Utility-based benchmark for measuring how well image captions preserve image-level information

Active Partial Rollouts in RL to tame long-tail generation, achieving 44% throughput improvement

Open-source vision-language model combining CLIP encoder with language model for multimodal tasks

Text-to-image generation with 1D binary latents achieving 32× token reduction

Math reasoning model trained with long chain-of-thought reinforcement learning

Fully open 3B language models with stellar performance, trained on AMD Instinct MI300X GPUs

LLM agents as research assistants for automated scientific discovery