Skip to main content

Research project / Blades

Investigating transfer between model representations.

Hidden state injection improved reported accuracy on a direct medical reasoning task for one Phi-to-MediPhi model pair. Transfer failed in the other three tested pairs.

Read the research Explore the code
MODEL COMPOSITIONBLADES / 03 Sourcehidden states Gatedinjection Targetmodel layers CONCEPTUAL INJECTION PATH
Conceptual hidden state injection. Evaluation scope and controls are described in the paper.

The public record

Blades: Hidden State Injection from Phi Reasoning Models into MediPhi

Andrew Young · March 24, 2026 · Preprint

Scope
Four model pairs; one successful Phi reasoning-to-MediPhi transfer. Reported medical-task accuracy: target 55.4%, source 62.5%, combined 69.6%.
Current limitation
Direct task evaluation, not a standardized MedQA result. Layer-table baselines, gate fitting and per-query source-model cost remain unresolved.
Available artifacts
Original preprint PDF available. Model Garage is linked in the article; run-level data and an exact reproduction revision are not identified here.
Read methods and reported results ↗Original paper PDF ↗

Editorial correction dated 2026-09-22; no new experiment or independent reproduction. The linked PDF is the original version and does not include these web corrections.