narphorium.com reposted
An interesting paper with implications for skill extraction, continual learning, and token efficiency - it provides even more incentive & value from running user-specific task evals because you can extract skills from them. arxiv.org/abs/2608.07885
AI Weekly's analysis
→
- A coding agent compiles training-trajectory patterns into a natural-language skill injected into a non-reasoning model's system prompt.
- Across four agentic benchmarks the skills recover 55%-100%+ of the reasoning gap for GPT-5.4-mini, beating reasoning mode on two of four.
- Runs emit 2.7-6x fewer output tokens and zero reasoning tokens, and skills work even when distilled from non-reasoning trajectories alone.
Read full analysis →
View on Bluesky →
