A research team has released SCRIBE, a training framework that makes AI agents more reliable at multi-step tool use by rewarding the right things at the right level of abstraction.
The core problem SCRIBE targets is credit assignment: when an AI agent fails a complex task, it's hard to know which step went wrong. Existing approaches use large language models as judges to score intermediate steps, but those judges tend to be inconsistent because they're evaluating open-ended behavior without clear rubrics. SCRIBE — short for Skill-Conditioned Reward with Intermediate Behavioral Evaluation — fixes this by maintaining a library of "skill prototypes," essentially structured templates that define what good execution looks like for specific sub-tasks. Each step in the agent's reasoning gets routed to a matching prototype, turning a fuzzy judgment call into a constrained verification check.
The results are hard to ignore. On the AIME25 math benchmark, SCRIBE pushed a Qwen3-4B model from 43.3% to 63.3% accuracy — a 20-point jump from a relatively compact model. That matters because most accuracy gains in this range come from scaling up model size or adding more compute, not from smarter training signals. SCRIBE also claims to be additive to existing low-level tool optimizations, meaning teams can stack it on top of other improvements rather than choosing between them.
The underlying insight — that mid-level skill mastery reliably precedes high-level planning ability — echoes how humans learn complex tasks, and suggests current agent training pipelines may be skipping a developmental stage that actually matters. Whether the skill-prototype library scales gracefully to messier real-world tool environments, rather than cleaner benchmark settings, is the question that will determine if this moves beyond the lab.