Trase has secured the number one position on Hugging Face's GAIA benchmark, a test measuring real-world AI agent capabilities including reasoning and tool use, outperforming submissions from Meta, Microsoft, and others.
- Trase achieved an overall success rate of 35.55% on the 450+ question benchmark.
- The system uses iterative self-improvement via STaR (Self-Taught Reasoner) fine-tuned on human-like action trajectories.
- It augments base LLM actions with code to compose multiple tool calls into single code actions.
- A top-level agent performs self-critique to determine if sufficient information has been gathered before switching tools or calling specialized agents.
The article notes that 90% proficiency is considered the threshold for replacing humans in complex tasks, a goal Trase aims to reach.