Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows
Curated executable tools, not one-off generated “skills,” drove most of the measured gain in financial agent workflows.
The paper introduces FinSkillBench, a 2,603-episode benchmark for verifiable finance tasks across portfolio construction, risk management, and fundamental analysis. Across the paired model analysis, curated skill packages raised mean scores by 16.2 points, while skills generated inside a single episode added only 0.5 points and used more tokens and turns. When the authors split the package apart, executable tools alone added 19.5 points, documents alone added 5.6 points, and the combination was subadditive. The result varied by workflow, leading the authors to treat the “skill premium” as a property of the model, resources, and harness together, not the model alone. ArXiv · AI/CL/LG's note
The paper introduces FinSkillBench, a 2,603-episode benchmark for verifiable finance tasks across portfolio construction, risk management, and fundamental analysis. Across the paired model analysis, curated skill packages raised mean scores by 16.2 points, while skills generated inside a single episode added only 0.5 points and used more tokens and turns. When the authors split the package apart, executable tools alone added 19.5 points, documents alone added 5.6 points, and the combination was subadditive. The result varied by workflow, leading the authors to treat the “skill premium” as a property of the model, resources, and harness together, not the model alone. ArXiv · AI/CL/LG's note
score 5