Agentic AI
New distillation method lets smaller AI models match capabilities of much larger ones
Researchers show that training student models on their own mistakes—rather than copying teacher models—helps close the performance gap on complex reasoning tasks.
A fresh approach to knowledge distillation enables smaller language models to narrow the capability gap with their much larger counterparts on complex multi-step reasoning and tool-use tasks. The technique, called SCoRe (Student-Centered Reasoning), trains smaller student models using data tailored to their own strengths and weaknesses rather than forcing them to mimic larger teacher models wholesale.
The problem with traditional distillation#
Existing methods for shrinking large models typically have students attempt to replicate the full sequence of steps taken by a larger teacher. However, when reasoning abilities or factual knowledge differ between teacher and student, errors can compound down the line. Small mistakes early in a reasoning chain can snowball into increasingly wrong answers by the end.
How SCoRe works differently#
Instead of copying complete teacher trajectories, SCoRe flips the process: the student model generates its own solution attempt, and the teacher reviews it only to correct the very first mistake. This creates training data that matches what the student is actually capable of, revealing its specific problem areas.
The training unfolds in two phases:
- Supervised fine-tuning on corrected trajectories—allowing the student to learn from its own corrected steps
- Short-horizon reinforcement learning that begins from the verified part of the solution (before the earliest error) and targets that specific error point, letting the student explore solutions independently rather than copying the teacher
This short-horizon RL design also improves training stability compared to longer reasoning sequences.
Empirical results#
Testing across 12 challenging benchmarks, researchers found that a 7-billion-parameter student model trained with SCoRe achieved performance comparable to a 72-billion-parameter teacher. This suggests the approach offers a genuine leap in efficiency for deploying capable agentic AI systems at smaller scale.
