GeoNF has been created, and a novel evaluation metric has been proposed to assess structural accuracy. AI

Researchers have developed improved radiomic models for predicting lung cancer in indeterminate pulmonary nodules. By augmenting the training data with nodules from later development stages and employing biology-aware harmonization techniques, the models showed significantly better performance. This approach addresses challenges posed by low malignancy rates in early nodules and variability in image acquisition protocols, leading to higher accuracy in cancer prediction. AI

Researchers have developed a new method for reconstructing hyperspectral images (HSI) from standard RGB images, a process that can significantly reduce costs while maintaining high spatial resolution. The proposed Correlation and Continuity Network (CCNet) leverages both local spectral correlations and global spectral continuity to improve reconstruction quality. This approach has demonstrated state-of-the-art performance on benchmark datasets, outperforming existing spectral reconstruction algorithms. AI

Researchers have introduced Cyst-X, a new benchmark dataset and federated learning framework designed to improve the early detection of pancreatic cystic neoplasms. This initiative aims to address the challenges in stratifying malignancy risk, which currently leads to either unnecessary surgeries or missed diagnoses. The Cyst-X dataset includes 1,461 MRI scans from seven international centers, and the developed deep learning model achieved an AUC of 0.85 in distinguishing high-risk lesions. Notably, the federated learning approach allowed for distributed training across institutions without sharing raw patient data, maintaining performance while preserving privacy. AI

Researchers have developed Deepcontour, a new framework that uses deep learning to optimize the construction of integration contours for solving large-scale Generalized Eigenvalue Problems (GEPs). This method employs a deep learning-based spectral predictor and Kernel Density Estimation to automatically design efficient contours, leading to significant speedups. The framework achieved up to a 5.63x performance increase on various scientific datasets while maintaining numerical accuracy. AI

Researchers have introduced the selective disk bispectrum (SDB), a new image descriptor designed to be rotation-invariant and preserve all information except orientation. This descriptor aims to bridge the gap between hand-crafted and learned image representations by offering a balance of descriptive power, efficiency, and interpretability. The SDB has theoretical guarantees for its accuracy and invariance, and has been empirically validated on classification tasks and image alignment. AI

Researchers have developed the GCA framework, which includes a new dataset called GCA-DS and an agent named Gulf Climate Agent (GCA). This framework is designed to support climate decision-making specifically for the GCC states by integrating regional climate knowledge with geospatial and forecasting tools. The GCA agent utilizes a modular pipeline that processes real-time data and geospatial information to generate insights and visualizations, demonstrating improved reliability over general-purpose LLMs through domain fine-tuning and tool integration. AI

Researchers have introduced HydraCIL, a novel approach to class-incremental learning designed for resource-constrained environments like embedded systems. This method decouples feature extraction from classifier training, allowing for lightweight, task-specific classifier heads to be created without extensive backbone retraining. Experiments demonstrate that HydraCIL achieves performance comparable to state-of-the-art methods while significantly reducing training time and energy consumption. AI

Researchers have developed an AI application designed to provide users with real-time feedback on the perceived level of peace in social media videos. The system utilizes large language models to analyze YouTube video transcripts, measuring five social dimensions identified in peace studies. This approach proved more effective than traditional sentiment analysis in correlating with human coders' assessments of peace-related content. AI

Researchers have introduced HarDBench, a new benchmark designed to evaluate the safety of large language models (LLMs) when used in collaborative writing scenarios. The benchmark focuses on "draft-based co-authoring jailbreak attacks," where malicious users could prompt LLMs to generate harmful content within incomplete drafts. HarDBench covers high-risk domains like explosives, drugs, and weapons, and includes realistic prompts to test model susceptibility. The researchers also developed a safety-utility balanced alignment approach to mitigate these risks without compromising the LLM's helpfulness on benign tasks. AI

A new paper from arXiv highlights significant biases in current AI content watermarking techniques. The research indicates that the effectiveness and detectability of watermarks vary considerably based on the statistical properties of the content itself, leading to disparities across languages, cultural visual traditions, and demographic groups. The authors propose a framework for more inclusive benchmarking, emphasizing cross-lingual detection parity, culturally diverse content coverage, and demographic disaggregation of metrics, arguing that these fairness evaluations should precede widespread deployment. AI

Researchers have developed a novel spatiotemporal graph transformer designed to model complex interactions in metal additive manufacturing. This framework represents the manufacturing process as a network, allowing for the integration of multimodal data and capturing both within-node feature dependencies and cross-node neighborhood interactions. Experiments demonstrate that this approach significantly outperforms existing models in characterizing process-quality relationships, with cross-layer interactions proving critical for accurate quality prediction. AI

Researchers have developed a new method for interpreting machine learning models with categorical inputs. This approach, based on functional ANOVA decomposition, provides a closed-form solution that is computationally efficient and works even with dependent features. The new framework also offers a natural generalization of SHAP values for categorical data, addressing a long-standing limitation in model explainability. AI

Researchers have developed ClinReadNet, a novel deep learning framework designed to assess the quality of low-dose abdominal CT images. The network mimics radiologists' reading processes by incorporating modules that focus on both local details and overall image context, and by using attention mechanisms to identify regions of interest. Experiments on the LDCTIQAG2023 dataset show ClinReadNet achieves state-of-the-art performance in image quality assessment. AI

Researchers have developed PF-Trans, a novel Transformer model designed for spectral reconstruction in remote sensing. This model integrates physics-based principles and frequency-domain analysis to effectively address spectral aliasing, a common issue in snapshot broadband filter array imaging. PF-Trans achieves state-of-the-art performance, demonstrated by a Peak Signal-to-Noise Ratio of up to 48.50 dB on a specific dataset. AI

Researchers have developed DUET, a novel framework using dual transformer encoders to improve offsite conversion rate prediction in recommendation systems. This approach tailors specific transformer architectures to distinct user behavioral data streams: one for dense click signals and another for sparse, delayed conversion signals. The framework's complementary embeddings are then combined for downstream ranking, demonstrating up to a 0.38% reduction in normalized entropy and leading to consistent improvements in conversion prediction accuracy during A/B testing. AI

Researchers have developed a new machine learning method called PrivacyCredit to address the challenge of using alternative data for credit risk prediction without compromising consumer privacy. This method allows financial institutions to build accurate credit risk models by securely incorporating data such as mobile phone communications, while ensuring that sensitive borrower information remains protected. Experiments on a real-world dataset demonstrated that PrivacyCredit achieves the same predictive performance as traditional methods that would require insecure data sharing. AI

Researchers have developed a deep learning system to improve the resolution of head CT scans by interpolating intermediate slices. This method effectively halves the through-plane spacing, enhancing 3D visualization and reducing noise. The system was evaluated using various loss functions, with MS-SSIM+L1 showing the most promising results, outperforming traditional interpolation techniques. AI

Researchers have introduced Retrieval-based Anomaly Detection (RAD), a novel framework that eliminates the need for task-specific training in anomaly detection. Unlike current methods that rely on costly encoder-decoder models for reconstruction, RAD utilizes a training-free memory-based retrieval system. This approach stores anomaly-free features and detects anomalies by matching test patches against this memory, demonstrating state-of-the-art performance on multiple benchmarks, even in few-shot settings. AI

Researchers have developed a new framework to improve the recognition of Kuzushiji, an ancient Japanese cursive script, even when historical documents are obscured by seals. The proposed system integrates document restoration techniques to counteract seal interference, enhancing character detection, classification, and ordering. This approach significantly reduces character error rates compared to existing methods, particularly on synthetic test data designed to simulate severe seal interference. AI

Researchers have developed MinhwaNet, a system designed to analyze Korean folk paintings. The model found that while it can accurately identify and localize auspicious symbols within the paintings, this object-grounding is insufficient for predicting the painting's genre. Instead, the model performs better when it fuses visual information with textual descriptions, indicating that the arrangement of symbols, rather than their mere presence, is key to genre classification. AI

Researchers have developed Profy, a weakly supervised system designed to aid piano practice by providing interpretable visualizations of motor skills. The system learns from aggregated listener ratings to identify time-aligned passages that differentiate expert and amateur performances. Profy outputs clip-level predictions with evidence scores, enabling learners to focus on specific sections for review and replay, rather than relying on a single global performance score. AI

A new arXiv paper investigates the paradox where smaller block sizes in LLM quantization can degrade model quality. Researchers found this is not an inherent limitation but stems from how statistical clustering interacts with scaling factors. The study proposes solutions like preventing scaling factor underflow and using targeted heuristics such as the 4-over-6 methodology to improve quality, emphasizing the need for tight coupling between hardware and software design for next-generation ML accelerators. AI

Researchers have developed a deep learning framework capable of elucidating molecular structures from one-dimensional NMR spectra. This AI model, inspired by natural language processing techniques, can accurately predict molecular structures with up to 40 non-hydrogen atoms, covering a significant portion of drug-like chemical space. The transformer-based architecture achieves 60.4% accuracy within the top 15 predictions, demonstrating a novel approach to overcoming the combinatorial complexity of structure generation from spectral data. AI