Hugging Face introduces Model Cards, standardized documentation sheets for AI models covering architecture, training data, performance metrics, limitations, and use cases. This tool enhances transparency and simplifies model evaluation before deployment.
Hugging Face adds probabilistic time series forecasting to Transformers library. New feature generates probability distributions for predictions beyond point estimates. Includes pre-trained models and fine-tuning support.
Hugging Face launches interactive ML demos directly on arXiv, enabling readers to test models and paper results without leaving the platform. Native integration of Hugging Face Spaces into scientific articles.
Hugging Face introduces contrastive search, a decoding method for transformers that generates human-level text. The technique balances likelihood and lexical diversity, improving quality over standard beam search and sampling approaches.
Hugging Face releases a guide to fine-tune Whisper, OpenAI's speech recognition model, on multilingual data using the Transformers library. The method covers audio preprocessing, training, and evaluation across multiple languages.
Hugging Face releases an evaluation tool to measure bias in language models. The Evaluate platform enables systematic testing of gender, racial, and other biases in LLMs through standardized benchmarks.
Hugging Face introduces DOI (Digital Object Identifier) for datasets and models on its platform. DOIs enable standardized academic citation and permanent traceability of AI resources, facilitating reproducibility and scientific integrity.
Hugging Face demonstrates ultra-fast BLOOM inference using DeepSpeed and Accelerate. Quantization and parallelization optimizations reduce latency and memory consumption. Benchmarks show significant gains on multi-GPU setups.
Hugging Face Optimum integrates Vision Transformers on Graphcore processors. The guide details optimization of vision models for IPU inference, with performance benchmarks and reproducible code.
OpenAI trains critique-writing models to describe flaws in summaries. Human evaluators detect significantly more flaws when shown AI-generated critiques. Larger models excel at self-critique, with scale improvements greater for critique than summary generation.
Hugging Face introduces inference acceleration through Optimum and Transformers Pipelines, reducing latency and memory consumption for language models in production.
OpenAI publishes a research agenda to assess economic impacts of code generation models. The study examines how these tools affect developer productivity, employment, and wages, with no definitive conclusions yet.
Stable-baselines3, a Python library for reinforcement learning, integrates with Hugging Face Hub. Trained models can be shared, versioned, and reused directly through the platform.
Hugging Face Infinity achieves millisecond latency inference on modern CPUs. Case study demonstrates model optimization and performance without GPU requirements.
Hugging Face publishes a guide to fine-tune XLSR-Wav2Vec2, a multilingual speech recognition model, for low-resource languages. Uses the Transformers library to adapt the pre-trained model with limited data.
OpenAI trained a system solving grade school math word problems with nearly twice the accuracy of fine-tuned GPT-3. The system achieves 55% accuracy on tests where 9-12 year olds score 60%.
OpenAI shows that fine-tuning on a small curated dataset improves language model behavior with respect to specific behavioral values. Pragmatic approach to align models without requiring massive data volumes.
Hugging Face releases a guide for distributed training of BART and T5 on Amazon SageMaker. Uses HF Transformers with multi-GPU/multi-node optimizations for summarization tasks. Includes code, benchmarks, and best practices.
Hugging Face integrates Ray Tune for automated hyperparameter optimization in the Transformers library. This integration enables researchers and practitioners to efficiently tune training parameters (learning rate, batch size, etc.) using distributed search algorithms.
OpenAI co-authors multi-stakeholder report (58 authors, 30 organizations) proposing 10 mechanisms to improve verifiability of AI system claims. Tools enable developers to provide evidence of safety, security, fairness, and privacy-preservation; allow users, policymakers, and civil society to evaluate AI development processes.
OpenAI introduces a method to assess neural network classifier robustness against adversarial attacks unseen during training. The UAR (Unforeseen Attack Robustness) metric measures a single model's ability to withstand unanticipated attacks and emphasizes the need for performance evaluation across diverse unforeseen attack scenarios.
OpenAI reports progress on stable and scalable training of energy-based models (EBMs) with improved sample quality and generalization. EBM generation refines answers iteratively, producing samples competitive with GANs at low temperatures while maintaining mode coverage guarantees of likelihood-based models.
OpenAI publishes a paper arguing that long-term AI safety requires social scientists to ensure alignment with human values. The company plans to hire social scientists full-time.
OpenAI discovers that gradient noise scale predicts the parallelizability of neural network training across tasks. Complex tasks have noisier gradients, enabling larger batch sizes in the future and removing a potential constraint on AI system scaling.
OpenAI develops an energy-based model that learns spatial concepts (near, above, between, closest, furthest) from only five demonstrations. The model transfers concepts across domains: concepts learned in 2D particle environments applied to 3D physics-based robot tasks.
OpenAI proposes an AI safety technique where agents debate topics with each other, with a human judge determining the winner. Approach aims to improve model alignment and interpretability through debate-based learning.
OpenAI releases Evolved Policy Gradients (EPG), a metalearning approach that evolves the loss function of learning agents. EPG-trained agents generalize to novel tasks unseen during training, such as navigating to objects placed on different sides of a room.
OpenAI presents an interpretable machine learning method where AIs teach each other using examples understandable to humans. The approach automatically selects the most informative examples to convey a concept, experimentally validated for both AI and human learning.
OpenAI releases seven unsolved research problems identified during internal work. The initiative aims to engage the scientific community on key AI challenges.
OpenAI presents a method for learning sparse neural networks through L₀ regularization. The technique reduces the number of active parameters without sacrificing performance, improving computational efficiency and model generalization.
OpenAI develops a hierarchical reinforcement learning algorithm that learns reusable high-level actions. Applied to navigation, the agent discovers primitives (walking, crawling) and rapidly solves tasks requiring thousands of timesteps.
OpenAI presents robotics techniques enabling robot controllers trained entirely in simulation to adapt to unexpected environmental changes during physical deployment. Shift from open-loop to closed-loop systems.
OpenAI releases an asymmetric actor-critic method for robot learning: a lightweight actor runs on-robot while a heavy critic trains offline. Approach tested on object manipulation tasks.
OpenAI combines domain randomization and generative models to improve robotic grasping. The approach trains models on varied synthetic data, then transfers them to real robots without additional fine-tuning.
OpenAI shows self-play enables AIs to discover complex physical skills (tackling, ducking, faking, kicking, catching) without explicit environment design. The environment automatically adapts to optimal difficulty. Results replicated on Dota 2.
OpenAI releases LOLA (Learning with Opponent-Learning Awareness), an algorithm that models other agents' learning and discovers collaborative strategies like tit-for-tat in the iterated prisoner's dilemma.
OpenAI shows that adding adaptive noise to reinforcement learning algorithm parameters frequently boosts performance. This exploration method is simple to implement and rarely decreases performance.
OpenAI created adversarial images that reliably fool neural network classifiers across varied scales and perspectives. This challenges a recent claim that self-driving cars would be hard to trick maliciously due to their multi-angle image capture.
User reports llama.cpp v9413 benchmarks on dual AMD MI50 with Debian Testing. Using Qwen 3.6-35B-GGUF, Vulkan reaches 977 t/s (PP) and 55 t/s (TG) at concurrency 1; ROCm-MTP delivers 92-115 t/s (TG) at concurrency 2. MTP (Multi-Token Prediction) significantly improves generation throughput.
Anvil, an open-source wrapper around llama.cpp, offers an alternative to Ollama with increased transparency: readable GGUF files, visible flags before execution, local fleet management, Hugging Face integration, and MCP support. Built to restore the local-first simplicity lost by Ollama.