Meta FAIR scaling law shows tool diversity improves agent training efficiency

Meta's Fundamental AI Research division has derived a scaling law showing that increasing tool and API diversity improves agent performance more than generating additional tasks on a fixed set of tools, Tech Times reported. The findings suggest that training pipelines should prioritize tool variety rather than task volume. [1]

Tech Times reported that the Meta FAIR researchers established the scaling law after evaluating 849 trained models across 458,752 generated tasks. The research team utilized a pipeline called SynthAPI to generate the training data and conduct the experiments. [1]

Agents trained using the SynthAPI pipeline outperformed NVIDIA's Nemotron-Agentic tool-use baseline by 30.5 percentage points, according to Tech Times. The results challenge standard agent training practices that allocate data budgets primarily to scaling task counts. [1]

Separately, MarkTechPost reported that Samaya AI released Sterling, a post-trained financial research model built on NVIDIA's Nemotron 3.5 Lightning architecture. The model contains 30 billion total parameters with 3 billion active parameters. [2]

According to MarkTechPost, Sterling achieved a 58.9 percent score on the FrontierFinance benchmark at an average cost of $0.61 per question. In comparison, Anthropic's Claude Fable 5.1 scored 55.9 percent at a cost of $6.89 per question. [2]