Researchers built an automated system to improve small AI models
The new Turnslide method helps small models master multi-step tasks while saving significant computational costs.
A new research paper introduces Turnslide, an automated framework designed to help small language models (SLMs) perform complex, multi-step actions. Small models are popular because they can run on private hardware, but they often struggle when they need to use several digital tools in a specific sequence.
Usually, training these models to handle tool-calling requires vast amounts of manual data, which is expensive and rare. Turnslide solves this by using a mathematical approach where every tool is treated as part of a finite-state machine. This allows the system to generate logical, valid sequences of actions without needing humans to write every example.
Our FSM-based generation significantly improves downstream accuracy over an unmutated baseline.
The framework generates these sequences and translates them into full training examples with only a single request to a larger model. By focusing on targeted distributions of task difficulty and length rather than just random variety, the researchers achieved higher accuracy.
In their testing, the team found that models trained with this method reached 70.7% accuracy on tasks. This outperformed existing methods that hit 63.4% or 53.7%, all while using up to 6.6 times fewer data tokens.
Think of it like teaching a student through a structured curriculum rather than handing them an entire library of random books. Is efficiency the final frontier for making powerful AI accessible on your own devices?
It is important to note that this research focuses on synthetic data generation for tool-calling capabilities. The study does not address how these models perform on tasks outside of defined API sequences or how they handle noise in real-world data environments.
Source: arXiv cs.CL