Machine Learning Research - The Batch | DeepLearning.AI (Page 2)

Flowchart shows data reordering, probability sampling, and effective gradient updating in reinforcement learning.

Machine Learning Research

Faster Reinforcement Learning: New technique auto-selects training examples to speed up fine-tuning

Fine-tuning large language models via reinforcement learning is computationally expensive, but researchers found a way to streamline the process.

Chart details ChatGPT conversations. Writing (28.1%), info-seeking (21.3%), and guidance (28.3%) lead.

Machine Learning Research

What ChatGPT Users Want: ChatGPT users now more likely to be young, female, and seeking info, study shows

What do ChatGPT’s 700 million weekly active users do with it? OpenAI teamed up with a Harvard economist to find out.

Central AI agent icon links to merchant, cart, and payment symbols, illustrating agentic payments process.

Business

Agents of Commerce: Google’s AP2 gives developers new tools to build agentic payments

Google launched an open protocol for agentic payments that enables agents based on any large language model to purchase items over the internet.

Energy-Based Transformer refines predictions step by step, lowering energy for higher context compatibility.

Machine Learning Research

Transformers Energized: Energy-Based Transformers (EBTs) use gradient descent to gradually predict the next token

A new type of transformer can check its work. Instead of guessing the next output token in one shot like a typical transformer, it starts with a rough version of the token and improves it step by step.

Diagram of Qwen3-Next architecture with Mixture of Experts, Gated Attention, and Gated DeltaNet layers.

Machine Learning Research

Qwen3-Next Accelerates: Alibaba’s new model uses hybrid attention layers and a sparse MoE architecture for speed and performance

Alibaba updated its popular Qwen3 open-weights models with a number of fresh, speed-boosting tweaks.

Diagram comparing sliding window attention and ATLAS memory, showing wider context tracking in ATLAS.

Machine Learning Research

10 Million Tokens of Input Context: ATLAS, a transformer-like architecture, can process a context window as large as ten million tokens

An alternative to attention enables large language models to track relationships among words across extraordinarily wide spans of text.

Charts showing PromptGuard 2 blocking attacks, AlignmentCheck detecting goal hijacking, and CodeShield finding insecure code.

Machine Learning Research

Cybersecurity for Agents: Meta releases LlamaFirewall, an open-source defense against AI hijacking

Autonomous agents built on large language models introduce distinct security concerns. Researchers designed a system to protect agents from common vulnerabilities.

Table comparing DINO, DINOv2, DINOv3, SigLIP 2, and PE on segmentation, depth estimation, tracking, and classification tasks.

Machine Learning Research

Better Image Processing Through Self-Supervised Learning: Meta’s DINOv3 gets an updated loss term and improved vision performance

DINOv2 showed that a vision transformer pretrained on unlabeled images could produce embeddings that are useful for a wide variety of tasks. Now it has been updated to improve the performance of its embeddings in segmentation and other vision tasks.

Charts showing how much information language models memorize based on model size and training dataset size; one graph uses random data, the other uses real-world text.

Machine Learning Research

Does Your Model Generalize or Memorize: Researchers find models with more parameters copy more bits from training sets

Benchmarks can measure how well large language models apply what they’ve learned from their training data to new data, but it’s harder to measure the degree to which they simply memorized their training data. New work proposes a way to gauge memorization.

Three AI-generated video clips: a man vaulting over a moving car, a gymnast flipping on a plane wing, and a rabbit ice skating in pink boots.

Machine Learning Research

Mixture of Video Experts: Alibaba’s Wan 2.2 video models adopt a new architecture to sort noisy from less-noisy inputs

The mixture-of-experts approach that has boosted the performance of large language models may do the same for video generation.

Diagram showing SWE-Smith AI pipeline for generating synthetic coding tasks from real repositories using multiple strategies

Machine Learning Research

Training Data for Coding Assistants: Stanford and Alibaba build bug fixing dataset and pipeline to train AI

A bottleneck in fine-tuning large language models for software engineering is building a dataset that can show them how to edit code, search for subroutines, write test scripts, control a terminal, manage a file system, and so on. Researchers built a pipeline that produces such data automatically.

Bar charts comparing GPT-5’s speed, price, and intelligence vs. top AI models like Claude, Gemini, and Grok in 2025 benchmarks.

Machine Learning Research

GPT-5 Takeoff Encounters Turbulence: OpenAI's new model hits turbulence with cost, performance, and API complaints

OpenAI launched GPT-5, the highly anticipated successor to its groundbreaking series of large language models, but glitches in the rollout left many early users disappointed and frustrated.