LABD: Loop-Driven Agentic Behavior Distillation
Graduation project: teaching small language models to correct themselves over multiple steps, through distilled agentic loops, SFT and GRPO, to get past the “Reasoning Cliff”.
Small and mid-scale language models (SLMs) struggle with iterative self-correction, showing a pronounced “Reasoning Cliff”: a sharp drop in quality when a task needs multi-step reasoning or recovery from feedback.
Prompting-based methods fail at this scale because self-correction in small models is not emergent; it has to be taught explicitly. LABD distills loop-driven agentic behavior into small models with supervised fine-tuning and reinforcement learning.