Alibaba's SkillWeaver AI Cuts Token Usage by 99% While Routing Complex Tasks to the Right Tools
Summary
Alibaba's new SkillWeaver AI framework slashes token usage by over 99% — from 884,000 to just 1,160 per query — using a three-stage pipeline that intelligently routes complex multi-step tasks to the right tools, with accuracy jumping as high as 92% on larger models.
Key Points
- Alibaba researchers introduce SkillWeaver, a new AI framework that uses a three-stage Decompose-Retrieve-Compose pipeline to route complex multi-step tasks to the correct tools, reducing token consumption by over 99% compared to loading an entire tool library.
- A key innovation called Skill-Aware Decomposition (SAD) uses an iterative feedback loop to refine task breakdowns by feeding retrieved tool hints back to the LLM, boosting decomposition accuracy from 51% to 67.7% on a 7B model and up to 92% on larger models.
- Testing on a custom 300-query benchmark with 2,209 real-world tools shows SkillWeaver slashes context usage from roughly 884,000 tokens to just 1,160 per query, though developers must build their own error recovery mechanisms since the framework currently lacks fault handling for failed multi-step tool chains.