training
30 articles about training in AI news
Meta's New Training Recipe: Small Models Should Learn from a Single Expert
Meta AI researchers propose a novel training recipe for small language models: instead of learning from many large 'expert' models simultaneously, they should be trained sequentially on one expert at a time. This method, detailed in a new paper, reportedly improves final model performance and training efficiency.
NVIDIA Advances AI Robotics with Simulation-First Training, Isaac & Jetson
NVIDIA showcased AI robotics advances using foundation models and synthetic environments for training, enabling scalable deployment in real-world sectors like agriculture and solar. Key platforms are the Isaac simulator and Jetson edge AI hardware.
Tiny 9M Parameter LLM Tutorial Runs on Colab, Demystifies Transformer Training
A developer shared a complete tutorial for training a ~9M parameter transformer language model from scratch, including tokenizer, training, and inference, all runnable on Google Colab in minutes.
OpenAI Finishes GPT-5.5 'Spud' Pretraining, Halts Sora for Compute
OpenAI has finished pretraining its next major model, codenamed 'Spud' (likely GPT-5.5), built on a new architecture and data mix. The company reportedly halted its Sora video generation project entirely, sacrificing a $1B Disney investment, to prioritize compute for Spud's launch.
Meta Halts Mercor Work After Supply Chain Breach Exposes AI Training Secrets
A supply chain attack via compromised software updates at data-labeling vendor Mercor has forced Meta to pause collaboration, risking exposure of core AI training pipelines and quality metrics used by top labs.
Video of Massive AI Training Lab in China Sparks Debate on Automation's Scale
A social media post showcasing a vast Chinese AI training lab has reignited discussions about job displacement, underscoring the tangible infrastructure powering the current AI surge.
HIVE Framework Introduces Hierarchical Cross-Attention for Vision-Language Pre-Training, Outperforms Self-Attention on MME and GQA
A new paper introduces HIVE, a hierarchical pre-training framework that connects vision encoders to LLMs via cross-attention across multiple layers. It outperforms conventional self-attention methods on benchmarks like MME and GQA, improving vision-language alignment.
VMLOps Launches 'Algorithm Explorer' for Real-Time Visualization of AI Training Dynamics
VMLOps released Algorithm Explorer, an interactive tool that visualizes ML training in real-time, showing gradients, weights, and decision boundaries. It combines math, visuals, and code to aid debugging and education.
Why Deduplication Is the Most Underestimated Step in LLM Pretraining
A technical article on Medium argues that data deduplication is a critical, often overlooked step in LLM pretraining, directly impacting model performance and training cost. This is a foundational engineering concern for any team building or fine-tuning custom models.
Fine-Tuning LLMs While You Sleep: How Autoresearch and Red Hat Training Hub Outperformed the HINT3 Benchmark
Automated fine-tuning tools now let you run hundreds of training experiments overnight for under $50. Here's how Autoresearch and Red Hat's platform outperformed HINT3, and the tools you can use today.
NVIDIA's PivotRL Cuts Agent RL Training Costs 5.5x, Matches Full RL Performance on SWE-Bench
NVIDIA researchers introduced PivotRL, a post-training method that achieves competitive agent performance with end-to-end RL while using 5.5x less wall-clock time. The framework identifies high-signal 'pivot' turns in existing trajectories, avoiding costly full rollouts.
Training-Free Polynomial Graph Filtering: A New Paradigm for Ultra-Fast Multimodal Recommendation
Researchers propose a training-free graph filtering method for multimodal recommendation that fuses text, image, and interaction data without neural network training. It achieves up to 22.25% higher accuracy and runs in under 10 seconds, dramatically reducing computational overhead.
LeWorldModel: Yann LeCun's Team Achieves Stable World Model Training with 15M Parameters, No Training Tricks
Researchers including Yann LeCun introduce LeWorldModel, a 15M-parameter world model that learns scene dynamics from raw pixels without complex training stabilization tricks. It trains in hours on one GPU and plans 48x faster than foundation-model-based alternatives.
Jensen Huang Predicts AI Training Shift to Synthetic Data, Compute as New Bottleneck
NVIDIA CEO Jensen Huang states AI training is moving from real-world to synthetic data, with compute power becoming the primary constraint as AI-generated data quality improves.
OXRL Study: Post-Training Algorithm Rankings Invert with Model Scale, Loss Modifications Offer Negligible Gains
A controlled study of 51 post-training algorithms across 240 runs finds algorithm performance rankings completely invert between 1.5B and 7B parameter models. The choice of loss function provides less than 1 percentage point of leverage compared to model scale.
Reasoning Training Fails to Improve Embedding Quality: Study Finds No Transfer to General Language Understanding
Research shows that training AI models for step-by-step reasoning does not improve their ability to create semantic embeddings for search or general QA. Advanced reasoning models perform identically to base models on standard retrieval benchmarks.
PRISM Study: Mid-Training on 27B Tokens Boosts Math Scores by +15 to +40 Points, Enables Effective RL
A comprehensive study shows mid-training on 27B high-quality tokens consistently improves reasoning in LLMs. This 'retention-aware' phase restructures 90% of weights, creating a configuration where RL can succeed.
Minimax M2.7 Achieves 56.2% on SWE-Pro, Features Self-Evolving Training with 100+ Autonomous Optimization Loops
Minimax has released M2.7, a model that reportedly used autonomous optimization loops during RL training to achieve a 30% internal improvement. It scores 56.2% on SWE-Pro, near Claude 3.5 Opus, and ties Gemini 3.1 on MLE Bench Lite.
Unsloth Studio: Open-Source Web App Cuts VRAM Usage for Local LLM Training and Dataset Creation
Unsloth has launched Unsloth Studio, an open-source web application that enables users to run, train, compare, and export hundreds of LLMs locally with significantly reduced VRAM consumption. It also converts files like PDFs, CSVs, and DOCXs into training datasets.
Kimi's Selective Layer Communication Improves Training Efficiency by ~25% with Minimal Inference Overhead
Kimi has developed a method that replaces uniform residual connections with selective information routing between layers in deep AI models. This improves training stability and achieves ~25% better compute efficiency with negligible inference slowdown.
OpenSWE Releases 45,000+ Executable Environments for Training SWE Agents, Achieves 66% on SWE-bench Verified
OpenSWE introduces a framework with over 45,000 executable environments for training software engineering agents, achieving 66% on SWE-bench Verified through quality filtering of multi-agent synthesized environments. The Docker infrastructure is open-sourced for full reproducibility.
Goal-Driven Data Optimization: Training Multimodal AI with 95% Less Data
Researchers introduce GDO, a framework that optimizes multimodal instruction tuning by selecting high-utility training samples. It achieves faster convergence and higher accuracy using 5-7% of the data typically required. This addresses compute inefficiency in training vision-language models.
The Coming Revolution in AI Training: How Distributed Bounty Systems Will Unlock Next-Generation Models
AI development faces a bottleneck: specialized training environments built by small teams can't scale. A shift to distributed bounty systems, crowdsourcing expertise globally, promises to slash costs and accelerate progress across all advanced fields.
SAPO: A One-Line Code Fix for Training Stable AI Search Agents
Researchers propose SAPO, a simple modification to stabilize reinforcement learning for search agents, preventing catastrophic training collapse. It delivers +10.6% performance gains with minimal code changes.
StyleGallery: A Training-Free, Semantic-Aware Framework for Personalized Image Style Transfer
Researchers propose StyleGallery, a novel diffusion-based framework for image style transfer that addresses key limitations: semantic gaps, reliance on extra constraints, and rigid feature alignment. It enables personalized customization from arbitrary reference images without requiring model training.
Stanford and Munich Researchers Pioneer Tool Verification Method to Prevent AI's Self-Training Pitfalls
Researchers from Stanford and the University of Munich have developed a novel verification system that uses code checkers to prevent AI models from reinforcing incorrect patterns during self-training. The method dramatically improves mathematical reasoning accuracy by up to 31.6%.
AI Research Accelerator: Autonomous System Completes 700 Experiments in 48 Hours, Optimizing Model Training
An AI system autonomously conducted 700 experiments over two days, reducing GPT-2 training time by 11%. This breakthrough demonstrates AI's growing capability to accelerate scientific research and optimize complex processes without human intervention.
PerContrast: A Token-Level Method for Training More Personalized LLMs
Researchers propose PerContrast, a method that estimates how much each token in an LLM's output depends on user-specific information. By upweighting highly personalized tokens during training, it improves personalization performance by over 10% on average with minimal cost.
Apple's Neural Engine Jailbroken: Researchers Unlock Full Training Capabilities on M-Series Chips
Security researchers have reverse-engineered Apple's Neural Engine, bypassing private APIs to enable full neural network training directly on ANE hardware. This breakthrough unlocks 15.8 TFLOPS of compute previously restricted to inference-only operations across all M-series devices.
ART Framework Automates Reward Engineering, Revolutionizing AI Agent Training
The new ART framework combines GRPO with RULER to automatically generate reward functions, eliminating the need for manual reward engineering in AI agent training. This open-source solution could dramatically accelerate development of capable AI agents across domains.