Self-Policy Distillation via Capability-Selective Subspace Projection
- Published
- Source
- arXiv
- Paper number
- 224
- Field
- LLMs / NLP
- arXiv ID
- 2605.22675
Key points
- Independence: It does not require an external teacher or reward model.
- Generalization: It works across heterogeneous domains and transfers between tasks.
- Selectivity: It targets the underlying capability signal rather than superficial patterns.
Paper links
External research summaries. These are not HDATF publications or measured product results.