PAPER PLAINE

Researchers discovered that breaking EEG brain signals into discrete chunks called microstates—rather than treating them as continuous streams—helps machine learning systems recognize patterns better. This microstate approach outperformed traditional methods across multiple tasks including sleep detection, emotion recognition, and motor control, while also making the AI's decisions easier for humans to interpret.

When jobs grow in countries where Nepalis work, more money flows home as remittances — but when Nepal's central bank tightens monetary conditions, remittances shrink. The analysis of 30 years of data shows remittances could reach 28% of Nepal's GDP by 2030, making the country's economic stability heavily dependent on foreign employment markets and sensitive to sudden external shocks.

Large language models are often tested on datasets they've already seen during training, making their scores meaningless—like letting students study the exact exam questions beforehand. Researchers propose creating "contamination-resistant" datasets that models can use during evaluation but cannot learn from during training, and show how to build them using differences between how Transformers train versus perform inference.

Vision-language AI models are being held back not by weak reasoning skills but by poor visual perception. Researchers found that training models in three separate stages—first visual perception, then visual reasoning, then textual reasoning—improves performance by up to 5.2% on visual math tasks while cutting reasoning explanations by a fifth, suggesting that better eyesight reduces the need for laborious thinking.

Researchers demonstrated a new way to control how quickly quantum information decays in qubits by engineering surfaces that manipulate low-frequency magnetic noise around them. Unlike previous approaches that focused on spontaneous emission, this method targets pure dephasing—the gradual loss of quantum coherence—using specially designed cobalt-iron-boron metasurfaces placed near nitrogen-vacancy centers in diamond. The technique opens a new path for protecting quantum information without relying on optical engineering.

A new training method called General Preference Reinforcement Learning (GPRL) lets AI models improve at open-ended tasks like writing and reasoning without collapsing into narrow reward-gaming behavior. The approach treats quality as multidimensional rather than a single score, and achieved 56.51% win rate on standard benchmarks while outperforming existing methods across multiple evaluation tests.

When new firms constantly enter a market while others fail, the overall system eventually settles into a predictable pattern—even though entry and exit rates are unequal. The research identifies three distinct phases in how market populations evolve and discovers that there's an optimal exit rate that minimizes how long it takes for the market to reach major milestones, showing that firm turnover isn't just random turbulence but can be deliberately shaped.

AdaGrad, a foundational algorithm used in machine learning optimization, can successfully navigate noisy training environments where extreme outlier values occur—without needing extra safeguards like gradient clipping that other methods require. This finding applies when the noise follows heavy-tailed distributions and the algorithm automatically adapts to the severity of the problem without advance warning.

Researchers created URGE, a new method that improves how diffusion models (AI systems that generate images) follow instructions at the moment of creation—without requiring expensive mathematical calculations. The method assigns lightweight weights to different generation paths and occasionally filters out the worst ones, producing better results than existing techniques while being simpler and faster to run.

Researchers created a large, high-quality benchmark dataset and a new training method that can classify protein structures more efficiently than existing approaches. The new method, called Masked Invariant Autoencoders, works by hiding up to 90% of a protein's structure during training and learning to reconstruct it—a strategy that scales better than current methods while achieving superior performance on protein fold classification tasks.

Drones lose track of their position when sensors cut out or vibrate unpredictably—problems that stumped earlier tracking systems. Researchers built a learning-based filter that adapts to these disruptions in real time, using a neural network to adjust how much it trusts past measurements versus new sensor data. On real drone flights, it stayed accurate longer than standard methods when sensors went dark.

Researchers built an artificial intelligence system that can predict the magnetic structure of materials by looking only at their atomic arrangement—without running expensive experiments or complex physics simulations. The model handles both simple magnetic patterns and the complex, twisted arrangements found in real materials, reconstructing experimentally measured structures with high accuracy.