Science – Page 10

Coding’s New Helper: How AI is Fixing Software Bugs

21.01.2026 by qfx

A structured taxonomy organizes tool modules designed to facilitate issue resolution within large language models, establishing a framework for systematic problem-solving.

A new wave of artificial intelligence tools is emerging to automate the tedious task of identifying and resolving issues in software code.

Small Data, Big Canopies: When Deep Learning Finds Trees With Just 150 Images

21.01.2026 by qfx

A comparative analysis of tree canopy segmentation techniques-including YOLOv11, Mask-RCNN, DeepLabv3, Swin-UNet, and DinoV2-demonstrates a range of approaches to defining canopy boundaries, each offering distinct performance characteristics in isolating tree structures from background elements.

A new study reveals that established convolutional neural networks can effectively detect tree canopies with remarkably limited training data, outperforming more recent vision transformer architectures in low-data remote sensing.

Powering Through Uncertainty: A Smarter Grid Restoration Approach

20.01.2026 by qfx

Through iterative policy improvement, a deep reinforcement learning agent integrated with OpenDSS optimizes load restoration, demonstrating a system’s capacity to adapt and refine its performance within a dynamic electrical grid environment.

Researchers have developed a new meta-learning framework that rapidly adapts to changing conditions, improving the resilience of power grids facing increasing renewable energy integration.

Beyond Synthetic Data: Refining Augmentation for Smarter Assistants

20.01.2026 by qfx

Improving the quality of training data is crucial for building more accurate and reliable intent recognition systems, and a new approach focuses on identifying and correcting ambiguous examples.

The Accuracy Paradox: How Training Shapes a Model’s Ability to Detect the Unknown

20.01.2026 by qfx

Across fifty-six ResNet-50 models-all structurally identical but trained with diverse strategies-a tenuous relationship emerges between in-distribution classification accuracy and the ability to reliably detect out-of-distribution data, as demonstrated by a low correlation when evaluating performance via mean AUROC scores across twenty-one detection methods and eight datasets.

New research reveals that simply improving a model’s performance on familiar data doesn’t guarantee it will reliably identify data it hasn’t seen before.

Beyond Intelligence: Why Language Models Excel at Mimicry

20.01.2026 by qfx

The study assessed translation quality by computing the similarity between embeddings of original texts and their “Jabberwockified” translations-generated using large language models-and found that these similarity values, alongside a distribution across 150 diverse 250-word passages spanning fiction, transcripts, and scripts, provide a metric for evaluating the fidelity of the translated content.

The stunning abilities of modern language models aren’t evidence of understanding, but rather a testament to the power of pattern recognition and data compression.

Can AI Grade Critical Thinking? Assessing Skills in Higher Education

20.01.2026 by qfx

The analysis of 21st-century competency reveals distinct error patterns in zero-shot large language model performance, highlighting inherent limitations in assessing complex skills without specific training data.

New research explores how large language models can be used to analyze college curricula and determine the extent to which they foster essential 21st-century skills.

Decoding the Future: A New Approach to Time Series Generation

20.01.2026 by qfx

Synthesized time series data underwent dimensionality reduction and density estimation using t-SNE, principal component analysis, and kernel density estimation to reveal underlying patterns and distributions within the data.

Researchers have developed a novel framework that leverages multi-scale modeling to generate highly realistic and scalable time series data.

Uncovering Hidden Structures in Complex Networks

20.01.2026 by qfx

The system identifies patterns within noisy data by first defining and enumerating cliques and bicliques-subgraphs representing fully connected nodes-and then selectively choosing those cliques to refine the pattern detection process.

A new method reveals high-level patterns in graph data by analyzing reordered adjacency matrices and simplifying common motifs.

Reasoning on a Budget: Smarter AI with Self-Generated Data

20.01.2026 by qfx

Across evaluations on AIME, GPQA, and NaturalPlan, a prompting technique utilizing preference data-designated by a red line in comparative analyses-consistently enhances the reasoning capabilities of models including Grok-3, Grok-3-mini, and GPT-4.1, demonstrating improved performance regardless of token budget limitations and suggesting an inherent ability to foster more robust anytime reasoning.

New research demonstrates a method for large language models to improve their problem-solving abilities during the reasoning process, even with limited computational resources.