Core LLM inference & training advances (Inference Looping, ScheduleFree+, DeepSeek Sparse Attention)

New training-free Inference Looping (ODE-based recurrence) boosts Qwen3-4B by 2.6% on MMLU-Pro without fine-tuning. ScheduleFree+ (Meta FAIR) cuts training time by 31%. DeepSeek Sparse Attention and other efficiency techniques continue to improve production inference.