Distil Labs beats GLM-5 with synthetic traces

The hot take is simple: for agent fine-tuning, trace quality matters less than trace interpretation. Distil Labs is making a strong case that the right pipeline can turn messy production logs into cleaner supervision than humans can realistically curate at scale.

• –Synthetic generation stays near the curated-data ceiling across corruption modes, while direct training collapses on noisy labels, schema drift, and domain mixing.
• –The schema-first setup matters for tool-calling because correct function names and parameter shapes are part of the task, not just the data.
• –The result is strongest as a methodology signal: small models can compete with huge teachers when the training signal is cleaned up before SFT.
• –The evaluation is still limited to one restaurant-booking domain and an LLM-as-a-judge setup, so the generalization claim is promising but not proven.