frontier models
Six Useful Skills Beat a Much Larger Bank
SGUID retains procedural skills that provide effective distillation signals rather than selecting by relevance alone.
Summary
SGUID retains procedural skills that provide effective distillation signals rather than selecting by relevance alone.
Fewer than 25% of retrieved skills helped in the study’s on-policy distillation setup. Six selected skills matched or beat banks up to eleven times larger on three of four models; a second round with three new skills did so on all four. Qwen3-8B rose from 64.3% to 66.3% in that second round.
Why it matters
SGUID retains procedural skills that provide effective distillation signals rather than selecting by relevance alone.
Limits and context
No additional limitation was separately recorded.
Key claims
SGUID retains procedural skills that provide effective distillation signals rather than selecting by relevance alone.
Evidence: source-2026-10-09-018
Sources
- arXiv preprint 2610.12367, original v1arXiv / submitting research authors · primary research
Corrections
No corrections have been recorded for this story.