MetaExp adds 14.8 points on AutomationBench to an agent model
The researchers report that the result holds when the training data does not include the test subject.
Claimed, not confirmed
Researchers used 4,956 task records from 1,000 task families to fine-tune the Qwen3.5-122B-A10B model. The model got 8.5 more points on SpreadsheetBench and 14.8 more on AutomationBench. The source reports no test by a different group. For a person who makes agents, a model can get new skill from use.
Sources
Posted
