Reasoning LLMs (RLLMs), with their ability to self-correct and perform multi-step reasoning, perform excellently on complex programming and planning tasks, but their training cost is extremely high. Researchers from MIT, NVIDIA, ETH Zurich, and others found that up to 85% of training time is consumed by the "rollout" stage — the model generates multiple candidate answers for scoring, where only a small fraction of GPUs are working, while the rest sit idle.
Core breakthrough: adaptive draft model
The team designed an adaptive system that, upon detecting some GPUs sitting idle, automatically enables a smaller, faster "Drafter" model to predict the large model's output, which the large model then verifies. The draft model intervenes only when needed, with no extra overhead.
The key is "adaptive" — the draft model is dynamically scheduled based on real-time load, not fixed-ratio allocation, which solves the unstable draft-hit-rate problem of traditional speculative decoding. Experiments show that, while preserving accuracy, this method boosts training speed 2-3×.
Why it matters
Training a frontier reasoning model already costs millions of dollars in electricity. This kind of "lossless" optimization means there's no need to compromise between accuracy and efficiency. Even more importantly, the lightweight draft model trained as a byproduct can itself be directly used for inference acceleration, killing two birds with one stone.
Industry implications
As LLM capability competition gradually hits the compute ceiling, training-efficiency optimization is moving from the margins to center stage. FlashAttention solved attention-compute memory issues, MTP (Majorization-Threatens-Posterior) improved speculative decoding, and MIT's work here proves that beneath the "GPU idle" tip of the iceberg there's still substantial optimization space. In 2026, LLM competition may not just be about the model itself, but about who can squeeze the most out of compute.