SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models
This work proposes SCRIBE, a framework addressing the training instability of tool-augmented language models in multi-step reasoning caused by credit assignment difficulties. SCRIBE constructs a structured skill prototype library to validate subgoals at an intermediate level of abstraction, transforming open-ended LLM evaluation into a constrained verification task that yields precise, low-variance process-level reward signals. It introduces structured supervision at the skill-abstraction level for the first time, decoupling high-level planning from low-level execution and revealing a co-evolutionary mechanism wherein mid-level skill mastery precedes the emergence of high-level planning capabilities. Experiments demonstrate state-of-the-art performance across multiple reasoning and tool-use benchmarks: Qwen3-4B achieves 63.3% accuracy on AIME25, a substantial improvement from 43.3%, significantly boosting success rates in complex, multi-turn tool interactions and showing strong compatibility with underlying tool optimization methods.