LLM
What’s new Adding the right metadata can make LLM pretraining faster and more effective—and it’s not just about URLs. * Fine‑grained signals work: Prepending detailed quality indicators helps models learn quicker. * Append-and-predict: Appending metadata and training the model to predict it as an auxiliary task boosts efficiency.
AI
How do we know if an AI model is truly aligned with human preferences—helpful, honest, safe, and instruction-following? This paper explores a surprisingly effective shortcut: judge the judges. Instead of grading a model’s open-ended answers (which needs lots of human effort or very strong AI judges), the authors
AudioAI
AI that pulls the vocals out—fast Ever wished you could solo a singer from any track without artifacts? This research shows how, using a latent diffusion model that separates vocals from full mixes. * Efficient: Generates in a compact audio latent space, then decodes to sound—speeding up training and
AI
How do we know an AI is truly helpful, honest, safe—and follows instructions? Today, we usually read its answers and score them, which can be slow, costly, and subjective. This paper finds a simple clue: when compared to a trusted reference, models that write well also judge well. The
QuantumComputing
What if we could let evolution design quantum circuits that compress data? This paper explores exactly that with quantum autoencoders—models that shrink high‑dimensional quantum or classical data while keeping the important bits. The authors introduce a neural architecture search framework powered by a genetic algorithm. Instead of hand‑
AIResearch
Long, well-sourced answers are hard for AI to learn: most training rewards short, easily graded Q&A. This work introduces Reinforcement Learning with Evolving Rubrics (RLER) - grading guides that co-evolve with the model, so feedback stays aligned with what the model actually explores during multi-step research. Built with
AI
Probing AI Isn’t So Simple To keep AI models honest, researchers train tiny “probes” that look inside a model’s activations to flag behaviors like deception or sycophancy. But real examples of these behaviors are rare, so teams often use synthetic AI-generated data instead. This study tested how well
MachineLearning
Training AI usually needs lots of human-labeled examples. This work shows a different path: make models learn from plenty of unlabeled data by asking them to stay consistent under strong "noise." Instead of simple tweaks (like small crops or word drops), they use powerful data augmentation: RandAugment for
RAG
Which LLM wins at RAG Q&A? RAG (Retrieval-Augmented Generation) helps reduce hallucinations by grounding answers in source documents. This study tested five 7B-class models on computer science literature Q&A to compare accuracy and speed. * GPT‑3.5 + RAG answered both yes/no and long-form questions effectively.
AI
AI agents are getting good at coding and research, but they still struggle with long, domain-specific projects. Training them is hard: full-time human annotation is too costly, while pure trial-and-error rarely finds successful paths. Apollo is a new sampling framework that blends lightweight, asynchronous human guidance with strict action-level filtering.