Deep Reinforcement Learning Gets a Bayesian Boost
A new method combines the power of deep learning with Bayesian principles to tackle complex reinforcement learning tasks with improved efficiency and accuracy.
Science
Why Won’t It Just *Do* What You Ask? Unpacking the Quirks of AI Language
New research reveals large language models often prioritize ease over accuracy, but surprisingly excel at remembering details over extended conversations.
Faster Aerodynamic Design with Graph Networks and Smart Data
![The study demonstrates that test Mean Squared Error (testMSE) scales predictably with training set size ([latex]D_D[/latex]), where [latex]D_D[/latex] represents the number of unique geometry-flow snapshots used as graph-based data, and this scaling behavior differs significantly across models of varying sizes.](https://arxiv.org/html/2512.20941v1/figs/scaling/mean-error-all-models.png)
Researchers have created a new dataset and scaling laws to accelerate aerodynamic simulations using graph neural networks, enabling efficient design even with limited data.
When Conversations Get Confused: A New Test for Chatbot Clarity
Researchers have created a benchmark and framework to help conversational AI better navigate ambiguity and ask clarifying questions during extended dialogues.
Beyond the Network Boundary: Adapting Traffic Analysis to New Environments
Detecting unusually large network traffic flows – ‘elephant flows’ – becomes significantly harder when models are moved between different network setups, and this research tackles that challenge.
When Memories Fade: Understanding Forgetting in AI
New research reveals how the depth of knowledge representation impacts a model’s ability to retain information when learning new tasks, offering a path towards more robust artificial intelligence.
What Large Language Models Still Don’t Know
A new method reveals critical weaknesses in today’s most powerful AI systems and highlights shortcomings in how we measure their abilities.
Squeezing Value from Spot Instances for Large Language Model Training
New research details a smart scheduling framework that minimizes costs and meets deadlines when fine-tuning massive AI models using fluctuating cloud GPU pricing.
Strength in Numbers: A New Defense Against AI Attacks
A novel system leveraging a mixture of experts significantly improves the robustness of machine learning models against carefully crafted adversarial inputs.
The Deep Learning Scaling Puzzle: Why Bigger Isn’t Always Better
![Internal feature learning in deep residual networks collapses with increasing depth-at a rate of [latex] 1/\sqrt{L} [/latex]-but this degradation is rectified by a depth-aware learning rate, [latex] \eta_1 = \eta_c n \sqrt{L} [/latex], which restores active learning across layers and enables consistent hyperparameter transfer and improved performance, as demonstrated by lower training and testing losses and higher accuracy even with varying network depths and widths.](https://arxiv.org/html/2512.21075v1/figures/Vanish_resnet_performence_acc_loss.png)
New research reveals how the dynamics of feature learning in deep neural networks explain both the successes and limitations of simply scaling up model size.