Brief

A developer successfully contributed a new skill to the popular ECC GitHub repository, which functions as an agent harness system. This repository has garnered over 211,000 stars, indicating significant community interest and adoption. The integration of the new skill highlights the collaborative and open-source nature of AI development, allowing for community contributions to enhance agent capabilities. AI

This article outlines a 45-minute diagnostic process for startups to audit and control their spending on large language models (LLMs). It emphasizes that LLM costs often escalate due to numerous small, unmonitored calls across various functions like retries, background jobs, and internal tools, rather than single expensive prompts. The audit involves mapping all LLM call paths, attaching costs to specific units of value, identifying waste from retries and tool calls, strategically assigning tasks to cheaper models where appropriate, and implementing budget guardrails with clear ownership. AI

This article highlights seven specific slash commands within Claude Code that can significantly boost developer productivity. The author claims these commands have saved them over 10 hours per month by streamlining common coding tasks. The piece suggests that many users are not fully leveraging Claude Code's capabilities, leading to wasted time. AI

This article addresses the "Claude Code Error 429: Rate Limit Exceeded," a common issue encountered when using Anthropic's AI models. It explains that this error signifies that too many requests have been made to the API within a given timeframe. The piece offers guidance on how to resolve this by implementing strategies such as exponential backoff, request queuing, and optimizing API calls to manage usage and avoid hitting rate limits. AI

A slow PyTorch training job may not be due to the model's complexity but rather the data loading process. The article explains how to identify if your GPU is being starved of data by a slow DataLoader. It suggests methods to diagnose and resolve these performance bottlenecks. AI

Researchers have developed an "Omnivorous Vision Encoder" to improve how AI models understand different visual data types. This new framework fine-tunes existing vision encoders, like DINOv2, to create a unified feature space. The goal is to ensure that an AI can recognize the same scene consistently, whether it's presented as a standard RGB image, a depth map, or a segmentation map. AI

Researchers have developed a framework to quantify the value of perception, prediction, communication, and common sense in decision-making systems. Their work defines these quantities in a decision-theoretic manner, with information-theoretic parallels to concepts like Shannon entropy. An interesting finding is that perception alone can have negative value, whereas its combination with prediction, or prediction by itself, is always non-negative. These definitions aim to answer practical questions for designing autonomous systems, such as the importance and optimal order of observing and predicting agent behaviors, and may also offer insights into cognitive and neural processes. AI

Researchers have developed a diffusion-based model, dubbed CH-aware DMT, to reconstruct synthetic EUV images of the sun from historical He I observations. This method aims to extend the availability of solar EUV imaging data into earlier periods before modern satellites like SDO were operational. The model was trained on co-aligned SOLIS He I and AIA 193 Å data and demonstrated strong performance in preserving EUV morphology and coronal hole structures. Its historical applicability was further validated by comparing reconstructions with data from SOHO/EIT, Yohkoh/SXT, and independent solar activity proxies, suggesting its utility for multi-decade analyses of coronal evolution. AI

A new research paper explores the impact of different Stage-1 training methods on vision-language models (VLMs). The study found that while Stage-1 training, such as supervised fine-tuning (SFT) or on-policy distillation (OPD), leads to similar in-domain performance, it significantly influences the entropy regime of the model. Specifically, OPD results in higher policy entropy and answer diversity compared to SFT, although these advantages diminish after the Stage-2 reinforcement learning phase. AI

Researchers have developed AQIFormer, a new transformer-based architecture designed to classify air quality using images. This model integrates front and rear traffic imagery with weather data, improving cross-city generalization and achieving 89.96% accuracy on a large dataset. AQIFormer demonstrates strong performance even on unseen cities, with minimal accuracy degradation when adapted with few-shot learning. AI

Researchers have developed a machine learning approach using Random Forest Recursive Feature Elimination (RF-RFE) to identify key indicators for measuring poverty and inequality in Nigeria. By analyzing household survey data, the study found that a small set of income sources, consumption categories, and household characteristics can accurately predict poverty status and welfare distribution position. This method could significantly reduce the data requirements for future surveys, enabling more efficient monitoring of poverty and inequality in low- and middle-income countries. AI

Researchers have developed a new framework for strawberry harvesting robots to improve their visual perception and self-recovery capabilities. The SRR-Net system integrates fruit detection, segmentation, and ripeness assessment with gripper alignment correction. This system uses a micro-optical camera for real-time feedback, enabling adjustments during grasping and predicting slippage to recover or abort harvesting cycles. AI

Researchers have developed BCG-FM, a novel foundation model for analyzing cardiac health through ambient mechanical biosignals. This model utilizes a piezoelectric sensor embedded in a bed surface to record ballistocardiography (BCG) data overnight, requiring no user effort. Pretrained on 2.75 million hours of recordings from nearly 146,000 individuals, BCG-FM achieved a 3.26-year Mean Absolute Error in biological age estimation and demonstrated clinically relevant discrimination across various health conditions. AI

Researchers have developed transformer-based models to analyze Nepali memes for hate speech and sentiment. The study focused on text extraction from memes, employing OCR and subsequent analysis with transformer architectures. Experiments showed that a decoder-only model excelled at binary hate speech detection, while a soft voting ensemble approach improved sentiment analysis performance by 15.8% in Macro F1-score. AI

Researchers have developed XAInomaly, a new framework utilizing a semi-supervised deep contractive autoencoder for anomaly detection in open radio access networks (O-RAN). This approach aims to learn normal network behavior and identify deviations indicative of anomalies. To overcome the 'black-box' nature of deep learning, the framework incorporates a reactive explainable AI technique called fastshap-C. AI

A new research paper introduces UnWeaver, a framework that simplifies Graph-based Retrieval-Augmented Generation (RAG) systems. UnWeaver disentangles document content into entities, which are then used to recover original text chunks, preserving source fidelity. The study argues that this entity-based decomposition offers a more distilled representation and reduces noise. Experiments indicate that VectorRAG performs comparably to current state-of-the-art graph-based solutions at a significantly lower cost. AI

Researchers have proposed a new knowledge representation system that combines dense embeddings with quantum-fuzzy logic. This hybrid approach aims to overcome the trade-offs between probabilistic and crisp inference found in current LLM and ontology integrations. The proposed neuro-quantum-fuzzy systems could enable knowledge representation that supports both classical and contextual reasoning. AI

Researchers have developed a novel Physics-Informed Neural Network (PINN) framework that integrates auxiliary knowledge from sources beyond experimental data. This new approach enhances parameter discovery by incorporating information from peer-reviewed literature and network structures, specifically applied to modeling microbial interactions. The framework demonstrated significant improvements in accuracy and predictive power for microbial community modeling, outperforming existing methods and revealing ecological insights. AI

This thesis introduces novel frameworks for group emotion recognition in real-world scenarios, prioritizing privacy by analyzing collective audio-video signals rather than individual cues. The proposed cross-attention multimodal architecture with Frames Attention Pooling (FAP) and a Variational Encoder Multi-Decoder (VE-MD) framework demonstrate competitive performance without relying on individual facial or vocal data. These contributions aim to advance affective computing by enabling privacy-safe group emotion analysis. AI

Researchers have introduced CHIMERA-Bench, a new benchmark dataset designed to standardize and advance computational antibody design. This benchmark addresses the lack of a common evaluation framework by providing a unified task, a curated dataset of antibody-antigen complexes, and a comprehensive protocol with novel epitope-specificity measures. It aims to enable fair comparison and development of deep generative methods for antibody design, testing their generalization capabilities across various splits. AI

Researchers have developed a new hybrid model for predicting customer churn on structured data, combining a feature-tokenized transformer (FT-Transformer) with XGBoost. This approach aims to capture complex feature interactions and improve probability calibration, addressing challenges like class imbalance and nonlinear relationships. Tested on a public bank churn dataset, the model achieved an F1 score of 62.10% and an AUC-ROC of 0.861, outperforming a standard Multi-Layer Perceptron baseline. AI

Researchers have developed a hierarchical feature engineering framework to automatically classify vocal hyperfunction subtypes using neck-surface acceleration data. This method integrates static, dynamic, ratio-based, and coupling features to distinguish between phonotraumatic (PVH), non-phonotraumatic (NPVH), and healthy vocal patterns. The framework achieved an AUC of 0.891 for PVH and 0.728 for NPVH, highlighting the importance of coupling features for accurate classification. AI

Researchers have developed ZIPP, a novel method for zero-shot image personalization that conditions text-to-image diffusion models on natural-language personas. This approach allows for personalized image generation without requiring any user-specific data or model weight updates, addressing the cold-start problem and context-dependent preferences. ZIPP utilizes a large language model to rewrite prompts from the perspective of a persona, and personas are mined at scale using a graph attention network trained on a large Reddit interaction graph. The system was evaluated on ZIPBench, a new benchmark, and demonstrated significant improvements in personalization and reduced subpopulation bias compared to generic generation and fine-tuned baselines. AI

Researchers have evaluated seven uncertainty quantification (UQ) methods on the AION-1 astronomical foundation model for predicting galaxy properties. Conformal prediction methods, particularly the Locally Valid and Discriminative (LVD) framework, demonstrated superior calibration and local validity compared to non-conformal baselines. The study suggests LVD is the preferred UQ approach for foundation model embeddings in astrophysics, offering more reliable uncertainty estimates for scientific inference. AI

Researchers have developed FIT-Print, a novel method for verifying ownership of open-source AI models that is resistant to false claim attacks. Existing fingerprinting techniques are vulnerable to adversaries falsely claiming ownership of independent models. FIT-Print addresses this by using targeted signatures derived from model outputs and feature attributions, achieving a 100% defense success rate against false claims and a 0.0% false alarm rate on independent models. AI

Researchers have developed a new framework for classifying airborne multispectral point clouds, which combine 3D spatial and spectral information. The proposed method utilizes a two-stream feature fusion approach with attention mechanisms to enhance the representation of complex spatial-spectral data. It also incorporates a joint loss function to address challenges like unbalanced sample distribution and spectral similarity between classes. Experiments on two datasets show the framework outperforms existing state-of-the-art methods. AI

Researchers have developed TriHead-GAN, a novel generative adversarial network designed to create synthetic carbon emission time series data. This model addresses the scarcity of high-frequency monitoring data, which hinders deep learning applications in climate policy and regulation. TriHead-GAN's unique triple-head discriminator ensures the generated data accurately reflects cross-variable correlations and realistic temporal variability, outperforming existing methods in experiments. AI

Researchers have developed AMN, an Adaptive Multi-Scale Fusion Network designed for precise nuclei segmentation in histopathology images. This dual-encoder framework uniquely combines a Swin Transformer and a ResNet-50 feature pyramid, using a learned gating mechanism to dynamically balance their contributions across different scales. AMN incorporates a multi-objective loss function that includes focal loss, boundary-aware loss, and an uncertainty-modulated classification term to improve accuracy and reduce overconfident errors. The model achieved state-of-the-art results on the CoNIC benchmark, outperforming eight other architectures and demonstrating robust generalization capabilities on the MoNuSeg dataset. AI

Researchers have developed SHIELD-IDS, an enhanced intrusion detection system designed to combat adversarial attacks on machine learning models. The system integrates gradient boosting models like XGBoost and LightGBM into a diverse ensemble, protected by a three-layer defense mechanism. Experiments show SHIELD-IDS maintains over 99% detection accuracy on clean data and demonstrates improved robustness against common adversarial attack methods. AI

Researchers have identified a new security vulnerability in brain-computer interface (BCI) systems that integrate with large language model (LLM) agents. This vulnerability, termed "brain-prompt injection," allows attackers to manipulate the agent's actions by subtly altering neural signals, even if monitoring systems remain unaware. The study proposes a "Route-Safety Audit Contract" to enhance security by defining a minimal log schema and endpoint specification, demonstrating its effectiveness in mitigating certain attacks. AI

A new study published on arXiv explores the effectiveness of satellite-based flood mapping using geospatial foundation models. Researchers found that the accuracy of these models is significantly influenced by land cover and the type of flood event, with cropland and riverine floods showing better detection. The study also highlighted that inconsistencies between different reference products can be mistaken for model errors, and identified 23 failure modes, suggesting pipeline engineering is a more critical factor than model capacity for operational reliability. AI

Researchers have developed SatIR, a novel retrieval system designed to improve the matching of patients to clinical trials. This system goes beyond simple semantic similarity by treating trial eligibility criteria as formal constraints that must be satisfied. SatIR integrates Satisfiability Modulo Theories (SMT), relational algebra, medical ontologies, and LLMs to convert complex clinical information into executable constraints, enabling more accurate and efficient trial matching. AI

Researchers have developed a novel method for transferring information between different types of single-cell biological data. By using adversarial fine-tuning on foundation models, their approach can translate spatial transcriptomics data into single-cell RNA sequencing data, even when the datasets are unpaired. This technique shows promise in recovering spatial information from scRNA-seq data and outperforms existing multi-omics translation methods. AI

Researchers have developed PolyBuild, a novel end-to-end method for extracting building polygon contours directly from high-resolution remote sensing images. This approach bypasses the need for computationally intensive post-processing steps common in existing methods. PolyBuild utilizes an Initial Contour Generation Module for initial extraction and a Contour Optimization Module, incorporating CNN and Transformer features, to refine the contours, achieving superior performance on multiple datasets. AI

Researchers have developed a novel control system for tilt-rotor drones, which are known for their advanced maneuverability but also their inherent instability. Initial attempts using direct neural network control with MLPs, LSTMs, and transformers proved unsuccessful in stabilizing the system. The team's main contribution is a hybrid approach that combines a neural network with a sliding mode controller, where lightweight networks learn specific system dynamics from flight logs, significantly improving robustness and reducing computational load. AI

Researchers have introduced the DIYHealth Suite, a new framework aimed at advancing AI-powered health management within home settings. This suite includes a large-scale multimodal dataset called DIYHealth-900K, designed to capture diverse real-world home care scenarios. It also features DIYHealthGPT, an adaptive foundation model utilizing a novel Hybrid Hyper Low-Rank Adaptation technique, and DIYHealthBench, the first benchmark specifically for evaluating foundation models on home care tasks. Experiments show DIYHealthGPT achieving state-of-the-art performance across 11 home care tasks. AI

Researchers have developed RAPID, a novel framework designed to make Vision Transformers (ViTs) more computationally efficient. This method intelligently prunes and merges tokens based on their layer-specific characteristics, addressing the quadratic complexity of self-attention. In earlier layers, RAPID removes redundant local patterns, while in deeper layers, it merges less critical tokens while preserving important ones, guided by attention weights. Experiments on ImageNet-1K showed RAPID achieving a better accuracy-compression trade-off than existing methods, especially under aggressive compression. AI

Researchers have developed a new method to interpret how models designed to detect suicide ideation internally represent psychological risk factors. This approach moves beyond simple accuracy metrics to analyze the model's internal representations using visualization and geometric analysis. The study found that topic-aware data augmentation significantly improves the clarity and distinctness of representations for factors like family issues and financial crises, suggesting it enhances both performance and interpretability. AI

Researchers have developed a new method called CEF-Log for using Large Language Models to detect malicious web server logs. This approach uses a structured five-step reasoning template to guide the LLM, improving its ability to analyze logs and generate legally sound explanations. CEF-Log demonstrated high accuracy with minimal examples, achieving an F1-score of 0.99 on a known dataset and showing a tenfold increase in sample efficiency compared to other methods. A new dataset, ForenWebLog, was also introduced to evaluate the system on more complex, real-world attack scenarios. AI

Researchers have developed a system for precision agriculture that uses Spatio-Temporal Graph Neural Networks (STGCN) and a Transformer-based model to forecast weather for the next 30 days across 1,359 locations in Nepal. The STGCN model demonstrated superior accuracy in predicting weather patterns. This system combines weather forecasts with soil data to provide localized crop recommendations and includes a Retrieval-Augmented Generation chatbot to answer farmers' questions in natural language, all accessible via a mobile application. AI

Researchers have developed a new framework to evaluate autonomous cyber defense agents that configure commercial Endpoint Detection and Response (EDR) systems. This framework addresses the challenge of a "sim-to-real" gap, where autonomous agents interact with complex, black-box EDR tools like Microsoft Defender XDR. The evaluation, conducted in a simulated Active Directory environment, revealed that commercial EDR telemetry is not optimized for benchmarking, and the autonomous EDR behavior can fluctuate during testing. AI

A new paper on arXiv explores the concept of instrumental convergence and power-seeking behavior in artificial intelligence. The author argues that the concern over AI posing an existential risk due to power-seeking rests on a strong interpretation of the instrumental convergence thesis. The paper examines existing defenses of this thesis and concludes that they do not sufficiently support the argument for power-seeking AI, with implications for AI governance and long-term risk studies. AI

Researchers have developed a new framework for fine-tuning diffusion models to enhance compositional control in landscape image generation. This method uses a four-dimensional compositional anchor vector, integrated via a decoupled cross-attention mechanism, to guide image creation. Evaluations show significant improvements in horizon detection and adherence to the rule of thirds, with precision found to be category-dependent. AI

A new study published on arXiv compares the quality of feedback provided by Large Language Models (LLMs), Small Language Models (SLMs), and human instructors on technical writing assignments. The research found that a locally hosted SLM, specifically a quantized Llama-3.1, performed comparably to GPT-4 and was preferred by students for readability and actionability in technical courses. However, human feedback was still favored for highly specialized writing tasks, suggesting a tiered approach where AI handles foundational feedback and instructors focus on conceptual guidance. AI

Researchers have introduced a new dataset and benchmark designed to improve urban traffic perception by aligning street-level and aerial drone views. This benchmark focuses on two key tasks: matching object tracks across these different viewpoints and predicting a bird's-eye view from monocular street-level imagery using aerial supervision. The dataset aims to advance research in cross-view perception and urban scene understanding, providing standardized evaluation tools and baseline implementations for these challenging tasks. AI

Researchers have developed RadOT-Eval, a novel framework for evaluating the accuracy of AI-generated radiology reports. This system breaks down reports into structured clinical evidence units and uses optimal transport to align corresponding pieces of information. RadOT-Eval demonstrated strong correlations with human-annotated error burdens, outperforming existing metrics and an LLM-based evaluator on independent datasets. AI

Researchers have developed a new training strategy for AI safety judges, aiming to improve their consistency and reliability. The strategy involves using dynamic rubrics generated from prompt-response-label triples to expose judges to varied evaluation criteria. A curriculum approach progressively introduces these dynamic rubrics after initial training on fixed rubrics, leading to a 12B model that achieves high accuracy and stability across different rubric formulations. AI

Researchers have developed and evaluated six strategies for training deep learning models to segment white matter hyperintensities and stroke lesions in MRI scans, particularly when dealing with partially labeled datasets. Their analysis, conducted on a large cohort of 2,052 MRI volumes, found that pseudolabeling was the most effective method for improving model performance. This approach demonstrates the potential for creating reliable automated segmentation tools to aid in monitoring cerebral small vessel disease and extracting biomarkers for clinical research. AI

Researchers have developed a new framework using Large Language Models (LLMs) to decode continuous emotional dynamics from brain activity. This approach moves beyond traditional discrete classification by employing multi-target regression to track overlapping emotional dimensions as continuous trajectories over time. By analyzing functional connectivity in fMRI data and using LLM-generated sentiment profiles from narrative text, the study demonstrates that dynamic neural network interactions better capture emotional states than static brain region representations. AI

Researchers have developed CURE, a new framework designed to improve the accuracy and reliability of AI-generated radiology reports. This error-aware curriculum learning approach enhances visual grounding and factual consistency without requiring additional data. By dynamically adjusting training to focus on more challenging samples, CURE significantly boosts grounding accuracy, report quality, and reduces instances of AI-generated hallucinations. AI