Huawei and UK Researchers Teach AI Models When Thinking Isn’t Worth the Tokens, Cutting Agent Costs 80% With a Router That Weighs a Few Hundred Kilobytes

A Huawei-led research team’s Hybrid Decision Model lets a frozen AI model decide whether reasoning is worth paying for, cutting agent token costs by 80% and response times by 72.5% on one benchmark, with a router that adds only a few hundred kilobytes.

Read more

NVIDIA’s Sigma Drafts Text in Parallel Blocks and Runs About 3.5 Times Faster Than Standard Decoding, Making Diffusion a Real Contender for Cheaper AI Inference

NVIDIA, Cornell and Georgia Tech researchers detail Nemotron-Labs-Sigma, a 3B/8B continuous diffusion language model that drafts text in parallel blocks, reporting up to 8x faster math generation after distillation and about 3.5x throughput over standard decoding, with clear caveats on scale and coding accuracy.

Read more